The Intersection of Industry Occupation and Job Tasks with Hote

Section 1: Introduction and Background

Overall Goals of the Dissertation

Twenty years after the introduction of luxury accommodations, hotel room cleaners continue to

face the uncontrolled ergonomic hazards associated with hotel room cleaning. Upon the

promulgation of Cal/OSHA 3345 in 2018 – an occupational health regulation for the prevention

of musculoskeletal injuries occurring to hotel housekeepers, then Cal/OSHA Chief Sum stated,

“Hotel housekeepers have higher rates of acute and cumulative injuries compared to workers in

other industries, and data shows those injuries have steadily increased…”. This regulation

requires employers to identify, evaluate and correct housekeeping-related hazards….”.1,2 Hotel

room cleaner reports of increased workload and consequential pain are supported by worker

surveys and biomechanical studies demonstrating a high risk of low back disorders by specific

hotel room cleaning job tasks.3-9 Research is needed to characterize musculoskeletal disorder

injuries (MSDs) occurring to hotel room cleaners and to identify tasks associated with MSDs.

This includes injuries occurring to parts of the body other than the lower back such as shoulders,

elbows, knees, etc.10 Researchers call for expanded analyses of hotel room cleaner injury risk

factors and application of ergonomic interventions to decrease exposure to MSD risks when

performing hotel room cleaning tasks 11,12

The goals of this dissertation are to identify hotel room cleaners as an occupation experiencing

work-related musculoskeletal injuries (Aim 1), identify task at time of injury (Aim 2), provide a

descriptive analysis of hotel room cleaner MSD injuries (Aim 3A) and explore the relationship

between task and injury outcomes (Aim 3B) using narrative text data from workers’

compensation claims. These goals attempt to address the data gaps described below. In carrying

out the aims of this dissertation, the occupation of hotel room cleaner and the task at time of

injury were identified and coded, making data analysis possible. Data analysis then described

worker and injury characteristics and associations between task and injury outcomes. The

findings from my dissertation analyses contribute new information to aid in targeting injury

reduction efforts by nature of injury (e.g., sprain or strain), cause of injury (e.g., lifting or

pushing or pulling), and part of body injured. Moreover, these findings lay the foundation for

hotel room cleaning task-specific worksite interventions. The ideal administrative data that

combines demographic, employment and occupation and injury information is workers’

compensation data. The source data for my dissertation is California workers’ compensation

claims data for hotel industry employees.

Significance of Data Gaps Addressed by Dissertation Goals

Although the Bureau of Labor Statistics (BLS) publishes injury rates and injury and worker

characteristics by industry, e.g., hotels, and by occupation, e.g., maids and housekeepers, the

BLS does not publish cross-referenced BLS occupation injury rate data with BLS industry data.

Therefore, what we know about hotel room cleaner work-related pain and injury comes mostly

from academic studies. Lacking from the literature are peer-reviewed studies using

administrative data such as workers’ compensation data.

The lack of occupational injury surveillance data and peer-reviewed administrative data studies

about hotel room cleaner MSDs is significant for several reasons and at the key intervention

levels of industry, occupation and worksite (hotel). First, at the industry level, the U.S. traveler

accommodation industry (NAICS 7211) ranks number one among all industries (Table 1.1) for

employing Maids and Housekeeping Cleaners, representing 59.8% of the total employment by

the top five industries for this occupation.13 Maids and Housekeeping Cleaners is the Census

Bureau’s standard occupation classification “SOC 37-2012” with job tasks most similar to those

of hotel room cleaners.14 In 2018, the hotel industry’s incidence rate for nonfatal occupational

injuries and illnesses was 1.5 times the rate for private industry, indicating an excess of injuries.15

Second, at the occupation level, Maids and Housekeeping Cleaners are a high-risk occupation for

MSDs, ranking among the BLS’s ten high-risk occupations for MSDs for all industries for 2017

with an incidence rate more than three times that for all workers with

MSDs.16 Third, at the worksite (hotel) level, the occupation of Maids and Housekeeping Cleaners

accounts for 23% of the workforce in the traveler accommodation industry, the largest proportion

overall, making this occupation a hotel’s largest workforce.13

For a high-risk worker population employed in an industry with high rates of nonfatal

occupational injuries and illnesses that is the leading employer for this workforce, the lack of

surveillance data is a serious obstacle to employer and public health injury prevention and

control efforts.

Table 1.1 Five Industries with the Highest Levels of Employment for Maids and

Housekeeping Cleaners. Occupational Employment and Wages, May 2019. Bureau of Labor

Statistics

Industries Employing 37-2012 Maids and Housekeeping Cleaners

Industry

Employment

Travel Accommodation

467,270

Services to Buildings and Dwellings

100,490

General Medical and Surgical Hospitals

95,050

Nursing Care Facilities (Skilled Nursing Facilities)

71,480

Continuing Care Retirement Communities and

Assisted Living Facilities for the Elderly

46,480

Total employment, top five industries

780,770

Part 1. Public Health Issue of Concern: Musculoskeletal Disorder Injuries and Hotel Room

Cleaners

a. What do we know about the hotel room cleaner workforce?

Who are hotel room cleaners?

In the United States, hotel room cleaners are a predominantly female workforce with a large

percentage of Hispanics/Latinas followed by Asian/Pacific Islanders and African-Americans.7,10

Using data from the U.S. Census Bureau’s standard occupation classification of 37-2012 Maids

and Housekeeping Cleaners, 85.3% of this workforce is female and the average age for female

hotel room cleaners is 44.7 years.17,18 Immigrants account for the majority of hotel room

cleaners.7 Studies identify musculoskeletal disorders due to awkward postures as a concern for

immigrant cleaners who face multiple barriers such as low literacy skills, few resources and

cultural factors.19 Latina and immigrant hotel room cleaners were reported to experience

occupational disparities including higher injury rates for hotel housekeepers who identified as

Hispanic compared to their white counterparts.10,20

There are a variety of occupational terms used for this workforce, e.g., hotel maid, hotel

housekeeper, guest room attendant or room attendant, all of which are occupations in a hotel’s

housekeeping department. An additional job classification, housemen/housepersons is often

discussed along with hotel room cleaners as it is the job classification in a hotel’s housekeeping

department that removes dirty linens from guest rooms and cleans extra dirty rooms.

Historically, housemen had been a predominantly male occupation; in recent years, this job

classification is referred to as “housepersons” as more women are filling these jobs. For the

purposes of my research, the all-encompassing term ‘hotel room cleaner’ will be used throughout

this dissertation except for where the text refers to other terms used in the scientific literature

and/or government sources.

What Job Tasks Does a Hotel Room Cleaner Perform?

Hotel room cleaning is physically demanding work. Tasks associated with hotel room cleaning

include: making beds, vacuuming guest rooms, dusting, cleaning tubs and shower walls,

scrubbing bathroom floors, cleaning mirrors, polishing furniture, removing trash, pushing carts

and moving furniture to perform room cleaning tasks, e.g., to vacuum, to clean under beds.21-23

Hotel room cleaners are required to clean a certain number of rooms during a shift, known as the

“daily room quota.” The daily room quota is broken down by the number, room status and type

of room the hotel room cleaner is assigned to clean. For example, there are two categories of

room status, each one with a corresponding level of housekeeping service: stay over (guest is

staying another night), and check outs (guest is checking out), with check outs requiring more

cleaning work as the guest room is prepared for a new guest. The type of room includes: king

(king-size bed), double doubles (a guest room with two double beds) and suites or VIP rooms

(may include a living room with a sofa bed and a kitchenette) with the latter two types requiring

more work and greater physical demand. Hotel room cleaners report working quickly, taking

shorter breaks or none at all and working long hours to meet these quotas.6

b. What do we know from existing research about work-related injuries of hotel room cleaners?

Work-related Injury Rates of Hotel Room Cleaners

A study of U.S. hotel workers was undertaken using employers’ records of 2,865 work-related

injuries during 55,327 worker-years of exposure and hiring data that provided job tenure and

self-identified gender and race/ethnicity information.10 The study authors, academics and hotel

union-based researchers, found that by occupation, room cleaners had the highest rate of

workrelated injuries overall, the highest rate of MSD injuries, and the second-highest rate of

acute trauma compared to all hotel workers studied. In addition, when comparing by

race/ethnicity among hotel room cleaners, Hispanic hotel room cleaners had an 82% greater

chance of an MSD injury than white female room cleaners studied. This is the only study

reporting work-related injury rates in the United States by hotel worker occupations.

Work-related Pain Prevalence Rates

Findings from housekeeper surveys performed by academic researchers and labor union staff,

describe high prevalence rates of work-related pain, yet low rates of reporting such pain to

employers, highlighting that underreporting of work-related injuries is prevalent among hotel

room cleaners.4-7 All of these studies took place after the increase in luxury accommodations

which occurred in the hotel industry in the late 1990s. These low rates of reporting work-related

pain to employers occur despite high rates of prevalence of severe bodily pain, of medication use

and of doctor visits.4-6 In a study by university faculty and a labor union, Lee and Krause found

that of 258 unionized hotel room cleaners surveyed from four San Francisco hotels, 77% had

work-related pain in the past year; 73% of cases reported going to a doctor and 53% stayed home

from work, yet only 50% reported the pain to management and even less, 23% filed a formal

injury report in the past year.5 In a study of 941 hotel room cleaners from union hotels in Las

Vegas about increased workload, Krause and co-authors found that 47% of those surveyed

reported severe pain in the past four weeks and 84% took medication for pain they experienced

in the workplace in the same time period.6 In the same study, 78% of hotel room cleaners

reported experiencing pain the past 12 months that they believed to be work-related; of these,

62% went to a doctor for the pain. Similar findings were described in a report by UNITE HERE,

the hospitality workers’ union representing workers in North America. From 600 hotel

housekeepers surveyed in the U.S. and Canada, 91% reported work-related pain. Of those

reporting pain, 66% took pain medication and two-thirds visited their doctor because of the

pain.4 In a study by Scherzer and co-authors using the responses from the same surveyed 941

unionized hotel room cleaners from five Las Vegas hotels, 31% reported pain to management and

20% reported a workers’ compensation injury, with 35% of claims denied.7

Causes of Work-Related Hotel Room Cleaner Pain and Injuries

There are several factors that contribute to the causes of occupational hotel room cleaner pain

and injuries. These include the physically demanding nature of the work, failure by the hotel

industry to control the hazards associated with hotel room cleaning, changes in hotel guest

accommodations that increase workload, green programs that reduce staffing and increase injury

risk, aspects of how the work is organized (room quota, speed-up), reduction in staffing in the

hotel’s housekeeping department and lack of recognition by the hotel industry of the impact of

these factors on hotel room cleaners’ health. These factors are described in detail below.

c. What do we know about hotel industry changes in guest accommodations?

Intensification of Competition in the Hotel Sector

Competition for hotel guests has resulted in changes over the past twenty years by the hotel

industry that include providing luxurious sleeping accommodations, promoting hotel ‘green

programs’ to guests and reducing housekeeping staff. These changes were introduced in hotel

guest rooms starting in 1999, a period known as the “Bed Wars” where each hotel company

“brands” (differentiates) its bed package (mattress, sheets, duvets, pillows) from those of other

hotel companies, e.g., the Westin’s Heavenly Bed, the Radisson’s Sleep Number Bed.24

Intensification of Hotel Room Cleaner Workload

Fancy bedding such as triple sheeting, thick duvets, high thread-count sheets and more pillows

are the standard today, increasing the number of pounds of linen per hotel guest room.25 Hotel

luxury accommodations include mattresses that weigh over 100 pounds and duvets that can

weigh 14 pounds.26,27 For hotel room cleaners, this means forceful lifts during bedmaking and

heavier linens to pull off when stripping the beds. While these changes in hotel accommodations

occurred, the hotel industry also adopted lean staffing models that reduced the number of hotel

room cleaners and housemen, two key job classifications in a hotel’s housekeeping department.28

As a result of staffing cuts, the tasks of housemen were added to those of hotel room cleaners.

For example, heavy loads of dirty linens previously removed by housemen, are now added to

housekeeping carts already laden with the usual load of clean linens, towels, amenities and

vacuum cleaners. These carts are pushed by hotel room cleaners from room to room down

carpeted hallways and in and out of elevators to clean rooms on different floors.26

Renovations are another way to keep hotels competitive with the most recent change being the

installation of floor to ceiling glass shower doors and walls. Hotel room cleaners complain of a

lack of additional time and long-handled tools to clean the glass doors and walls. Instead,

cleaning is performed by extended reaches of the hotel room cleaner’s arm(s) above the

shoulders, adding forceful exertions and strain. Testimony from hotel room cleaners in support of

a state-level regulation proposed by the California Department of Occupational Safety and

Health to prevent musculoskeletal injuries in hotel housekeeping raised the following concerns:3

“Candy Hu, Hotel Housekeeper, Unite Here Local 2, ... “the hardest part of her job is cleaning

the bathroom because she must stand on her tiptoes and reach up high to clean the top part of the

glass shower doors.”

“Irma Perez, Hotel Housekeeper, Unite Here Local 2850 Oakland, … “it is important that

employers provide hotel housekeepers with the proper tools to do their jobs safely, and that they

also provide training for their housekeepers and supervisors on how to use those tools properly.”

“Olga Manrique, Hotel Housekeeper, Unite Here Local 19, … “many housekeepers must reach

up high to clean walls and other surfaces, and as a result of not having the proper tools, they

injure their arms and backs.”

Increased workloads associated with changes in hotel accommodations and hotel renovations

described above, coupled with the room quota system, contribute to setting the pace at which

tasks get done; pace then determines whether a slower, safer room cleaning work practice is

performed versus a quicker, less safe one.

Another way of marketing a particular hotel company’s brand is by offering guests the hotel’s

‘green’ program. Hotel green programs were established ostensibly to reduce water, energy, and

detergent usage, with over 6 million Starwood Hotel guests taking part in such programs since

2009.29 These green programs depend on a hotel guest’s refusal of daily room cleanings in

exchange for receiving food and beverage coupons or loyalty points in a hotel’s guest rewards

program where you can apply points towards a hotel stay. When comparing employment and

injury data for 2013 and 2017, a study of 23 hotels in nine U.S. cities that practiced green

programs found a decrease of 700,000 total work hours (the equivalent of 350 full-time jobs) and

a 49% increase in injuries.30

d. What is the public health significance of musculoskeletal disorder injuries?

According to the Centers for Disease Control, “musculoskeletal disorders (MSD) are injuries or

disorders of the muscles, nerves, tendons, joints, cartilage, and spinal discs”.31 The World Health

Organization characterizes musculoskeletal injuries as those accompanied by pain and restriction

of movements, resulting in permanent work disability, income loss and limitations in social

interactions.32 The National Institute of Occupational Safety and Health (NIOSH) defines a

subset of musculoskeletal injuries to be “work-related musculoskeletal disorders (WMSDs)”

where “the work environment and performance of work contribute significantly to the condition

or where MSDs are made worse or longer lasting by work conditions”.33

Approximately one-third of all occupational injuries in the United States are musculoskeletal

injuries, which also account for the majority of lost or restricted work-time cases.34 WMSDs are

costly, result in lost productivity and substantial human suffering. For example, for the United

States in 2007, the average medical cost of an individual work-related MSD case was $2,924

dollars; the total estimated costs for WMSDs nationally was $3.4 billion, with adjustments for

underreporting.35 An additional $1.1 billion is estimated in costs for lost workdays of injured

workers and replacing employees permanently disabled for the same time period.35 Examples of

how WMSDs interfere with workers’ routine activities range from bathing to childcare and can

include financial disruption resulting in car and home loss.36 Ergonomic interventions are

recognized as the next step towards reducing worker MSD injuries and employer workers’

compensation costs.34 Since my focus is on work-related hotel room cleaner MSDs, these

injuries will be referred to simply as MSDs, not WMSDs, throughout this dissertation.

The Importance of Low Back Disorders

Research about musculoskeletal disorder injuries occurring to hotel workers from four job titles –

hotel housekeeper, kitchen worker/cook, dishwasher/steward and banquet server - identified the

back as the most common (40%) part of body injured, with other research finding the lower back

reported as the most frequent (63%) site for severe pain.6,10 Low back disorders are the most

frequent type of work-related musculoskeletal injury and can negatively impact a worker’s

quality of life.37,38 High costs related to low back disorders are sustained by employers and

employees.37 Globally, work-related low back pain accounts for 818,000 disability-adjusted life

years lost each year.38 Disability-adjusted life years lost is a measure that represents a more

complete picture of the magnitude of health problems on a population level, making it useful for

health planning.39 These findings taken together indicate there is substantial potential to reduce

low back pain world-wide by reducing occupational risk factors for musculoskeletal injuries.38

The Intersection of Task and Hotel Room Cleaner MSD Injury Risk

Overlap exists among the most ergonomic risk factors for MSD injuries and those frequently associated

with hotel housekeeping tasks: repetitive motions/tasks, bending, extreme reaches, heavy lifting; pushing,

pulling, carrying heavy objects, and prolonged awkward postures.33,34 In addition to the aforementioned

risk factors, forceful exertions, contact stress, excessive work rate, and inadequate recovery time between

housekeeping tasks are also recognized as contributing towards risk of work-related MSDs.1

A biomechanical study as far back as 1992 identified “extreme postures” in occupational

bedmaking as associated with MSD injuries and attributed the trend to use large, heavy beds that

are low in height as possibly increasing MSD risk.40 Subsequent studies by the same authors

confirmed these earlier findings and broke down bedmaking into six isolated tasks: “bedding on,

bedding off, lift middle, lift corner, push bed and pull bed” and made recommendations for

decreasing spinal loading.41,42 The latter study concluded that static methods grossly

underestimated the amount of spinal loading and that safe lifting limits could be exceeded during

occupational bedmaking.

Biomechanical analyses of hotel room cleaning work using the Lumbar Motion Monitor (LMM)

found a 79% job risk for a low back disorder as a result of performing hotel room cleaning tasks

overall, a 69% risk from bedmaking tasks and a 66% risk from bathroom tasks.8 The LMM

evaluates the likelihood (risk) that performing a job overall or a specific task may result in a low

back disorder. The LMM uses the following five measures of trunk movement: 1) forward

bending, 2) side-to-side bending, 3) twisting, 4) the force exerted at the moment of lifting and 5)

a lift rate, making it a ‘dynamic’ assessment method.9 Jobs with an overall risk greater than 60%

are high-risk, followed by mid-risk jobs with a risk greater than 30% and less than 60% and

lowrisk jobs are those with risk less than 30%.9 It is important to note that in this evaluation,

none of the hotel room cleaning tasks fell into the low-risk zone (<30%).

The next step beyond estimating low back injury risk by task is to be able to associate tasks with

injury outcomes such as nature of injury and part of body injured. Such an analysis could

contribute new knowledge to target injury reduction efforts by nature of injury, task and part of

body location. These efforts lay the foundation for task-specific worksite interventions. To

perform such an analysis would require reviewing task and injury data from a source that

combines job task information with injury and medical information. For work-related MSDs,

workers’ compensation data is a recommended source, albeit with limitations.

Part 2. Public Health Research Approaches to Injury Control: Application to the

Prevention of Hotel Room Cleaner MSD Injuries and Identification of Research Gaps

Theoretical Framework of Injury Control and Risk Reduction

Three frameworks useful for injury prevention and control guided this dissertation: Haddon’s

Matrix, the WHO’s Public Health Approach and the Unified Model put forth by Lett,

Kobusingye and Sethi.

Haddon’s Matrix

Haddon’s Matrix provides the researcher with a framework to look at a variety of potential

causative factors of an injury event instead of only providing a description of an accident with no

thought about etiology.43 From the same article, Haddon points out that descriptions of accidents

have “… a long history and close relationship….to notions of personal vulnerability and

invulnerability.” Haddon’s inclusion of the physical and cultural environments shifts the focus

from solely the ‘host,’ which in the context of occupational injuries is the worker, to focus on the

‘vector’ which is the energy/hazard the worker experiences. From there, then the shift is to the

‘environmental factors,’ namely, workplace (physical) and work organization (cultural). For the

occurrence of MSD injuries to hotel room cleaners, Haddon’s matrix has many applications. For

the vector, also known as the “etiologic agent”, one hotel room cleaning example is the

mechanical force exerted to lift mattresses and the movement of the hotel room cleaner’s trunk

(back) to make the bed as demonstrated by LMM biomechanical evaluations.43 Another example

is the speed with which hotel room cleaners move from one task to another. The physical and

cultural environments are critical factors for the occurrence of MSD injuries to hotel room

cleaners as illustrated in the previous paragraphs about how physically demanding the work is

(biomechanical studies, worker surveys) and the intensification of hotel room cleaner workload

due to changes in the hotel industry’s luxury accommodations and business model regarding

staffing.

By focusing on the etiology – the agent (vector) and environmental factors, then you can focus

on what Haddon calls “control opportunities.” These control opportunities present themselves in

the three phases of Haddon’s time dimension. The recognition of control opportunities and the

inclusion of a time dimension sets Haddon’s Matrix apart by establishing that injuries are not

only predictable events but also preventable ones.44 Injury prevention comprises the first phase –

prevention of the etiologic agent reaching the host. For hotel room cleaner MSDs, it is not

possible to prevent the force needed to lift mattresses to make the beds from reaching the hotel

room cleaner. Instead, we look at the second phase that “involves the interaction of the etiologic

agents and the susceptible structures”.43

In phase two, the amount of upper extremity and back exertion is statistically significantly

reduced by using a mattress lift tool and fitted sheets together.45 The mattress lift tool and fitted

sheet in this study are interventions and are examples of focusing on the etiologic agent

(mechanical energy), the physical environment (weight of the mattress) and the cultural

environment (using a flat sheet, more common in hotels, than a fitted sheet). This framework

provides an emphasis on interventions for changing the workplace and work organization and not

solely changing the actions of the host (the worker), e.g., training the worker in safe lifting.

Although safe lifting techniques by the worker are important, Haddon’s phase two emphasizes

solutions and creates opportunities for innovation, like using a mattress lift tool with a fitted

sheet reducing the number of mattress lifts from 15 to 7.8 lifts, a 48% reduction in lifting.45

Haddon’s final phase, the third phase, focuses on “maximizing salvage, once damage has been

done to the susceptible structures”.43 It is this minimization of an injury’s damage to the worker

in the third phase where Haddon emphasizes “…..that make the problem [damage to the host] of

social concern….and must be reduced [the damage]. The problem is not a priori “to prevent

accidents” per se.” Based on Haddon’s emphasis, the third phase shifts the focus to reducing the

damage of the injury which would support researchers studying workers’ compensation and

rehabilitation strategies. For example, one study found modified-work programs to be an option

for an earlier return to work which benefited workers categorized as temporarily and

permanently disabled and provided cost savings to the employer.46

Haddon’s Matrix helps to identify important research gaps by breaking down the hotel room

cleaner MSD injury event into etiologic factors and intervention points using the three time

dimensions. Haddon’s Matrix helps to identify important areas of research and possible data

gaps. For example, research exists about the injury experience of the host (hotel room cleaner)

from university and union research efforts and other academic authors but not from other sources

such as government agencies like the Bureau of Labor Statistics and only limited ‘grey’

publications using administrative data such as workers’ compensation claims.5,6,10,11,20,47 For the

second phase that focuses on the interaction of the host with vectors (etiologic agents), when

reviewing the literature, only two peer-reviewed studies identify hazards and solutions to reduce

injury risk factors of hotel room cleaning. A study from 2005 found that a handled shower

cleaning tool resulted in more neutral postures for cleaning bathroom showers, although not

specific to the hotel industry, the findings are applicable.48 More recently, a 2019 bedmaking

study found that using a mattress lifting tool and a fitted sheet together significantly reduced the

count of mattress lifts and upper extremity and back exertions.45

As demonstrated above, more studies are needed about interventions specific to tasks performed

by hotel room cleaners that contribute to the risk of MSDs. Breaking down the job of hotel room

cleaning into tasks can lead to identifying 1) cause – the tasks that are high-risk for a MSD injury

and 2) interventions specific to those high-risk tasks for MSDs. Biomechanical studies referred

to earlier using the LMM technology, identify bedmaking and bathroom cleaning as tasks with a

greater likelihood, 69% and 66% respectively, of causing a low back disorder injury.8,9 The

LMM technology is unable to identify risk for other parts of the body besides the lower back. In

a study using administrative data, the back, upper extremities and shoulders were the leading

parts of body affected for hotel room cleaners with MSDs, respectively.10 The role of task and its

impact on upper extremities, shoulders and other possible parts of body requires further study.

The Public Health Approach

The World Health Organization’s (WHO) public health approach is an epidemiological

framework consisting of four hierarchal elements also known as steps: “surveillance; risk factor

identification; designing and piloting interventions, followed by evaluation; and large-scale

program implementation”.49,50 The Public Health Approach, while overlapping with Haddon’s

Matrix about risk factor and intervention identification, adds two population-level components:

surveillance and large-scale program implementation. The inclusion of surveillance in the Public

Health Approach is important because it is the first step in defining a public health problem that

affects a population instead of only a specific host and/or a specific event in time. For this

dissertation, the population of interest is hotel room clean workers and their work-related

injuries. Without surveillance data, the effectiveness of carrying out the next population-focused

step – a large-scale program – would be limited, at best or worse yet, may result in erroneous

conclusions and inaccurate subsequent actions in the application step.

My dissertation contributes to the Public Health Approach framework by identifying a population

at risk for MSD injuries - hotel room cleaners. Since my dissertation Aim 1 focuses on

identifying the occupation of hotel room cleaner in workers’ compensation MSD claims data, my

improved research methods identify the population at risk in this data. This enables researchers

to carry out surveillance studies of changes over time in the number, nature of injury, cause of

injury, part of body injured and other variables analyzed in Aim 3 of this dissertation.

The Unified Model

The Unified Model (Figure 1) combines the Public Health Approach with Haddon’s Matrix (Lett

et al., 2002) and provides a clear construct for injury investigation and science-based injury

prevention interventions. The Unified Model combines the epidemiological elements and ‘injury

as a disease’ concept of Haddon’s Matrix with the Public Health Approach’s focus on scientific

methods to perform necessary surveillance, followed by identification of risk factors and

evaluation of interventions before broad implementation. Taken together, the unified model

provides for a comprehensive and precise application of injury control measures and policies.

Figure 1 Unified Model. A Unified Framework for Injury Control: the Public Health Approach and

Haddon's Matrix Combined (Lett et al., 2002)

The Unified Model drives the aims of my dissertation, which are to improve research methods to

identify hotel room cleaners and task at time of injury in workers’ compensation claims to access

this data for statistical analyses and thereby contribute reliable information for a comprehensive

and coherent approach to injury control and prevention. This data is needed to counter the failure

by the hotel industry to control the hazards resulting in hotel room cleaner MSD injuries. The

ultimate goal of my dissertation is to develop methods that will aid researchers to produce

science-based evidence necessary for effective intervention development and program

implementation. As Lett et al., point out, without a comprehensive and coherent model like the

Unified Model with its three axes, erroneous assertions at best, false assertions at worst, about

causality and preventive measures can impede (continue to impede) valid interventions and

effective programs for injury control.

The Unified Model, and the select model elements from above, guide the dissertation to answer

the following methodological and research questions:

• Can we establish that as an occupation, hotel room cleaners experience MSD injuries?

• Can we characterize hotel room cleaner MSD claims by demographic factors of the host

such as gender and age?

• Can we characterize hotel room cleaner MSD claims by work-related factors of the host

such as job title category, e.g., housekeeper, maid and job tenure?

• What are the leading nature of injury categories found in the claims data?

• What are the leading cause of injury categories found in the claims data?

• What are the leading part of body injured categories found in the claims data?

• Can we characterize hotel room cleaner MSD injuries by task at time of injury?

Workers’ Compensation Surveillance Data: Application of the Public Health Approach

If we review again the four steps of the Public Health Approach: “1) surveillance, 2) risk factor

analysis to identify causes, 3) develop and test interventions, and 4) large-scale program

implementation and evaluation,” we will see that this framework coincides with the use of

workers’ compensation data for my dissertation.44,49,50 Each dissertation aim complements each

step of the framework with the exception of the last step which is beyond the scope of my

dissertation but is clearly an aspiration of my dissertation, namely that my dissertation research

can contribute to implementation of large-scale intervention efforts. Below are examples of how

my dissertation aims complement the first three steps of the Public Health Approach and

contribute towards the fourth step:

1) identify hotel room cleaners as an at-risk occupation for MSD injuries and thereby improve

occupational injury surveillance; 2) provide task at time of injury data to contribute towards

identifying job tasks associated with injury factors and outcomes and provide data for statistical

analyses and 3) perform descriptive analyses to characterize MSD injuries by several factors,

including task at time of injury, to contribute science-based evidence necessary for intervention

development and subsequent injury prevention program implementation. In summary, my

analysis will establish that hotel room cleaners experience MSD injuries, identify tasks

associated with MSD injury factors, and describe MSD injury characteristics. With the new

research methods, occupational injury surveillance using California workers’ compensation

claims data of hotel room cleaners will be possible and as interventions are developed, piloted

and evaluated, ideally as part of a large-scale injury control program, their impact can be

measured over time through surveillance efforts including changes in nature and cause of injury,

part of body injured and tasks associated with MSD injuries. Without the research methods to

identify occupation and task at time of injury, the above surveillance, intervention development

and program implementation and evaluation efforts would be impossible or at least seriously

limited or carried out in error. Workers’ compensation data provides information about the

claimant related to their occupation, injury, years worked and disability benefits. Medical billing

data can also be linked to the workers’ compensation data to provide information such as medical

diagnoses, treatment and medical benefits.

Part 3: Workers’ Compensation Data: An Approach to Address Hotel Room Cleaner Injury

Data Gaps

Without surveillance data, it is not possible to substantiate the intersection of MSD injury risks in

the occupation of cleaner with the MSD injury risks in the hotel industry, namely, hotel guest

room cleaning job tasks. This is precisely where the first half of the title of my dissertation

originates: “The Intersection of Industry, Occupation and Job Tasks with Hotel Room Cleaner

Musculoskeletal Disorder Injuries: A Methods Approach to the Analysis of California Workers’

Compensation Data.” It is the lack of surveillance data by the Bureau of Labor Statistics about

the epidemiology (the magnitude, characteristics, rates and trends) of hotel room cleaner MSD

injuries that is an obstacle to intervention development, intervention evaluation and program

implementation as part of injury control and prevention efforts. Workers compensation data –

characterization of hotel room cleaner MSD injuries and identification of task at time of injury –

can provide injury details that BLS data cannot and can be used for surveillance with limitations.

Some of these limitations include low rates of workers filing WC claims -- only 25% of workers

with musculoskeletal injuries filed a WC claim and of Las Vegas hotel room cleaners reporting

pain, only 20% filed a claim.7,51 In the latter case, this may reflect low rates of reporting

workrelated pain to employers by hotel room cleaners ranging from 31% to 50% and

discouragement to report pain or file a claim, due to 35% of hotel room cleaner WC claims being

denied.5,7 As a surveillance system, the “capture” rate of upper extremity MSD injuries by the

workers’ compensation system has been estimated at less than 8% when compared to other

sources of data.36 Numerous factors for not filing a workers’ compensation claim have been cited

in the literature, making it necessary to recognize the serious limitations of this data while

utilizing the important information it can provide about occupational injuries.52

The approach that I have chosen to address the data gaps related to MSD injuries and hotel room

cleaners is to use workers’ compensation (WC) claims data and specifically claims from the state

of California. Workers’ compensation records provide data variables such as “industry,

occupation, nature of injury and cause, hospitalizations, medical treatments and costs, disability

duration…”, among others, that comprise a robust database of information provided by

employers, healthcare providers and insurance companies.53 NIOSH’s Center for Workers’

Compensation Studies recommends WC research activities for the purposes of “supporting

primary-prevention efforts, developing interventions to improve safety, and evaluating the

primary-prevention effects of public policy toward workplace safety (such as OSHA

enforcement, OSHA regulations, and targeted safety interventions)”.54

Application of Workers’ Compensation Data for Task Analysis of Hotel Room Cleaner

Musculoskeletal Disorders

WC analysis at the task level could provide insight into the cause of injury and contributing

factors. WC data is a source of narrative text information for a specific industry and/or

occupation that potentially can be distilled down to the tasks associated with a certain nature of

injury and the circumstances at the time of the injury.

More directly relevant to the topic of this dissertation is a task analysis performed by the

California Division of Occupational Safety and Health on a subset of data from a study of hotel

room cleaner MSDs using 2009-2012 claims data from California’s Workers’ Compensation

Information System.11,47 In a manual review of a sample of 2,000 hotel room cleaner MSD

claims cases, task was identified in 66% of cases reviewed and five broad task categories were

identified – bedmaking, bathroom cleaning, vacuuming/room cleaning, cart handling and other.

Next, Neidhardt reviewed over 7,000 MSD hotel room cleaner claims cases and was able to

break down task categories further into subtasks such as scrubbing the tub as part of the

‘bathroom cleaning’ category. Although challenges exist for identifying occupation and task in

California workers’ compensation data, there remains significant potential for using this data for

injury control.11,47 Research is needed to improve accessibility to this information in claims data.

California Workers’ Compensation Data: A Priority for Hotel Room Cleaner MSD Injury

Research

There are several reasons for making California a priority when selecting state workers’

compensation data for research about this specific workforce. California is the state with the

highest employment of Maids and Housekeepers – with 101,370 employed in all industries as of

May 2019 shown in Table 1.2.13 In 2017, California ranked first among all U.S. states in the

amount of total workers’ compensation benefits paid, totaling $12.1 billion.55

Table 1.2 Five States with the Highest Employment Level for Maids and Housekeeping

Cleaners. Occupational Employment and Wages, May 2019. Bureau of Labor Statistics

To understand the scale of the number of WC claims filed annually in California, over 680,000

First Report of Injury Forms (FROI) were filed in 2019 as part of workers’ compensation

claims.56 Moreover, California is the only state with a hotel room cleaning musculoskeletal

injury prevention standard, which took effect on July 1, 2018; a pressing need exists for baseline

data about these injuries using the states’ Workers’ Compensation Information System.1

Exploration of Methodological Issues in Workers’ Compensation Data: Filling Research

Gaps

Employment o

f Maids and Housekeeping

Cleaners, by State, May 2019.

State

Employment

California

101,370

Florida

80,000

Texas

76,040

New York

47,990

North Carolina

32,530

Using California workers’ compensation injury claims data, the goals of this dissertation are to

identify and describe work-related musculoskeletal injuries occurring to hotel room cleaners

from 2007-2016 and identify task at time of injury from 2012-2016 to explore the relationship

between task and injury outcomes, e.g., nature of injury and part of body injured. In order to

accomplish these goals, existing research methods were evaluated and modified and new

methods were created and tested. The aims of this study address these areas of future research

important for injury prevention and control, worksite interventions and occupational injury

surveillance. The first two aims test research methods for identifying: 1) the occupation of hotel

room cleaners in hotel worker MSD injury claims data and 2) the task associated with the hotel

room cleaner MSD injury incident. The third aim examines the findings of descriptive analyses

that characterize hotel room cleaner MSD injuries utilizing select variables drawn from the data

extraction from California’s Workers’ Compensation Information Systems listed in Appendix E,

provided to the principal investigator by California’s Department of Workers’ Compensation.

This dissertation is a Proof of Concept study as it tests the feasibility of using modified or new

research methods to identify hotel room cleaner MSD claims and task at time of injury in a

dataset of WC hotel worker MSDs. Information in Section 3 Methods details the application of

previously validated occupation and task terms from other studies to the custom extraction that

created the dissertation administrative dataset and compares the results (evidence) of improved

and/or new research methods. Proof of Concept studies have been useful in administrative data

studies by demonstrating the feasibility of applying varying methods such as term extraction to

free text data in medical records, the use of validated algorithms for identifying multiple

morbidities for chronic disease surveillance and testing the efficacy of data linkage systems

across state and national jurisdictions to study healthcare outcomes of injury-related

hospitalizations.57-59 More recently, natural language processing techniques were found to be

useful in coding interviews with public health content, either as a first step in coding text data or

as an accuracy check after coding was completed.60 A Proof of Concept approach to evaluate the

potential of natural language processes in combination with human coding was appropriate to

access narrative next data in workers’ compensation claims and to improve research methods for

the purpose of quantitative analyses.

Dissertation Aims

The three aims of this dissertation research are:

Aim 1: To create a new variable ‘Hotel Room Cleaner’ by testing and modifying existing

methods to identify hotel room cleaners in the narrative text field for ‘occupation’ using pooled

2007-2016 California workers’ compensation claims data for hotel industry workers with

musculoskeletal disorder (MSD) injuries. This creates the dissertation dataset comprised of hotel

room cleaner MSD injury claims. This data is needed for analysis in Aim 3A and also provides

the data from which a task sample is created for analysis in Aim 3B.

Aim 2: To develop and test methods to create a new variable that identifies the task associated

with the hotel room cleaner MSD injury incident using narrative text information from the

‘injury description’ field in claims data for a subsample of the dissertation dataset. Identifying

task will provide the data needed for analyses in Aim 3B from 2007-2016 and contribute towards

intervention-related research.

Aim 3: To perform basic descriptive statistical analyses to characterize hotel room cleaner MSD

injuries by nature of injury (e.g., sprain or strain), cause of injury (e.g., lifting or pushing or

pulling), part of body injured, age, job tenure, and task at time of injury.

Aim 3A: To perform descriptive statistical analyses to characterize MSD injuries in hotel

room cleaners by demographic (age, gender), injury (nature, cause, cause of injury

category, part of body, part of body category) and work (tenure) variables.

Aim 3B: To perform three descriptive statistical analyses to characterize hotel room

cleaner MSD claims stratified into seven broad task categories by injury characteristics:

1) nature of injury, 2) cause of injury and 3) part of body.

In carrying out the dissertation aims, different elements of the three axes of the Unified Model

were applied: ‘T’ for time elements; ‘Q’ for epidemiologic elements (disease concepts); and ‘P’

for Public Health Approach elements (systematic science-based methodology). The integration of

my aims with the three axes is summarized below:

Aim 1: Identification of hotel room cleaner occupation in California workers’ compensation

MSD claims data.

Unified Model Element: Axis P - Surveillance, Axis Q – Host.

Aim 2: Identification of task at time of injury.

Unified Model Element: Axis P - Risk Factor Definition, Axis Q - Vector, Axis T - Event.

Aim 3: Descriptive statistical analyses to characterize MSD injuries by nature of injury (e.g.,

sprain or strain), cause of injury (e.g., lifting or pushing or pulling), part of body injured, age, job

tenure and task at time of injury.

Unified Model Element: Axis P - Risk Factor Definition, Axis Q - Host, Axis Q -Vector, Axis T -

Event.

Dissertation Sections

Section 1 introduces the public health issue of concern, namely, musculoskeletal disorder injuries

occurring to hotel room cleaners. Background is provided that contributes to theories of

causation of the MSDs by describing the changes in the hotel industry’s guest room

accommodations, resultant intensification of hotel room cleaning workload and associated

prevalence of workplace pain and injuries. This section also puts forth a framework for the use

of public health research approaches for injury control to prevent hotel room cleaner

musculoskeletal disorders in light of research gaps. The potential of workers’ compensation data

to address those data gaps is demonstrated.

Section 2 builds on the literature referenced in the introduction section to support the aims,

research questions and selection of workers’ compensation data for my dissertation research. In

this section, a thorough review of the literature is provided that demonstrates gaps in

occupational injury surveillance systems that fail to identify injured hotel room cleaners, the lack

of scientific evidence about interventions to prevent MSD injuries occurring to hotel room

cleaners and the inaccessibility of narrative text data in workers’ compensation claim databases

for injury and task information. The literature review critiques what data is needed and argues for

innovative research methods to bridge existing data gaps.

Section 3 describes the methods used to identify the occupation of hotel room cleaner (Aim 1)

and the task at time of injury (Aim 2) from narrative text data. This section identifies the

challenges of testing and modifying existing methods and creating new ones with innovative

artificial intelligence techniques of natural language processing and text classification models.

Section 4 explores the MSD data by performing descriptive analyses (Aim 3) in two parts. The

first part (Aim 3A) describes the MSD injuries by the following variables: demographics (age,

gender), injury characteristics (nature, cause, cause of injury category, part of body, part of body

category) and work descriptors (job tenure at time of injury, hotel room cleaner job title). The

second part (Aim 3B) stratifies the MSD claims into seven task at time of injury categories and

describes the data by the following injury characteristics: 1) nature of injury, 2) cause of injury

(lifting, carrying, twisting, etc.) and 3) part of body.

In Section 5, the study findings are discussed along with the study’s strengths and limitations.

The public health significance is presented. Conclusions based on a critical review of the study’s

aims and methodology are offered. Policy implications and recommendations for areas of future

research are put forth. Potential worksites interventions are suggested based on study findings.

Section 2: Literature Review

Utility of Workers’ Compensation Data for Identifying Occupation and Task for Injury

Analysis

Numerous studies using WC data succeed in identifying the leading types of injuries (nature of

injury), the factors contributing to the cause of injury and parts of body affected stratified by

industry; by occupation; by injury type at the state level, by part of body and occupation, for

example 25% of WC claims by welders concerned eye injuries and helpers/laborers had eye

injury rates twice that of workers overall; and by injury type and occupation.61-68 WC data can

identify categories of workers not easily accessible in administrative datasets due to the nature of

their employment. Multiple approaches using California WCIS data identified residential

workers employed by private households by identifying tasks associated with day laborer work,

e.g., maintenance and landscaping and domestic work, e.g., housekeeping and childcare.69

Findings from unpublished studies explored methods of defining MSDs and occupation using

WCIS data for hotel housekeepers and analyzing by task.11,47 Leading cause of injuries were

strain or injury, NOC; pushing or pulling; lifting; and repetitive motion, with the low back area

and shoulders the leading parts of body injured.11

Workers’ compensation claims data can provide layers of information that, with further research,

will uncover high-risk industries, operations, and occupations, along with corresponding hazards,

injuries and parts of body affected. These findings can then lead to the discovery of high-risk

tasks. An investigation of agriculture workers using four years of workers’ compensation data

from Colorado identified that the largest proportion of injury claims were due to livestock-

worker interaction.70 For these claims, the leading nature of injury was contusions occurring to

the upper and lower extremities. Douphrate and co-authors followed up these findings with a

study focusing on livestock-handling injuries only using ten years of WC data.71 Study results

found dairy farms having the second highest injury claim rate compared to the two other

employers studied, “cattle dealers and cattle/livestock raisers”. Narrative text injury descriptions

from WC claims detected that 48% of the animal-related injuries occurred in the dairy parlor. By

identifying the location of where the injuries occurred on a dairy farm (dairy parlor), then the job

task (milking) was identified. In the proceedings from a 2012 NIOSH-sponsored conference on

the research applications of workers’ compensation data, conference facilitators concluded that

“best practices for the use of workers’ compensation data to identify intervention needs could be

developed”.72 One underlying assumption is, having reached the threshold of a workers’

compensation claim, these injuries resulted from high-risk tasks. This assumption is supported by

studies that found an increased likelihood of workers filing a workers’ compensation claim is

associated with: work-relatedness confirmed by a physician, involved surgery and/or days away

from work; high levels of pain and high physical job demands; and restrictions on physical

activity, decreased health status, days away from work equal to or greater than a week.36,51,52

Implementing recommended interventions based on these cases is a priority.

Task Analysis: Pathway to Interventions

Task analysis has identified potential intervention points for varying industries and occupations.

Analysis of narrative text data from WC claims of hospital workers categorized patient handling

injuries into seven subtasks with 40% occurring from either lifting or transferring the patient and

32% from repositioning, pulling or catching the patient.73 By pinpointing the task and the

prevalence of injuries associated with that task, researchers identified interventions appropriate

for the task and prioritized them by greatest need.

Task analysis can be used more readily to identify the maximum workload of an overall task or

the workload of a single task.74 Task analysis, triggered by research involving workers’

compensation claims, can evolve over time with each new study building on the findings of

earlier studies and thereby creating a trajectory of data analysis, hazard assessment, intervention

development and evaluation, program implementation and potentially, policy initiatives. This

trajectory echoes the theoretical framework of my dissertation, the Unified Model, that ranges

from “surveillance to risk factor identification, followed by intervention evaluation and program

implementation.”44

Example of the Unified Model Put into Practice: Dairy Parlor Milking Task Research Further

research by Douphrate and colleagues exemplifies the above trajectory and is presented here as

an example. Increasing trends in dairy farming towards large herd operations in the U.S. resulted

in the industrialization of parlor milking systems and milking practices, raising concerns about

ergonomic hazards and the lack of solutions beyond stop-gap measures, e.g., lighter equipment,

rubber mats.75,76 Research identified high-risk milking practices affecting the upper extremities,

that were then associated with differences in the configuration of large herd dairy parlors.77,78

Symptoms of work-related MSDs in the upper extremities (shoulders and wrists/hands) were

found for 80% of workers studied in large-herd dairy parlors and were associated with working

conditions, primarily static postures, repetitive tasks, lack of breaks, extended reaches and

working while injured.79 In the course of developing milking task interventions, five main

milking tasks were identified, with attachment of the milking unit to the cow’s udder as the most

physically demanding.80 Study authors focused on the hazards of the milking unit (equipment)

and identified several features e.g., weight, spread of the milk tubes, shape of the teat cup, that

suggest possible varying levels of corresponding muscle activity. The authors call for design

changes, identified with the help of worker input, to reduce worker muscle fatigue, noting that

for milking tasks that were highly repetitive, small reductions in muscle loading over time could

delay or prevent the progression of MSD injuries.

The lack of occupational safety and health regulations targeting the agriculture industry in

general and dairy farming in particular contribute to the existence of unsafe working conditions

in the dairy industry.81 A review of sixteen intervention studies in the dairy industry found that

none were of sufficient sample size to evaluate their impact on injury rates.82 The researchers

propose strict evaluation methods be included in study design, beyond measuring knowledge

gained. Also, to improve safety climate, transformational leadership is needed that integrates

worker health and safety in dairy management’s plans along with other goals, e.g., production.83

This research trajectory serves as an example for other industries about the potential of WC data

to trigger intervention studies.

Task Analysis of Hotel Room Cleaner Occupation

Ergonomic task analysis involves “… defining the problem, analyzing problem jobs and

implementing ergonomic solutions”.84 Identifying work areas and jobs, compiling list of job

duties or job description and determining job exposures is a first step as part of analyzing jobs

before selecting the method for task analysis and performing the task assessment.84

Biomechanical evaluations using the Lumbar Motion Monitor (LMM) of hotel room cleaning

work, referred to earlier, generated broad task categories and corresponding specific subtasks

(Appendix A) as part of enumerating risk levels for the likelihood that a task results in a

workrelated low back disorder. Hotel room cleaner task lists serve as a framework for

intervention points for reduction of injury risk.9 Allread et al’s three main task categories are

“Bedmaking”

(seven subtasks), “Bathroom cleaning (six subtasks)” and “Other” which includes vacuuming,

dusting, replenishing from cart and checking furniture and drawers. Based on three hotel room

cleaning evaluations at different hotel properties, there was no single subtask considered in the

“low-risk” category for a low back disorder. The results of all three evaluations for the hotel

room cleaning job overall were in the “high-risk” category. Neidhardt manually reviewed 2,000

WCIS claims after excluding 665 (25%) due to lack of job task information associated with the

injury. Neidhardt used these task categories for WCIS data analysis: “Cleaning bathrooms,

shower-tub, mopping floors”; “Making bed, lifting mattress, moving-rearranging bed”;

“Vacuuming, spot cleaning floor, room cleaning”; Pushing or pulling cart, sudden stop of cart, re-

stocking cart”; “Injuries associated with stairs, steps, ladders”; “Other”. The leading four tasks at

time of injury were bedmaking, bathroom cleaning, stairs/steps/ladders and pushing or pulling

cart, respectively. For each task categories, there were 4-5 subtasks. For example, pushing or

pulling was the task associated with the injury for “vacuuming”; pushing on carpet for “operating

linen cart”; bending to clean for “cleaning tub or shower”; pulling changing linen for “handling

linen”; and tucking under or lifting mattress for “handling a mattress”. These findings are

supported by Allread et al’s LMM evaluations that identify lifting rate and forward bending as

key risk factors for low back injuries in hotel room cleaners.9 Allread et al found the following

hotel room cleaning tasks to be “high-risk”: bedmaking, vacuuming, bathroom cleaning and

dusting.9 A peer-reviewed study echoes the above findings and to date, is the only one that

associates specific hotel room cleaning tasks with self-reported hotel room cleaner ergonomic

problems.6 Survey results by Krause et al found the following leading concerns .. “linen cart too

heavy (84%), heavy bedspreads, or comforters on beds (74%)” and …“vacuum cleaner too heavy

(62%) and vacuum cleaner needs repair (62%).”

Interventions exist for musculoskeletal injury risk factors of bathroom cleaning with scientific

findings from over fifteen years ago. Researchers found that a 21-inch long handled cleaning tool

provided a more neutral work posture for the wrist, shoulders and trunk compared to using a

sponge.48 In addition, awkward postures of the lower extremities were reduced: there was less

climbing into bathtubs and less kneeling to clean bathtubs. This resulted in safer foot placement

and less wear and tear on the knees. A Cal/OSHA publication also released in 2005, “Working

Safer and Easier for Janitors, Custodians and Housekeepers,” is an eighty-six-page manual of

valuable information, illustrating tools and safe work practices for cleaners using ergonomics for

injury prevention.85 These include fitted sheets, long-handled tools and safe work practices for

moving furniture and lifting mattresses. To date, only one intervention study exists

recommending bedmaking aids and fitted sheets which together were found to reduce lifting by

48%.45

With limited injury data on the task, cause of injury and part of body injured specific to

performing hotel room cleaning work, it is less likely that employers will invest in programs

applying existing interventions or exploring the development of new ones when there is a lack of

administrative data documenting the outcome of work-related job hazards, namely occupational

injuries. Regulations can help employers focus separate job tasks, identify risk factors and

develop programs to put into practice solutions such as worksite interventions. Hotel room

cleaner tasks are defined in Cal/OSHA 3345 Hotel Housekeeping Musculoskeletal Injury

Prevention standard (Appendix B) as “… [those]…related to cleaning and maintaining sleeping

room accommodations”… “include, but are not limited to, the following: (1) sweeping, dusting,

scrubbing, mopping and polishing of floors, tubs, showers, sinks, mirrors, walls, fixtures, and

other surfaces; (2) making beds; (3) vacuuming; (4) loading, unloading, pushing, and pulling

linen carts; (5) removing and supplying linen and other supplies in the rooms; (6) collecting and

disposing of trash; and (7) moving furniture”. California’s promulgation of this regulation, the

only state to do so, requires employers to implement an MSD injury control plan for hotel room

cleaners defined as “…. an employee who performs housekeeping tasks and may include

employees referred to as housekeepers, guest room attendants, room cleaners, maids, and

housepersons.” Plan elements include performance of a worksite evaluation, injury investigation,

identification of solutions, worker and supervisor training, and the active participation of workers

and their union representatives.1

The Potential of and Challenges to Using Narrative Text Data

Similar to the various uses for WC data for occupational injury research targeting occupation and

task, studies using narrative text data vary in their purpose, e.g., text extraction, validation of

coded data; by approach/technique; and by method of quality control assessment, with complete

information detailing the latter two elements lacking.86 In a review article of 28 peer-reviewed

papers, among the authors’ conclusions, use of systematic methods to search narrative text data

provided important information not previously accessible in coded data and computerized

methods can be effective to classify less complex data into categories, saving complicated text

for manual methods.86 Identification of tractor overturns as a cause of injury for tractor fatalities

increased from 49% using standardized surveillance codes to 54% when incorporating narrative

text data as part of Kentucky Fatality Assessment and Control Evaluation (FACE)

investigations.87 By providing more comprehensive information than the standardized variables

in the FACE dataset, a logistic regression analysis of the narrative text data found increased odds

of workers “dead at the scene” when driving a tractor in “muddy conditions” (OR: 9.1, p<0.01)

and driving a tractor “with a front loader attached” (OR: 5.8, p<0.02). These findings detected

new contributing factors and were added as variables in the FACE program. The limitations of

the data include injury scenarios with sample size too small to be included due to lack of power

for statistical analyses. Combining pre-coded data (gender, industry) with results from coding

injury narratives in WC claims allowed for identifying electrical injuries not commonly

associated with certain industries with electrical hazards, e.g., services (33.4%) and with the

female workforce (27.7%), drawing attention to a probable lack of safety trainings for at risk

workers.88 These same injury narratives provided robust findings about the cause of electrical

injuries when the authors applied case-selection algorithms using pre-coded cause of injury data

variables and a tailored six-category taxonomy to the “accident” and injury narrative

descriptions. The taxonomy included activity, source, initiating process, mechanism of injury,

vector and voltage, providing new insights into non-fatal electrical injuries such as 55% occurred

during manual tasks and 24% had appliances and office equipment as the leading source, with

the exception that no information on vector was found.

Yamin et al (2016) created a systematic multi-level classification procedure to decide to which of

three categories a claim belonged using de-identified WC data (n = 4,268) from metal fabrication

employers.89 Study researchers removed claims with MSDs and slips /trips/falls using a keyword

auto-coding program, with “all other injuries” the only claims remaining. This category

accounted for 68% of all claims. The remaining claims were classified by two coders into three

categories: machine-related, non-machine related and possible machine-related. After achieving

83% agreement on the initial coding, the remaining claims were resolved jointly and the potential

machine-related claims were excluded along with the non-machine claims. Using job tasks from

previous hazard evaluations, eight task categories were defined and narrative text descriptions

were manually classified, with 36% (n = 1,053) being machine-related and the rest of the injuries

occurring during secondary tasks. The authors identified that serious injuries, e.g., amputations,

fractures, severe lacerations occurred at various points in the work process, not in only one task,

pointing to a need for broader injury control efforts.

Researchers found the need for guidelines for standardizing text entered into injury descriptions

in order to accurately classify the data.86 Emergency room reports had a high occurrence of word

variation belonging to two groups: 1) spelling errors (11-18.8%), including multiple errors in

each instance and 2) word truncations, abbreviations, acronyms (>20%), with varying success of

different text normalization methods.90 Similar word variation issues were found in other studies

mentioned earlier.87,88

Researchers utilized and/or created various methods to improve or ensure the validity of study

results. Lack of sufficient information to identify underlying causes was a limitation of WC data

when used alone, but in conjunction with pre-coded data and a coding taxonomy tailored to the

study objectives, e.g., etiology of injuries, WC data provides useable information.88,89 Using

California WC data, Riley and Mujano created an index measuring the researchers’ confidence

level on employment status of the claimant.69 The index counted the number of sources in the

WC data that indicated the claimant was a residential worker including the narrative text fields of

Occupation and Injury Incident Description fields, finding 14.3% of claims with a high

confidence level and 85.4% with moderate level. Using NIOSH’s Industry and Occupation

Computerized Coding System computerized auto-coder to add 2010 Census Occupation Codes,

numerous occupations at risk for carpal tunnel syndrome were able to be identified using CA

WCIS data.91

Application of AI/Machine Learning Methods to Public Health Research

Semi-automated review, human and machine-combined and natural language processing (NLP)

techniques are increasingly utilized for public health research.60 Occupation injury research leads

the field, accounting for 46% of papers summarized in a systematic review of machine learning

processing methods (MLP).92 Vallmur points to the ability of MLP methods to accurately classify

text data into broad categories, to identify injury patterns and causal mechanisms. A semi-

automated method classified WC narrative text data for 15,000 claims by BLS event codes:

68% using Naïve algorithms only and 32% manually for more challenging narrative data.

Maximum accuracy rates for the classified data was 87%.93 These same researchers in a

following study proposed human-machine pairing as a best practice for filtering large

administrative datasets like workers’ compensation claims.94 An important first steps is to have

multiple coders of the narrative text data using a statistical scoring system to resolve

differences.60 Authors of this study recommend computerized methods applied to data where

there is a high confidence level that the data is classified correctly and to enhance computerized

methods with manual coding. For narrative data classified by the model, a best practice resulting

in a relatively high level of accuracy, is to manually code the classified data with the 30% lowest

predictive probabilities and perform computerized coding for the remaining 70% of the data.94

Also, study authors found that algorithms succeeded in coding repetitive, simple narrative text

consistently in a systematic manner. The use of training data and multiple iterations to fine-tune

the model is advised.93 Natural language processing techniques can be combined with human

coding to: verify the accuracy of manually coded data or to classify the data and then use the

NLP findings to generate codes.60 The quality of the source data is key since the model and

algorithms are dependent on the data, impacting the generalizability of applying the model.92-94

Vallmur pointed to data consisting of small sample sizes or specific content areas as limits to the

generalizability for application to other populations as the data influences the models ability to

recognize patterns and those patterns reflect the study population only.92

Section 3: Methods

Study Design

This study is a quantitative retrospective record review of administrative data from the California

Workers’ Compensation Information System (WCIS) that describes the frequency of

musculoskeletal disorders (MSDs) among hotel room cleaner claimants and assesses the

association between type and frequency of MSDs and the job task performed at time of injury. In

addition, artificial intelligence (AI) methods were applied to identify narrative text data extracted

from the WCIS dataset, demonstrating the feasibility of utilizing natural language processing

(NLP) techniques to transform narrative text data into coded data suitable for quantitative data

analysis. As a result, two categories of data – occupation and task at time of injury – were joined.

Analysis of this data along with demographic, work, and injury information about MSDs

occurring to hotel room cleaners is now available for public health injury prevention and control

efforts. This information contributes to filling the data gap explained in Section 1 regarding the

lack of published U.S. government occupational injury surveillance data about maid and

housekeeping cleaners employed by the hotel industry. Natural language processing techniques

provided systematic, valid, and replicable methods to filter thousands of WC claims to identify

occupation and task in narrative text data and code the claims swiftly (Appendix I). This proof of

concept study using natural language processing techniques, combined with human coding,

increased the accessibility of narrative text data in injury incident descriptions of administrative

data.60 The methods applied in this dissertation are innovative and contribute towards improving

research methods for the quantitative analysis of workers’ compensation claims. This

combination of NLP techniques and human coding was exemplified in the working relationship

between the AI contractor and myself. Their work was guided by my coding parameters and

algorithms, based on decisions I made in response to challenges in the data, e.g., over two

thousand occupation terms from which to differentiate those pertinent to hotel room cleaner job

duties; categorizing task data when generic terms such as “cleaning” or “regular duties” were the

only task terms included in the injury descriptions; distinguishing between a “task” and an

“activity” in the injury description – whether moving a bed is a “moving furniture” task or an

activity within the “bedmaking” task. These decisions were the result of different methods I

applied (detailed in this section) and my expertise acquired from over fifteen years of combined

experience working for the leading hotel workers’ union in North America. First as a hotel

housekeeping injury researcher (four years) and then as director for health and safety (11+ years)

Working with artificial intelligence (AI) experts, my parameters were incorporated into their

model, based on the limits of the NLP techniques, the data and the study timeline.

Data Source

The California Department of Industrial Relations’ Division of Workers’ Compensation (CA-

DWC) provided the secondary data for this study from its Workers’ Compensation Information

System (WCIS), an administrative database of claims. WCIS data includes coded and narrative

text data about an injury or illness, affected employee (including demographic characteristics)

and employer.

Occupational Injury and Illness Data Information Forms: Sources of WCIS Data Elements

In California, when a work-related injury occurs, employers are required to submit either an

“Employer’s Report of Occupational Injury or Illness” or a “Doctor’s First Report of

Occupational Injury or Illness” (DFR) to its insurance claims processor. The Doctor’s First

Report form (Appendix C) includes information about the employer and details about the

injury/illness incident, physical examination, test findings, diagnosis and treatment).95 The

Employer’s Report of Occupational Injury and Illness form (Appendix D) includes information

about the employer, the employee’s employment history, wages and hours, injury circumstances,

age, gender, and injury severity. Information from either form fills the “First Report of

Occupational Injury and Illness” submitted electronically by the claims administrator to the CA-

DWC’ Workers’ Compensation Information System within ten working days after learning of

the claim.96

Data Elements of the WCIS Database

The First Report of Occupational Injury and Illness (FROI) completed by the employer’s

insurance claim administrator described above comprises the data elements in the WCIS

database. Data is coded using the terminology, rules and codes from the Workers’ Compensation

Insurance Organization (WCIO) found in Appendix H.

Data Collection

Data Extraction from California’s Division of Workers’ Compensation

The California Division of Workers’ Compensation (CA-DWC) provided a custom data

extraction based on my data request which was prepared with input from subject matter experts,

including those from the California Department of Industrial Relations (CA-DIR) and the

CADWC.† I also submitted a data matrix that listed the research justification and corresponding

study aim for each of the requested sixty-five data elements (Vossenas Dissertation WCIS Data

Matrix, Appendix E). A Memorandum of Understanding was signed between the City University

of New York on behalf of the Graduate School of Public Health and Health Policy and

California’s Division of Workers’ Compensation. The last WCIS data file was transmitted

electronically from CA-DWC on 6/27/19. The City University of New York University

Integrated Institutional Review Board approved this study on February 11, 2019 (Appendix F).

Data Sample

WCIS Administrative Data Sample – Hotel Industry Musculoskeletal Disorder (MSD) Claims

This dissertation focuses on coded and narrative text injury and demographic data of California

workers’ compensation claims subsetted from a WCIS administrative data sample for years 2001

through 2016 for hotel industry workers with work-related musculoskeletal disorder injuries.

Sixteen years of data were requested to provide a sample size large enough to generate a

sufficient number of records to: 1) identify the occupation of hotel room cleaner, recognizing that

this job title accounts for about 23% of the hotel workforce; 2) identify the task at time of injury

using seven task categories, recognizing that previous research coded task as “unknown” for

33% of WCIS cases studied, and 3) provide disability and medical billing data associated with

the claims in the study dataset, recognizing that only 47% of claims are estimated to have

medical billing data.47 Study years were limited to begin in 2001 as the WCIS database was

created in mid-2000.

Based on discussions with California workers’ compensation subject matter experts, the

following First Report of Occupational Injury (FROI) data elements were considered to be highly

populated and were included in the data extraction: date of injury, nature of injury, part of body

injured, cause of injury, accident description/cause, date of hire, and gender. These data elements,

along with “Occupation,” “Date of Birth” and “Date of Return to Work” comprised the essential

data elements of the study dataset.

Hotel Industry Definition

All claims in the study dataset were for individuals employed in occupations in the

hotel/motel/casino industry referred to throughout this dissertation by the all-encompassing term

“hotel industry.” The hotel industry was defined using industry source codes: two Workers'

Compensation Insurance Rating Bureau (WCIRB) Class codes, ten Standard Industrial

Classification (SIC) codes and four North American Industrial Classification System (NAICS)

codes (Appendix G). No information about employers or worksites were requested.

Musculoskeletal Disorder Case Definition

The CA-DWC used a combination of Worker Compensation Insurance Organizations (WCIO)

codes for the case definition for an MSD injury (Appendix H). I provided this definition in the

data application to the CA-DWC and in subsequent communications with the agency’s data

programming staff. This definition follows methods used with WCIS data for hotel room cleaner

MSD injuries by Cohen et al and replicated in a study about patient handling MSD injuries of

healthcare workers.11

The MSD injury case definition requires that a claim includes: one of the specified nature of

injury subcategories, plus one of the specified cause of injury subcategories, plus one of the

specified parts of body injured subcategories. Every claim included in the WCIS sample contains

at least one specified subcategory of each of these three variables in Figure 3.1.

Figure 3.1 Injury Variables that Comprise the Definition of an MSD Injury

Cause of Injury Part of Body

Variable:+ + Injured

Variable: Specified

Subcategories Specified

Subcategories

The specified subcategories by type of variable are listed below:

1. Nature of Injury Specified Subcategories and Codes. Dislocation (16); Hernia (34);

Inflammation (37); Sprain (49); Strain (52); All other specific injuries, NOC (59); All Other

Occupational Disease Injury, NOC (71); VDT-related Disease (76); Carpal Tunnel Syndrome

(78); All Other Cumulative Injury, NOC (80); Multiple Physical Injury Only (90); Multiple

Injuries Including Both Physical and Psychological (91). The definitions of the nature of injury

categories are available via a link in Appendix H. For example, “Sprain (49)” is defined as

“49. Sprain or Tear Internal derangement, a trauma or wrenching of a joint, producing pain and

disability depending upon degree of injury to ligaments” and “Strain (52)” is defined as “52.

Strain or Tear Internal derangement, the trauma to the muscle or the musculotendinous unit from

Nature of Injury

Variable:

Specified

Subcategories

violent contraction or excessive forcible stretch. 53. Syncope Swooning, fainting, passing out, no

other injury” (Nature of Injury, Appendix H).

2. Cause of Injury Specified Subcategories and Codes. Strain or Injury By: Twisting (53),

Jumping (54), Holding or Carrying (55), Lifting (56), Pushing or Pulling (57), Reaching (58),

Using Tool or Machinery (59), Strain or Injury by, NOC (60), Wielding or Throwing (61),

Repetitive Motion – Carpal Tunnel Syndrome (97); Striking Against or Stepping On: Sanding,

Scraping, Cleaning Operations (67); Rubbed or Abraded by: Repetitive Motion (94), Rubbed or

Abraded, NOC (95); and Miscellaneous Causes: Cumulative, NOC (98) (Cause of Injury,

Appendix H).

3. Part of Body Injured Specified Subcategories and Codes. Neck: Codes 20-23, 25; Upper

Extremities: Codes 30-39 (all codes); Trunk: Codes 40-47 (all codes), 61-63 (all codes); Lower

Extremities: Codes 50-58 (all codes); Multiple Body Parts: Insufficient Info to Properly Identify

- Unclassified (65), Multiple body parts (90), and Body systems and Multiple Body Systems

(91) (Part of Body Injured, Appendix H)

For example, a room cleaner lifts a mattress and feels a strain in her upper arm. This observation

would meet the definition of an MSD injury because it includes: a strain (nature of injury/

subcategory “strain”) + lifting (cause of injury/subcategory “lifting”) + upper arm (part of body

injured/subcategory “Upper Extremities”). A strain caused by lifting and felt in the head would

not be included because the head is not one of the specified part of body injured subcategories.

Study Dataset

Setting Parameters

There were 79,052 claims in the Hotel Industry MSD dataset for years 2001-2016 provided by

CA-DWC (Figure 3.2). Based on the results of frequency analyses of the data using SAS®

software, Version 9.4, the data were limited to certain factors. To evaluate gender segregation in

the data, the occupation terms “Housekeeper” and “Room Attendant” accounted for the majority

of claims with occupation terms similar to hotel room cleaning job titles. The proportion of these

two occupations terms that had female claimants were averaged together, resulting in a finding

that 96% of claimants were females. With too few claims with male claimants for a comparison

analysis, the dataset was limited to females. “California” as Employee State accounted for 99.2%

of claims; the data was limited to California residents. The dataset was limited to years 20072016

(n=33,303 claims) as this was when the distribution of claims was the highest (70%).

Validation and Cleaning the Data

Standard methods of data cleaning were performed such as data validation and removing data

that did not meet study inclusion criteria. Timestamps were removed from data fields and

variables were renamed by putting “Clean” in front of the variable name, e.g., “CleanDOB.”

Industry was confirmed using two separate variables from the WCIS dataset: Industry Code and

Class Code (Data Element Nos. 25 and 59, respectively, Vossenas Data Matrix (Appendix E). A

total of 10,303 claims that did not meet the criteria for hotel industry and were not coded as

missing, were excluded (only meets Class Code criteria, not a select Industry Code or select

Missing Industry: n=5,714; only meets Industry Code criteria, not a select Class Code or select

Missing Class: n=4,585; neither meets an Industry nor Class Code nor a Missing Industry or

Class Code: n=4). For the remaining 23,070 claims, they were classified as either: 1) confirmed

as hotel industry if both Class and Industry Code variables met the criteria (n=19,652), 2) where

the Class Code criteria was met and the Industry Code was missing (n=2,926); and 3) where the

Industry Code criteria was met and the Class Code was missing (n=492). This recoding permits

the stratification of the data by level of hotel industry verification (one hotel industry code or

two) and the identification of claims where one or both hotel industry codes were missing. In

addition, this recoding allows for performing an analysis to assess the impact of hotel industry

verification level on the accuracy of identifying hotel room cleaners in the occupation data (Aim

1) and/or identifying hotel room cleaner tasks in the task data (Aim 2).

To verify that all claims met the definition of an MSD injury case, frequencies were run on the

Nature of Injury, Cause of Injury and Part of Body Injured variables. All claims met the MSD

injury case criteria. A new variable “Study ID” was created. Study Record numbers, Jurisdiction

Claim Numbers (JCN) and Study ID were checked for duplicates. A file was created matching

the new Study ID with the corresponding Study Record and JCN variables, followed by

removing the latter two variables from the dataset.

De-identification of the Study Dataset: Numeric Data

The claims data provided by CA-DWC were not de-identified. I followed the methods described

in the Safe Harbor §164.514(b)(2) clause of the Health Insurance Portability and Accountability

Act (HIPAA) Privacy Rule and de-identified the study dataset for zipcode, city, date, gender and

age variables and for narrative text data.97 The “Clean” date-related variables were used to

generate new date-related study variables: Age Time of Injury, Years Job Tenure Time of Injury,

and Months Job Tenure Time of Injury. Other Clean date-related variables were renamed as the

following new study variables: Year Hire, Year Injury, Year Death, Year Disability Benefit, Year

Last Worked, and Year Return to Work. Subsequently, all original and Clean date-related

variables were removed from the dataset. For claimants with an “Age Time of Injury” response

over 89 years, data were recoded into the response category of “≥90 years.” The variable

“Gender Code” was removed from the dataset as all claimants in the study dataset are female.

“Employee Postal Code” was truncated to the first three digits, creating the new variable

“Zipcode.” The Employee Postal Code and “Employee City” variables were removed.

Study Dataset Samples

The study dataset was subsetted into several sequential data samples as illustrated in Figure 3.2,

to ultimately create two analytic data samples: “Hotel Room Cleaner Occupation MSD” and

“Hotel Room Cleaner Task at Time of Injury.” These two study samples correspond to Aims 1

and 2, respectively. Both aims are concerned with methods to code narrative text data and are the

outcome of the application of list maintenance and natural language processing methods by the

AI experts. After removing non-hotel industry claimants from an earlier iteration, the Female

Hotel Industry MSD Claimant Sample was created for years 2007-2016 (n=23,070 claims).

From this data sample, the AI experts applied a “list maintenance” method that resulted in

achieving Aim 1, the identification of the occupation of hotel room cleaner in hotel industry

MSD workers’ compensation claims data. As a result, the “Female Hotel Room Cleaner

Occupation MSD Claims” analytic data sample (n=12,125 claims) was created.

To achieve Aim 2, the identification of the task at time of injury, I subsetted the Aim 1 analytic

data sample to years 2012-2016 (n=6,952 claims), resulting in the “5-Year Female Hotel Room

Cleaner MSD Claims” data sample. The most recent five years of data were selected to reflect

more current changes in hotel accommodations and hotel housekeeping work practices.

De-Identification of the Task Data Sample: Narrative Text Data

Before sending the task data sample to the AI experts, I de-identified the narrative text injury

descriptions that comprise the task data. I followed the Safe Harbor HIPPA guidance about

deidentification of free text data which states “The de-identification standard makes no

distinction between data entered into standardized fields and information entered as free text

(i.e., structured and unstructured text) -- an identifier listed in the Safe Harbor standard must be

removed regardless of its location in a record if it is recognizable as an identifier.”97 Efforts were

made to remove the gender terms “she” and “her,” along with names (worker, supervisor, hotel),

hotel room numbers, dates of injury, adjudication case numbers and diagnoses from the task data

used to de-identify the data.

Application of Natural Language Processing Methods to Task Data Sample

The AI experts applied NLP methods to the task data sample and achieved Aim 2, resulting in the

“Female Hotel Room Cleaner Task at Time of Injury, Years 2012-2016” analytic data sample.

After further refining the parameters of the coded task data sample (restrict claims to

Hotel Room Cleaner =1 “Yes,” n = 6,952; remove the “Unknowns,” n = 3,013; remove claims

with a “<70% confidence level,” n = 1,310), the sample size is reduced to 2,629 claims (includes

1 missing injury incident data [task]). As a result of achieving Aims 1 and 2, the analytic data

samples for Aims 3A and 3B are created.

Figure

3.2

Flowchart

Steps

Construct

Hotel

Room

Cleaner Occupation

MSD

and Task

Time

of Injury

Analytic

Data

Samples

Exclude

males

JCA-WCIS

Administrative

Data

Sample|

Bah

neces

Hotel

Industry

MSD

Cums

years

2001-2006

2001-2016

945679

n=79.052

Exclude:

Female

Hotel

Industry

MSD

Claims

‘non-botel

industry

claims

‘Gndustry

not

verified)

910303,

2007-2016

33373

enero

(industry

verified)

eaten

——

10,987

See

aac

Res

ee,

ahi

wee

claims

with

task

coded

3.013

Sire

cit

aoe

—

ESSE

eae

Aim

Analytic

Data

Sample

Description of Artificial Intelligence Methods for Classifying Narrative Text Data

Wiseyak Inc. is a healthcare technology firm located in Bellevue, WA and Kathmandu, Nepal

that uses artificial intelligence (AI) and natural language processing)-based solutions for

healthcare data challenges. Wiseyak Inc. was hired to create AI solutions using NLP techniques

to classify and code narrative text fields of thousands of claims by occupation and by task. I

worked with experts from Wiseyak starting in 2019, some of whom who were based in

Kathmandu, Nepal. We collaborated via email, Skype and Google Meet. Wiseyak Inc. is referred

to as the “AI team” throughout my dissertation.

The AI team applied two categories of methods to the narrative text variables (Figure 3.3) that

originated from the CA-WCIS dataset: “Occupation_Descr” and “Accident Description/Cause.”

These methods transform narrative text data into quantitative data, adding three new study

variables suitable for data analysis: “Hotel Room Cleaner Occupation” and “Hotel Room Cleaner

Job Title” (Aim 1, both variables) and “Task at Time of Injury” (Aim 2).

Aim 1: Application of List Maintenance Method to Occupation Description Narrative Text Data

The AI team used a “list maintenance” method to identify occupation terms from the occupation

description narrative text data found in the “Occupation_Desc” variable. This method classified

the occupation terms, using the “Occupation Terms List” and corresponding response codes for

the newly created “Hotel Room Cleaner Occupation” variable (answering the question, Is the

claimant a Hotel Room Cleaner?). Although the AI team applied the list maintenance method, it

is not considered an “AI” or “NLP” method. The same method was applied to classify the

occupation terms by a second list, the “Hotel Room Cleaner Job Title List” and corresponding

response codes for the newly created “Hotel Room Cleaner Job Title” variable (claimant’s

category of hotel room cleaning job title, e.g., maid, room attendant, housekeeper). The

occupation terms ranged from one to five words each.

I provided the AI team the lists of occupation terms and response codes for the two occupation

variables. According to Wiseyak, a fixed number of unique data with numerous misspellings of

the occupation terms made using Language Processing methods impractical.† The occupation

description terms were usually limited to one to three terms. Examples of spelling variations

and/or errors in the occupation terms include, for example: Room Attendant, Room Attendent,

Room Attendand, Room Attendance, and Rm Attendant. The above details and more are

provided from the AI team in a report by Shankar Poudel included in Appendix I.

Figure

3.3

Flowchart of

Methods

Transform

Narrative

Text

Data

into

Quantitative

Data

Machine

Learning

Processing

Methods

Tana

CAC

Sec

seule

Toate

ona

oer

Testa

oer

(reomot_)

cesisn

coi

car

eva

Canpog

Noe

—s

deity

tithe

acon

eae

conten

See

Fine-Tune

Model

ee,

Hotel

Room

Cleanse

Sob

Tales

TppiyCatricd

Citar

once

Tee

Classify

soo‘

Terms

clase

‘New

Variable:

“Task

Time

Injury

Aim 2: Application of Machine Learning Processing Methods to Injury Description Narrative

Text Data

Unlike the occupation description narrative text data, the injury description narrative text data

consist of several words or a couple of sentences, making it possible to use Machine Learning

Processing (MLP) methods to classify this data. Several MLP methods were applied, each with

its own distinctive utility and role (Table 3.1). The objective of the AI methods is to create a

language model that generates a key vocabulary of terms from the injury descriptions. Then a

classification model uses this vocabulary to classify the injury description narrative text data into

coded data. The utility of the MLP methods are discussed in the literature, in particular, in

articles recommended by the AI team about the use of Support Vector Machines (SVMs) for text

classification and Term Frequency-Inverse Document Frequency (TF-ID) to assess the

importance of key words in the data.98-100 More details are provided from the AI team in a report

included in Appendix I.

Table 3.1 Application of Machine Learning Processing Methods to Classify Injury

Description Narrative Text Data (Aim 2), excerpts from Appendix I.

Application of Machine Learning Processing Methods:

Classification of Injury Description Narrative Text Data for

Task at Time of Injury Variable

Methods

Key Features

Pre-concept

The concept at the onset of the classification process to

classify the data into 9 classes (codes). Sub-classes, e.g.,

making a sofa bed comes under bedmaking task (class “2”),

are classified later in “layering” phase post-application of

the classification model.

Pre-processing

Prepare the text to maximize its significance for the

vocabulary that generates the language model. Remove

“stop” words such as “the” and “a” and apply “stemming”

to reduce words like “pushing” and “pushes” to “push.”

Language Model

“Term Frequency-Inverse Document Frequency” vectorizer

is used to calculate the importance of each word for

classification in the training dataset. Then “high

importance” words were filtered to create vocabulary that is

used to classify each injury description in the model.

Classification Model

The one vs. rest multiclass linear SVM algorithm is used to

create a ML model as it produced results similar to those

generated by complex algorithms, with the added

convenience of being able to use “Calibrated Classifier.”†

Fine-tuning Classification Model for

Improved Accuracy

Methods applied to enhance accuracy result in small

improvements: 1) regularization (c-parameter), 2) kernel,

3) gamma parameter, 4) margin.

Calibrated Classifier

Application of “Calibrated Classifier” produces a

confidence level (probability) for each class, namely that

the model accurately predicted which class the injury

description belongs to.

Application of Subclass Layers

Sub-classes are created now that the data has Class and

Confidence Levels, e.g., making a sofa bed changes from

class “2, Bedmaking” to subclass “21, Making a Sofa Bed”

The sub-class data are now added to rest of the class data.

Study Variables

Hotel Room Cleaner Occupation (New Variable)

Hotel Room Cleaner Occupation is defined as whether or not the claimant is a Hotel Room

Cleaner and is a new variable with the following values: 1) yes; 2) no; 3) maybe; 4) questions;

5) misclassified; 8) missing; 9) unknown.

Hotel Room Cleaner Job Title (New Variable)

Hotel Room Cleaner Job Title is defined as the claimant’s job title within the occupation of Hotel

Room Cleaner and is a new variable with the following values: 1) housekeeper; 2) room

attendant; 3) room cleaner; 4) maid; 5) houseperson; 6) not applicable.

AgeTRUE Time of Injury (New Variable)

AgeTRUE Time of Injury is defined as the claimant’s age at time of injury incident. This is a new

variable created using the data from the Age Time of Injury variable and removing claimants

aged less than 18 years. The Age Time of Injury variable is a new variable created during the de-

identification of data phase. It was created using the SAS INTCK procedure that computed the

time difference in years between Employee Date of Birth and Date of Injury.

Age Group (New Variable)

Age Group is defined as the age group interval in years to which the claimant belongs and is a

new variable with the following values:

18-24

25-34

35-44

45-54

55-64

65-69

70-74

75+

YearsTRUE Job Tenure (New Variable)

YearsTRUE Job Tenure is defined as number of years of service with employer at time of injury

incident. YearsTRUE Job Tenure is a new variable created using data from the

Years_JobTenureTimeInj variable and moving negative and extreme values to missing. The

Years_JobTenureTimeInj variable is a new variable created during the de-identification of data

phase. It was created using the SAS INTCK procedure that computed the time difference in

years between Date of Hire and Date of Injury.

MonthsTRUE Job Tenure (New Variable)

MonthsTRUE Job Tenure is defined as number of months of service with employer at time of

injury incident. MonthsTRUE Job Tenure is a new variable created using data from the

Months_JobTenureTimeInj variable and moving negative and extreme values to missing. The

Months_JobTenureTimeInj variable is a new variable created during the de-identification of data

phase. It was created using the SAS INTCK procedure that computed the time difference in

months between Date of Hire and Date of Injury.

Tenure Group (New Variable)

Tenure Group is defined as the tenure group interval to which the claimant belongs and is a new

variable created with data from the MonthsTRUE Job Tenure variable with the following values:

≤6 months

6 months-1 year

1-2 years

2-5 years

5-10 years

10-20 years

20+ years

Task at Time of Injury (New Variable)

The variable “Task at Time of Injury” identifies the specific job task the hotel room cleaner was

performing when injured. This new variable is based on narrative text data from the “Accident

Description/Cause” variable coded into seven task categories from the Cal/OSHA Title 8. Hotel

Housekeeping Musculoskeletal Prevention standard, Section (b). Definitions: ‘“Housekeeping

tasks” means tasks related to cleaning and maintaining sleeping room accommodations including

bedrooms, bathrooms, kitchens, living rooms, and balconies. Housekeeping tasks include, but are

not limited to, the following: (1) sweeping, dusting, scrubbing, mopping and polishing of floors,

tubs, showers, sinks, mirrors, walls, fixtures, and other surfaces; (2) making beds; (3)

vacuuming; (4) loading, unloading, pushing, and pulling linen carts; (5) removing and supplying

linen and other supplies in the rooms; (6) collecting and disposing of trash; and (7) moving

furniture.”1 These housekeeping task categories were used verbatim as “Task Category” codes.

Three subclassifications were created for emerging hazards for Task Category 2 - Bedmaking:

“21” rollaway; “22” sofa bed; and “23” Murphy bed. Additional codes were created: “88”

missing; “9” unknown for data where task was not determined; “91” cleaning rooms (generic use

of the term “cleaning”); “92” non-hotel room cleaner task; “93” other hotel room cleaning task;

“94” walking; “97” regular duties; “98” pain/injury/worker motion/diagnosis; “99” not

applicable. For data analysis, three task category codes can be collapsed into one “9” unknown

category: “91” cleaning rooms (generic use of the term “cleaning”); “97” regular duties; and

“98” pain/injury/ worker motion/diagnosis. Subject matter experts1 recommended adding

“unknown” as a response category since previous research coded task as unknown for 33% of

WCIS cases studied.47 With “Unknown” as a response category, an analysis of differences

between claims with and without an identified task at time of injury is possible.

Nature of Injury (WCIS Data Element Number (DN) 35)

Nature of injury codes for “specific injury” or “cumulative injury or disease” from the California

WCIS that relate to musculoskeletal disorders were used in descriptive analyses to characterize

the MSD injuries by nature of injury. These subcategories were used by earlier researchers to

identify MSDs in hotel housekeepers and healthcare workers.11 These subcategories include

Dislocation (16), Hernia (34), Inflammation (37), Sprain (49), Strain (52), All Other Specific

Injuries, NOC (59), All Other Occupational Disease Injury, NOC (71), VDT-related disease (76),

Carpal Tunnel Syndrome (78), and Multiple Injuries Including Both Physical and Psychological

(91). All other nature of injury codes were excluded.

Cause of Injury (DN 37)

Cause of injury codes from the California WCIS subcategories that relate to musculoskeletal

disorders were used in descriptive analyses to characterize the MSD injuries by cause of injury.

These subcategories and codes were used by preceding researchers as stated earlier to identify

MSDs in hotel housekeepers and healthcare workers.11 These subcategories include Strain or

Injury by: Codes 53-61, 97; Striking Against or Stepping Down: Sanding, Scraping, Cleaning

Operation (67); Rubbed or Abraded: Repetitive Motion (94), Rubbed or Abraded, NOC (95); and

Miscellaneous Causes: Cumulative, NOC (98). All other cause of injury codes were excluded.

Cause of Injury Category (New Variable)

Cause of Injury Category is defined as the category of injury causes to which the claimant’s

injury was attributed. The Cause of Injury Category is a new variable based on the categories

from the California WICO injury codes found in Appendix H, with the following values: Strain

or Injury By; Striking Against or Stepping Down; Rubbed or Abraded; and Miscellaneous

Causes.

Part of Body Injured (DN 36)

Part of Body (DN 36) codes from the California WCIS subcategories were used in descriptive

analyses to characterize the MSD injuries by the part of body injured. These subcategories and

codes were used by preceding researchers as stated earlier to identify MSDs in hotel

housekeepers and healthcare workers.11 These subcategories include Neck: Codes 20-26; Trunk:

Codes 40-49, 60-63; Upper Extremities: Codes 30-39; Lower Extremities: Codes 50-58; and

Multiple Body Parts: Insufficient Info to Properly Identify-Unclassified (65), Multiple Body

Parts (90), Body Systems and Multiple Body Systems (91). All other part of body codes were

excluded. The Part of Body Variable is comprised of Part of Body values from the included

subcategories.

Part of Body Category (New Variable)

Part of Body Category is defined as the category to which the claimant’s injured part of body

belongs. The Part of Body Category is a new variable based on the categories from the California

WICO injury codes found in Appendix H, with the following values: Neck; Trunk; Upper

Extremities; Lower Extremities; and Multiple Body Parts.

Data Analysis

Application of Methods by the AI Team to Transform Narrative Text Data into Quantitative Data

Aim 1: Identification of Hotel Room Cleaner Occupation

I provided the AI team with an excel spreadsheet containing 23,070 hotel industry MSD claimant

records for years 2007-2016 (all females), a list of codes for response values for the research

question, “Is this a Hotel Room Cleaner?” and a list of codes to classify the hotel room cleaner’s

job title, e.g., housekeeper, maid, etc. The AI team used a list maintenance method described

earlier in this chapter to classify the records by both lists, thereby creating two new variables:

“Hotel Room Cleaner Occupation” and “Hotel Room Cleaner Job Title.” I received a csv file

from the AI team with the same 23,070 records coded for these two new variables.

Aim 2: Identification of Task at Time of Injury

I provided the AI team with an excel spreadsheet containing 8,442 records for years 2012-2016

(all females) that was comprised of records where the Hotel Room Cleaner Occupation was

coded as 1) Yes, 3) Maybe and 4) Questions and a list of codes for response values for the

research question, “What was the Task at Time of Injury?”. The AI team applied Machine

Learning Processing methods described earlier in this chapter (Table 3.1) to classify the injury

description narrative text data. This transformed the narrative text data into quantitative data,

thereby creating the new variable: “Task at Time of Injury.” I received a csv file from the AI

team with the 2012-2016 data coded for this new variable.

Quality Control Sampling of Narrative Text Data Classified into Task Categories

Before analyzing the Aim 3B analytic data sample, I performed quality control checks on the

coded responses for a sample of each of the seven task categories, i.e., Task Category 1:

Cleaning, Task Category 2: Bedmaking, Task Category: 3 Vacuuming, etc. After performing a

quality control check on a 20% sample of Task 1: Cleaning with 1,235 claims and at the

recommendation of doctoral faculty, I decided to create an algorithm a priori to decide on the

sample size for quality control checks based on the number of claims per task category. Below is

the algorithm I devised apriori:

Use a 15% sample for quality control checks for tasks with >100<200 claims;

Use a 10% sample for quality control checks for tasks with >200 claims; and

Use a 20% sample for quality control checks for tasks with <100 claims.

After all task categories were sampled, then the total number of records sampled was used to

enumerate an averaged percent of records sampled for the task dataset overall.

I evaluated each task category on key indicators of accuracy of the classification of task data that

I devised: % claims hotel room cleaner (HRC)-related, % claims classified correct and % claims

with ≥70% confidence level. The AI team recommended ≥70% confidence level as a standard

cut-off point for accuracy of the data, as this is considered a best practice in the literature

(Marucci-Wellman et al., 2017). Having observed accurate coding in the data at levels <70%, I

also evaluated the accuracy of each sample by three coding confidence levels: ≥70%, 60-69%

and <60%. This allowed me to determine if I should consider adjusting the cut-off point for

including task data on a task category basis, based on that task category’s specific results for

accuracy by confidence level category, e.g., task x cut-off is 60-69% because 100% of claims in

the sample at this confidence level were coded correctly.

In addition to the above indicators, I calculated two estimates for each task category: estimated N

of claims with ≥70% confidence level and estimated N of claims with correct task coding at the

≥70% confidence level. The first estimate is calculated by multiplying the percentage classified

correctly by the total number of claims for a specific task category. Using the ≥70% confidence

level as a cut-off point, this first estimate would equal the number of claims in the task dataset

for each task category that would be included for Aim 3B data analysis. The second estimate is

calculated by multiplying the total number of claims from the first estimate by the percentage of

claims with a ≥70% confidence level that are coded correctly. For example, suppose there are

1,000 claims for task x with 80% of those claims are in the ≥70% confidence level category and

the accuracy rate for the ≥70% confidence level category is 90%. The first N would yield 800

claims estimated to be added to the data analysis for task x using the ≥70% confidence level

category as the cut-off inclusion criteria. Of those 800 claims, the second N would yield 720

claims estimated to be correct. After both sets of estimates were calculated for all tasks, the total

number of records estimated to be correctly classified in the ≥70% confidence level category was

used to enumerate an averaged accuracy rate for estimated records overall for the task dataset.

Descriptive Analyses: Demographics, Injuries and Task Analysis (Aims 3A and 3B)

Aim 3: Descriptive Analyses: Study Population, Injury Characteristics and Task at Time of Injury

To perform basic descriptive statistical analyses to characterize MSD injuries in hotel room

cleaners by age at time of injury, age group, years job tenure, months job tenure, tenure group,

nature of injury (e.g., sprain or tears), cause of injury category (e.g., Strain or Injury By), cause

of injury (e.g., twisting), part of body category (e.g., Upper Extremities, Trunk), part of body

(e.g., upper arm, low back area), and task at time of injury.

Aim 3A: Descriptive Analyses: Study Population and Injury Characteristics, Years 2007-2016

Descriptive statistical analyses were performed to characterize the hotel room cleaner study

population for years 2007-2016. The study population was limited to female claimants as

explained earlier in this chapter with a response of Yes=1 for the hotel room cleaner occupation

variable. The total number of hotel room cleaner MSD injury claimants (12,125) was identified

with a breakdown by Hotel Room Cleaner Job Title.

Pearson correlation coefficients with P values and number of observations were generated for

AgeTRUE Time Injury with MonthsTRUE and YearsTRUE Job Tenure variables. As part of

performing the SAS Correlation Procedure, the mean, standard deviation and minimum and

maximum values were generated for these three variables. With minimum and maximum values

identified for AgeTRUE Time Injury (18-90 years), it was possible to calculate the maximum

possible job tenure by years (72 years) and by months (864). All other values, including negative

values and extreme values, were classified as missing for the tenure variables. For the maximum

age value, ≥90 years was determined to be the maximum age according to requirements of the

Safe Harbor §164.514(b)(2) clause of the Health Insurance Portability and Accountability Act

(HIPAA) Privacy Rule.97 All age values ≥90 years were classified as “90” years as part of the

de-identification of the dataset phase described earlier in this chapter. The SAS Frequency

Procedure was used to analyze the distribution of female hotel room cleaner MSD claimants by

age group and tenure group. Number of observations and percentages were reported.

Descriptive statistical analyses were performed using the SAS Frequency Procedure to

characterize MSD injuries in hotel room cleaners by injury variables (nature of injury, cause of

injury category, cause of injury, part of body category, part of body). All injury variables are

categorical variables. Subcategories of these three variables were determined by the distribution

of the study data. Number of observations and percentages were reported.

Aim 3B: Descriptive Analyses: Task at Time of Injury, Years 2012-2016

Descriptive statistical analyses were performed using the SAS Frequency Procedure to

characterize MSD injuries in hotel room cleaners using injury variables stratified by seven task at

time of injury categories and counts of claims by task at time of injury is reported. The seven

task categories are described in the Variable Definition for the ‘Task at Time of Injury’ variable

an. Injury variables include: cause of injury, e.g., lifting, carrying, twisting, etc.; nature of injury,

e.g., strain, sprain, contusion; and part of body injured. Subcategories of these three variables

were determined by the distribution of the study data for each task category. Injury variables

were sorted by seven task categories and percentages of injury are reported, along with sample

size for each task.

Section 4: Results

Part 1: Transformation of Narrative Text Data into Quantitative Data

This section describes the steps taken to transform narrative text data for occupation and task into

quantitative data, focusing on the creation of new variables and response codes; the subsetting of

claims data into sequential data samples based on the responses and corresponding decisions

made based on the data; and quality control checks on each new data sample and key variables.

The utility of the variables for this study and their potential for future research will be discussed

in Section 5, Conclusions and Discussion.

Aim 1. Identification of Hotel Room Cleaner Occupation

Creation of the “Female Hotel Industry MSD Claims” Data Sample

To achieve Aim 1, the identification of the occupation of hotel room cleaner in hotel industry

MSD workers’ compensation claims data, the Female Hotel Industry MSD Claims data sample

was created, following the methods detailed in Section 3 Methods. As a result, there are 23,070

hotel industry MSD claims for all occupations for female claimants in the dataset for study years

2007-2016.

Creation of the “Hotel Room Cleaner Occupation” Variable and Response Codes

A new variable “Hotel Room Cleaner Occupation” was created and identifies whether the

claimant’s occupation is related to the job of hotel room cleaning based on the occupation listed

in the narrative text field from the Occupation_Desc variable. This variable determines whether

or not a claim is added to the “Female Hotel Room Cleaner Occupation MSD Claim” data

sample and whether or not it is counted as an MSD injury having occurred to a female hotel

room cleaner. There were 2,216 unique occupation terms identified for all claims in the dataset.

I selected job titles from the definition of “Housekeeper” from the Cal/OSHA Title 8. Hotel

Housekeeping Musculoskeletal Prevention standard to use to code all claims: “Housekeeper”

means an employee who performs housekeeping tasks and may include employees referred to as

housekeepers, guest room attendants, room cleaners, maids, and housepersons” (emphasis

added).1 These job titles were used to determine which of the 2,216 occupation terms in the

dataset pertained to the occupation of hotel room cleaner. I identified a total of 184 unique

occupation terms that included one of the above five housekeeper job titles. There were several

iterations of each of the five job titles due to spelling errors, and non-standard nomenclature, e.g.,

multiple abbreviations for the same term.

All claims were coded to answer the research question “Is this occupation term a hotel room

cleaner?” The response categories are 1) “hotel room cleaner” – observations that meet the hotel

room cleaner definition; 2) “non-hotel room cleaner” – observations that do not meet the “hotel

room cleaner” definition; 3) “possible-hotel room cleaner” – observations that may meet the

“hotel room cleaner” definition, referred to as “Maybe” category; 4) “questionable-hotel room

cleaner” - observations that are uncertain if they meet “hotel room cleaner” definition, referred to

as “Question” category; 5) misclassified text; 8) missing and 9) unknown.

I developed an algorithm to aid in the decision-making about coding occupation terms that were

unclear if the claim was a hotel room cleaner, resulting in the “Maybe” and “Question”

categories. To illustrate the coding algorithm, for each claim, the occupation term was compared

to the five occupation categories and coded following these guidelines: exact match, code “yes,”

(1); no match, code “no,” (2); includes one of the five occupation categories plus another term,

code “maybe,” (3); and includes one of the five occupation categories plus another term that is

not clear if it is related to hotel room cleaning, code “questions,” (4). For example, the

occupation term “housekeeper” is an exact match, code yes; “housekeeper support,” code as

maybe; “housekeeper sup,” code as questions because it could be housekeeper support or

housekeeper supervisor; “housekeeper supervisor,” code as no; “worker,” code as unknown (9),

and “9050 HOTELS - ALL EMPLOYEES,” code as misclassified text (5). Empty fields were

coded as “Missing,” (8). Where questions arose on certain occupation terms such as “Executive

Housekeeper” or “Turndown Attendant,” other sources of information were consulted such as job

descriptions in employment websites, e.g., “Indeed,” job classifications from union contracts

with hotel employers, O*Net Summary Reports for Maids and Housekeeping Cleaners

(372012.00),21 scientific literature, and government research studies.

Creation of the “Hotel Room Cleaner Job Title” Variable and Response Codes

Another new variable “Hotel Room Cleaner Job Title” was created using the same five variable

response codes described above: “housekeepers, guest room attendants, room cleaners, maids,

and housepersons.”1 This variable answers the study research question, “What is the job title of

the injured hotel room cleaner?” This variable determined whether or not the claimant was coded

as working in a hotel room cleaner job title, allowing for future data analysis by job title.

Based on the job title term, the claim was then coded by hotel room cleaner job title. A “Not

Applicable,” 6 response category was created for all non-hotel room cleaner responses. In

addition, two subclassifications: “Not Applicable, Maybe,” (63) and “Not Applicable, Question,”

(64) were created to correspond to the response codes “Maybe,” (3) and “Question,” (4) for the

Hotel Room Cleaner Occupation. This makes it possible to perform an analysis in the future to

identify how similar the injury and task experience is for claims with Maybe and Question codes

compared to claims identified as hotel room cleaners. Then, depending on the findings of such an

analysis, this provides the opportunity to recommend consideration of Maybe and Question job

titles as occupation terms with hotel room cleaning-related MSDs.

Testing AI Team’s Methods on Hotel Room Cleaner Occupation and Job Title Response Codes

The AI team requested training data – examples of occupation terms from the narrative text

“Occupation_Desc” variable. I provided them a “training” subsample of 1,000 claims consisting

of 100 randomly selected claims for each of the ten study years. This training subsample was

shared with the AI team to filter and code the claims by hotel room cleaner occupation and job

title category, the two new variables related to occupation narrative text data. The AI team found

that numerous misspellings of the occupation description terms made the use of automatic

language processing methods impractical (Appendix I). Instead, as described earlier, Wiseyak

Inc. used a “list maintenance” method to classify the occupation description terms into the hotel

room cleaning occupation codes. I provided feedback to the AI team, who further refined the

method and re-coded the training data. The AI-coded training data were found to be 98.4%

accurate for coding the Hotel Room Cleaner Occupation variable and 99.4% accurate for coding

the Hotel Room Cleaner Job Title variable. The entire data sample was sent to Wiseyak Inc. to

filter and code the Hotel Room Cleaner Occupation and the Hotel Room Cleaner Job Title

variables using the codebook finalized after coding the training data.

Results of Quality Control Checks on Coded Hotel Room Cleaner Occupation and Job Title Data

A CUNY Graduate School of Public Health and Health Policy (GSPHHP) student researcher

was hired by the CUNY Research Foundation to assist me in performing a 20% quality control

check on the coded responses. In addition, I manually reviewed the AI-coded data for responses

to the Hotel Room Cleaner Occupation and Job Title variables: 33% (7,698) of the entire dataset

for claims with all possible responses for Hotel Room Cleaner Occupation (n=23,070) were

reviewed along with corresponding responses for Hotel Room Cleaner Job Title.

Table 4.1 depicts adjustments made to the codebook as a result of the 20% quality control check

and two manual reviews described above. Occupation terms decreased for response codes

Unknown (n=7) and No (n=4) causing an increase in the Maybe, Question and Misclassified

claims. An improved coding scheme resulted with an increased number of potential hotel room

cleaner occupation terms to research in the future. The data sample was recoded according to the

Table 4.1 Hotel Room Cleaner Occupation Term Counts:

Pre and Post Quality Control Checks

Hotel Room Cleaner Occupation Term Counts: Pre

and Post Quality Control Checks

Occupation Term

Original

Coding Decision

Revised Coding

Decision

Yes

184

1,795

1,791

Maybe

139

141

Question

Misclassified

Unknown

Missing

Total

2,216

Creation of the “Female Hotel Room Cleaner Occupation MSD Claims” Analytic Data Sample

All claims not meeting the criteria for a hotel room cleaner were removed from the data sample.

Of the remaining claims, 48% (5,829) of the responses for the Hotel Room Cleaner and Hotel

Room Cleaner Job Title variables were reviewed. The “Female Hotel Room Cleaner Occupation

MSD Claims Study Dataset” contains 12,125 claims for 2007-2016 and is the analytic data

sample for Aim 3A (see Part 2, Descriptive Analyses: Female Hotel Room Cleaner MSD Claims

Data).

Aim 2. Identification of Task at Time of Injury

Creation of the “5-Year Female Hotel Room Cleaner MSD Claims” Data Sample

To achieve Aim 2, the identification of the task at time of injury in female hotel room cleaner

MSD workers’ compensation claims data, the “5 Year- Female Hotel Room Cleaner Task at Time

of Injury” data sample was created from the Female Hotel Room Cleaner Occupation MSD

dataset. Study years were reduced to the most recent five years of data, 2012-2016, to reflect

current changes in hotel accommodations and hotel housekeeping work practices.

Creation of the “Task at Time of Injury” Variable and Response Codes

A new variable “Task at Time of Injury” was created and identifies the specific job task the hotel

room cleaner was performing when injured. This new variable is based on narrative text data

from the “Accident Description/Cause” variable. This variable determines whether the claim is

included in this study’s task analysis or for future research. Task codes were created using

seven categories for “Housekeeping tasks” defined in the Cal/OSHA Title 8.3345 Hotel

Housekeeping Musculoskeletal Injury Prevention standard, section (b) Definitions.1 See Study

Variables, Section 3 Methods and additional response codes in the next paragraph for details. All

claims were coded to answer the study research questions, “What is the task at time of injury?”

Two qualified coders with experience in occupational injury research (my dissertation committee

sponsor and myself) coded the same 550 cases drawn from a random sample stratified by year.

This number of cases represents approximately a 15% sample of the estimated final task data

sample. Issues of concern were flagged, disagreements in coding task data that did not meet the

seven task categories were resolved, and new task codes and an initial codebook were created.

A “missing” code (88) was created. “Emergent hazards” were identified such as “making a

rollaway,” resulting in the creation of subclassifications of the relevant seven task categories,

e.g., “rollaway (21), “sofa bed” (22), “Murphy Bed” (23) as subclassifications of “making bed”

task (2). The task category “Unknown” (9) was created for data where task was not determined;

multiple subclassifications of “unknown” task codes were created (see details in the “Study

Variables” section). This follows subject matter experts’ recommendations to add “unknown” as

a response category for the task variable since previous researchers coded task as “unknown” for

33% of WCIS cases studied.47

Creation of Artificial Intelligence (AI) Model to Classify and Code Task Data

I provided training subsamples for the AI team to develop a NLP model to filter claims and code

task at time of injury. The training subsample consisted of raw task data from the Accident/

Cause variable (injury description) for 545 hotel room cleaner claims, randomly selected and

stratified by year and broken down by each task category, e.g., 50 claims with task data about

software features “Filter,” “Find” and “Sort” to select and de-identify the training data before

sending to the AI team.

Based on my feedback and additional training sample data (n=300 claims), the AI team further

refined the model, generating three iterations of the model. The larger the amount of training

data, the greater the accuracy of the NLP application (Appendix I). A “percent confidence level”

column was added by the AI team to indicate the probability that the model predicted the

accurate task code (classification) based on the narrative text in the injury description, e.g., the

model coded a claim for task at time of injury as “bedmaking,” (2) and the narrative text data in

the injury description corresponded to a bedmaking task. The limits of the model to code

subclassifications of task categories or to classify the subcategories of unknown responses, e.g., a

narrative text field with only a diagnosis or a part of body injured and no description of the task,

were identified by the AI team and myself. This resulted in the recognition that the “unknown”

(9) code would capture this data broadly, leaving reconciliation of coding the unknown data to

more specific subclasses to be performed possibly in the future. When comparing the coded

responses for Task at Time of Injury variable to the training sample of 200 claims, 90% of the

claims were coded accurately. The codebook was finalized with variables, response labels and

codes, coding rules and justification for codes. The AI model was deemed ready to be run on the

task data sample.

Results of De-identification of Task Data Sample

Before sending the task data sample to the AI team for coding task, the data sample was

deidentified (see Section 3 Methods). Gender terms such as “she” and “her” were extensive in

the narrative text injury descriptions and were the most frequent type of identifiable data. Even

where a worker’s name or the term “she” instead of employee’s name was not used, “her” was

used widespread in the entire task dataset. Paired examples of before and after de-identification

narrative text injury descriptions are included below, the employee’s name removed as needed:

“[employee name] was sweeping balconies when she began to feel pain”

“ee was sweeping balconies when ee began to feel pain”

“while pulling the sofa bed forward with both hands in a guest room, she injured her

right shoulder”

“while pulling the sofa bed forward with both hands in a guest room, ee injured right

shoulder”

“[employee name] was cleaning room [number] shower head and felt a sharp pain in her

left calf”

“ee was cleaning room x shower head and felt a sharp pain in left calf”

De-identification for gender terms, in particular, was a lengthy process as the task data sample

sent to the AI team to code included of 8,442 claims, of which 1,490 were “Maybe” and

“Questions” claims.

Results of Quality Control Sampling of Narrative Text Data Classified into Task Categories After

receiving the coded task data from the AI team, I removed 1,490 claims with hotel occupation

coded as Maybe or as Questions as these claims were included so they could be coded for task

for possible future research. Of the 6,952 hotel room cleaner claims remaining (82% of the

original task data sample sent to the AI team), 3,013 (43%) were coded as unknown and were

removed. The remaining 3,939 claims comprise the task data sample that was used to perform

quality control checks on the Task at Time of Injury variable response codes.

Table 4.2 provides the results of the quality control checks on task data for each of the seven task

categories and two subtasks, and by key indicators. Overall, 14.9% of the coded task data was

sampled, with percent quality control checks varying by size of task category as explained in

Section 3 Methods. Over two-thirds (68.3%) of the coded task data sample had claims in the

≥70% confidence level category. The overall estimated accuracy of the coded task data is 84.5%.

Based on these results, I decided to remove all claims that had a task confidence level < 70% to

improve coding accuracy which is supported by the literature (Marucci-Wellman et al., 2017)

and was recommended by the AI team. Also, 89% of all claims (using a weighted average) were

deemed as related to hotel room cleaning tasks.

Table 4.2 Results of Quality Control Sampling of Narrative Text Data Classified into Numeric Task Categories, Years 2012-2016

N % %

TASK % QC QC HRC- classified

CATEGORY Code N sample sample related correct

≥70%

conf.

level

correct

≥70%

claims

only

≥70%

claims

correct

clean

1,235

20%

247

90%

55%

71.4%

70.7%

19.1%

19.2%

881

622

bedmaking

1,424

10%

143

89.5%

80%

69.2%

91.9%

68.8%

46.4%

984

904

vacuum

186

15%

89.7%

83%

27.6%

100%

85.7%

71.4%

cart

593

10%

88.5%

84%

80.3%

89.8%

100%

44.4%

476

427

supplies

20%

70.6%

53%

29.4%

60%

50%

trash

138

15%

85.7%

62%

61.9%

100%

n.a.

move furniture

154

15%

82.6%

61%

47.8%

73%

n.a.

54.5%

missing

n.a.

Subtasks

Rollaway bed

100%

62.5%

100%

sofa bed

103

20%

100%

n.a.

103

claims

≥70%

conf.

level

claims

6069%

conf.

level

claims

<60

conf.

level Estimate

claims N*

Estimate

N**

Murphy bed 23 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. TOTAL 3,939 586 2,691 2,273

% TOTAL

SAMPLED

OVERALL

14.9%

% ESTIMATED

ACCURACY RATE

OF ESTIMATED

TASK ANALYTIC

DATA SAMPLE

84.5%

*Estimate N >=70% confidence level claims only = result of Column E (n of sample) x Column H (% claims >=70%)

**Estimate N >=70% confidence level claims correct = result of Column I (%correct of >=70% confidence

level claims) x Column L (Estimate N >=70% confidence level claims only)

n.a.

Creation of the “Female Hotel Room Cleaner Task at Time of Injury” Analytic Data Sample,

Years 2012-2016

Based on the results of the quality control checks on the coded task data, the accuracy and

quality of the coded task data was considered acceptable. The “Female Hotel Room Cleaner

Task at Time of Injury Dataset” consists of 2,629 claims and is the analytic data sample for Aim

3B (see Part 2, Descriptive Analyses: Female Hotel Room Cleaner MSD Claims Data). The steps

to arrive at this dataset from the initial Female Hotel Industry MSD Claims dataset (n=23,070)

and then sequentially from the Female Hotel Room Cleaner Occupation MSD Claims dataset

(n=12,125) was presented in Figure 3.2, page 54, Section 3.

Part 2: Descriptive Analyses of Female Hotel Room Cleaner MSD Injury Claims Data

Aim 3A: Female Hotel Room Cleaner Claimant Demographic, Work and Injury Characteristics

Research Question 1: Can we establish that as an occupation, hotel room cleaners experience

MSD injuries?

Of the 23,070 female hotel industry MSD claims for years 2007-2016, the majority of the

claimants, 52.6%, were hotel room cleaners (n = 12,125).

Research Question 2: Can we characterize hotel room cleaner MSD claims by demographic

factors of the host such as gender and age?

Gender

All claimants were female. The decision to include only female claimants was made when setting

parameters for the hotel industry study dataset. The decision was based on the high percentage of

females in the job titles referred to above and the proportion these two job titles accounted for in

claims with hotel room cleaner job titles (Study Dataset, Section 3 Methods).

Age

The mean age of the study population was 43.6 years (SD 10.9) with ages ranging from 18 to 90

years (Table 4.3, page 89). By age group (Table 4.4, page 89), the highest proportion of hotel room

cleaner MSD claims were for claimants aged 45-54 years (33.1%), followed by claimants aged 35-

44 years (28.9%) and 25-34 years (16.8%). The number of claims increased with age group,

peaking in the 45-54 years age group, then decreasing by more than half for those 55-64 years.

The lower numbers observed in the latter age group may reflect the age distribution in the hotel

room cleaner population overall. Unfortunately, I do not have access to such populationlevel data.

Research Question 3: Can we characterize hotel room cleaner MSD claims by work-related

factors of the host such as job title category, e.g., housekeeper, maid, and job tenure?

Job Title Category

The job title of Housekeeper accounts for 58.2% of all claims, followed by Room Attendants with

39.5% (Table 4.5).

Table 4.5 Breakdown of Female Hotel Room Cleaner Occupation by Job Title, 2007-2016

(n=12,125)

Hotel Room Cleaner Job Title

Housekeeper

7,055 58.2

Room Attendant

4,785 39.5

Room Cleaner

84 0.7

Maid

85 0.7

Houseperson

116 0.9

Total

12,125 100

Job Tenure

Findings for mean job tenure are in Table 4.3, page 89 and are expressed using two metrics:

“years” (the time difference in years between Date of Hire and Date of Injury) and “months”

(the time difference in years between Date of Hire and Date of Injury) computed using SAS

INTCK procedure (see Section 3, Study Variables, page 60). The mean job tenure at time of injury

using a “years” metric was 5.2 years (SD 6.6), ranging from 0-62 years. Using a

“months” metric, the mean job tenure at time of injury was 67.9 months (SD 79.9), ranging from

0-746 months. The mean job tenure amount in months is the equivalent of 5.7 years, capturing

five more months of job tenure than the years metric. By job tenure group (Table 4.6), hotel

room cleaners with 2-5 years job tenure accounted for the highest proportion of claims (21.3%).

The 5-10 years and the ≤6 months job tenure groups each accounted for approximately 18% of

all claims. Further research is needed to understand the significance of the tenure group findings

due to the low to mid-tenure groups’ representation in the distribution of hotel room cleaner

claims. These findings may reflect the overall distribution of tenure among the hotel room

cleaner population or could indicate an area of further research.

Table 4.6 Distribution of Female Hotel Room Cleaner MSD Claims by Job Tenure, Californi

Workers’ Compensation Claims, 2007-2016 (n=11,559)

Job Tenure N %**

7 months-1 year 1,170

10.1

1-2 years

1,491

12.9

2-5 years

2,464

21.3

5-10 years

2,116

18.3

10-20 years

1,672

14.5

20+ years

564

4.9

Missing = 566.

** Percentages may not add up to 100% due to rounding.

Age Group by Tenure Group

The correlation between age and tenure (Table 4.3, page 89) was 0.44 (P < 0.0001) for both

tenure variables (month and year), indicating a medium correlation. Figure 4.1 presents a

grouped bar plot of the seven tenure groups by the eight age groups to better understand age and

tenure in the study sample. The ≤6 months tenure group accounted for the majority of claims for

age groups 18-24 and 25-34 years and ranked second for the 35-44 years age group. The 2-5

years tenure group ranked first for the 35-44 years age group and second for the 45-54 age group.

The 5-10 years tenure group ranked first for the 45-54 age group. The 10-20 years tenure group

ranked third for the 45-54 years age group and accounted for the majority of claims for the 55-64

years age group. The findings by age and tenure groups indicate that lower tenure groups account

for substantial proportions of mid-range age groups. Further analysis is needed to understand age

and tenure group patterns in the data. There is a need to determine whether the prevalence of

injuries for the ≤6 months tenure group across various age groups reflects the actual distribution

of tenure by age group or instead reflects a trend of higher injury prevalence for hotel room

cleaners with less job experience.

Research Question 4: What are the leading nature of injury categories found in the claims

data?

Nature of Injury

By Nature of Injury (Table 4.7, page 90), “Strain or Tear” injuries accounted for close to twothirds

of all injuries (64.9%), followed by “Sprain or Tear” injuries with 11.9%

Research Question 5: What are the leading cause of injury categories found in the claims data?

Cause of Injury Category

By Cause of Injury Category (Table 4.8, page 91), “Strain or Injury By” accounted for the overwhelming

majority of injuries (92%). By individual Cause of Injury (Table 4.9), “Strain or

Injury By, NOC” accounted for 27.7% of claims, followed by “Pushing or Pulling” (18.4%), “Repetitive

Motion” (16.6%) and “Lifting” (14.5%).

Table 4.9 Distribution of Hotel Room Cleaner Claimants’ MSD Injuries by Cause of Injury,

California Workers’ Compensation Claims, 2007-2016 (n=12,125) Cause of Injury

N %*

Strain or Injury By, NOC

3,352

27.7

Pushing or Pulling

2,231

18.4

Repetitive Motion

2,017

16.6

Lifting

1,761

14.5

Cumulative, NOC

885

7.3

Twisting

767

6.3

Reaching

635

5.2

Other**

477

3.9

*Percentages may not add up to 100% due to rounding.

**All remaining categories, each of which account for <3% of injuries.

Research Question 6: What are the leading part of body injured categories found in the claims

data?

By Part of Body Category (Table 4.10), “Upper Extremities” was the most frequent category, with

40.5% of all claims, followed by “Trunk” (33.2%) and “Lower Extremities” (13.4%).

Table 4.10 Distribution of Hotel Room Cleaner Claimants’ MSD Injuries by Part of Body Category,

California Workers’ Compensation Claims, 2007-2016 (n=12,125) Part of Body Category

N %

Upper Extremities

4,909

40.5

Trunk

4,030

33.2

Lower Extremities

1,626

13.4

Multiple Body Parts

1,201

9.9

Neck

168

1.4

Other*

191

1.6

*All remaining categories, each of which account for <1% of injuries.

By individual Part of Body (Table 4.11), Low Back Area’ accounts for the highest proportion of claims

(27.2%), followed by “Shoulders” (14.3%). Taken together, “Multiple Body Parts,”

“Wrist” and “Knee” account for about one-quarter of remaining hotel room cleaner MSD claims.

Table 4.11 Distribution of Hotel Room Cleaner Claimants’ MSD Injuries by Part of Body, California

Workers’ Compensation Claims, 2007-2016 (n=12,125) Part of Body

N %*

Low Back Area (Lumbar and Lumbo-Sacral)

3,292

27.1

Shoulder(s)

1,732

14.3

Multiple Body Parts (incl. Body Systems and Body Parts)

1,201

9.9

Wrist

877

7.2

Knee

762

6.3

Hand (excl. Wrist and Fingers)

459

3.8

Multiple Upper Extremities

445

3.7

Lower Arm

360

3.0

Ankle

346

2.8

Upper Arm (excl. Clavicle and Scapula) 296

2.4

Upper Back Area (Thoracic Area) 248

2.0

**All remaining categories, each of which account for <2% of injuries.

Additional Tables: Aim 3A Demographic and Injury Characteristics

Table 4.3 Mean Age and Job Tenure Characteristics of Hotel Room Cleaner Claimants with MSD

Injuries, California Workers’ Compensation Claims, 2007-2016 (n=12,125)

Age at time of injury

Mean

(SD)*

Minimum

Maximum

12,109**

43.6

10.9

Job tenure at time of injury (months)

11,558***

67.9

79.9

746

Job tenure at time of injury (years)

11,588***

5.2

6.6

*SD, standard deviation.

**Missing = 16 (0.13%)

***Missing = 567 (4.68%)

Other**

2,107 17.4

Percentages may not add up to 100% due to rounding.

Table 4.4 Distribution of Hotel Room Cleaner MSD Claims by Age Group, California Workers’

Compensation Claims, 2007-2016 (n=12,109)* Age Group

N %*

18-24

597

4.9

25-34

2,039

16.8

35-44

3,496

28.9

45-54

4,009

33.1

55-64

1,782

14.7

65-69

129

1.1

70-74

0.3

75+

0.2

*Missing = 16 (0.13%)

Table 4.5 Breakdown of Hotel Room Cleaner Occupation by Job Title, 2007-2016 (n=12,125)

Hotel Room Cleaner Job Title N %

Housekeeper 7,055 58.2

Room Attendant 4,785 39.5

Room Cleaner 84 0.7

Maid 85 0.7

Houseperson 116 0.9

Total 12,125 100

Table 4.6 Distribution of Hotel Room Cleaner MSD Injuries by Job Tenure, California Workers’

Compensation Claims, 2007-2016 (n=11,559)

Job Tenure

%**

7 months-1 year

1,170

10.1

1-2 years

1,491

12.9

2-5 years

2,464

21.

5-10 years

2,116

18.3

10-20 years

1,672

14.5

20+ years

564

4.9

Missing = 566 (4.7%)

** Percentages may not add up to 100% due to rounding.

Table 4.7 Distribution of Hotel Room Cleaner MSD Injuries by Nature of Injury, California Workers’

Compensation Claims, 2007-2016 (n=12,125)

Strain or Tear

7,864 64.9

Sprain or Tear

1,444 11.9

All Other Specific Injuries, NOC

994 8.2

All Other Cumulative Injuries

933 7.7

Inflammation

340 2.8

Multiple Physical Injuries Only

250 2.1

Other **

300 2.5

*Percentages may not add up to 100% due to rounding.

**All remaining categories, each of which account for <1% of injuries.

Table 4.8 Distribution of Hotel Room Cleaner MSD Injuries by Cause of

Injury Category, California Workers’ Compensation Claims, 2007-2016

(n=12,125) Cause of Injury Category

N %

Strain or Injury By

11,156

92.0

Miscellaneous Causes

885

7.3

Rubbed or Abraded

0.6

Striking Against or Stepping Down 13 0.1

Table 4.9 Distribution of Hotel Room Cleaner Claimants’ MSD Injuries by Cause of Injury,

California Workers’ Compensation Claims, 2007-2016 (n=12,125) Cause of Injury

N %*

Strain or Injury By, NOC

3,352

27.7

Pushing or Pulling

2,231

18.4

Repetitive Motion

2,017

16.6

Lifting

1,761

14.5

Cumulative, NOC

885

7.3

Twisting

767

6.3

Reaching

635

5.2

Other**

477

3.9

*Percentages may not add up to 100% due to rounding.

**All remaining categories, each of which account for <3% of injuries.

Table 4.10 Distribution of Hotel Room Cleaner Claimants’ MSD Injuries by Part of Body Category,

California Workers’ Compensation Claims, 2007-2016 (n=12,125) Part of Body Category

N %

Upper Extremities

4,909

40.5

Trunk

4,030

33.2

Lower Extremities

1,626

13.4

Multiple Body Parts

1,201

9.9

Other*

359

3.0

*All remaining categories, each of which account for <2% of injuries.

Table 4.11 Distribution of Hotel Room Cleaner Claimants’ MSD Injuries by Part of Body, California

Workers’ Compensation Claims, 2007-2016 (n=12,125) Part of Body

N %*

Low Back Area (Lumbar and Lumbo-Sacral)

3,292

27.1

Shoulder(s)

1,732

14.3

Multiple Body Parts (incl. Body Systems and Body Parts)

1,201

9.9

Wrist

877

7.2

Knee

762

6.3

Hand (excl. Wrist and Fingers)

459

3.8

Multiple Upper Extremities

445

3.7

Lower Arm

360

3.0

Ankle

346

2.8

Upper Arm (excl. Clavicle and Scapula) 296

2.4

Upper Back Area (Thoracic Area) 248

2.0

Other** 2,107

17.4

*Percentages may not add up to 100% due to rounding.

**All remaining categories, each of which account for <2% of injuries.

Aim 3B: Female Hotel Room Cleaner Claimant Injury Characteristics by Task at Time of Injury

Female Hotel Room Cleaner Claimant Task at Time of Injury Dataset, 2012-2016

The Female Hotel Room Cleaner Task at Time of Injury dataset includes 2,628 claims with task at

time of injury identified for the study years 2012-2016; one claim was removed due to injury

missing incident data. All claims in this dataset meet the threshold of belonging to the ≥70%

confidence level category, signifying that the narrative text data describing the injury incident

was correctly classified by task category. This dataset is also referred to as the Aim 3B analytic

data set and answers the research question, “What was the task at time of injury?” Results from

descriptive statistical analyses to characterize MSD injuries in hotel room cleaners stratified by

seven task categories are presented in Tables 4.12-4.15 on the following pages.

Research Question 7: Can we characterize hotel room cleaner MSD injuries by task at time of

injury?

For the Task at Time of Injury dataset overall, bedmaking claims accounted for the highest

percentage of all claims (43.5%), followed by cleaning surfaces (29.9%) and handling linen carts

(17.2%). The remaining 9.4% of claims were distributed among the four remaining tasks in the

following order: collecting and disposing trash (3.3%), vacuuming (2.7%), moving furniture

(2.3%) and moving linen and supplies (1.1%) in and out of guest rooms.

Table 4.12 Breakdown of Task at Time of Injury, Hotel Room Cleaner MSD Claims (n = 2,628)*

Task Category

N**

Cleaning Surfaces

785

29.9

Making Beds

1143

43.5

Vacuuming

2.7

Handling Linen Carts

453

17.2

Moving Linen/Supplies

1.1

Collecting/Disposing Trash

3.3

Moving Furniture

2.3

Aim 3B: Injury Variables Stratified by Seven Task Categories

Nature of Injury Variable

For the Task at Time of Injury dataset overall, by Nature of Injury category, “Strain or Tear” was

the most frequent type of injury accounting for 75.3% of all claims. By task at time of injury, the

Nature of Injury frequencies varied (Table 4.13). “Strain or Tear” injuries were the leading nature

of injury for all tasks ranging from 65.3% while vacuuming to 82.1% for moving linen/supplies.

“Sprain or Tear” injuries ranked second for all tasks except for vacuuming where it ranked third

after “All Other Specific Injuries, NOC” (12.5%). “Sprain or Tear” injuries were most frequent

when handling linen carts (14.6%). “All Other Cumulative Injuries” occurred most often when

vacuuming (5.6%), followed by cleaning surfaces (2.4%). “Inflammation” injuries were the

highest for collecting/disposing trash (3.5%). Hotel room cleaners moving linen/supplies in and

out of guest rooms had the highest frequency of dislocation injuries (3.6%). Both “Inflammation”

and “Dislocation” occurred across cleaning surfaces, bedmaking, handling linen carts and

vacuuming tasks. Moving furniture was the only task associated with “Carpal Tunnel

Syndrome.” Hernia injuries occurred at low percentages yet were associated with three tasks:

vacuuming (1.4%), collecting/disposing trash (1.2%) and handling linen carts (0.9%).

Table 4.13 Nature of Injury by Hotel Room Cleaner Tasks, California Workers’ Compensation MSD Claims, 2012-2016 (n=2,628)

Cleaning Surfaces (n = 784)

Making Beds (n = 1,140)

Vacuuming (n = 72)

Strain or Tear

74.2

Strain or Tear

76.7

Strain or Tear

65.3

Sprain or Tear

13.0

Sprain or Tear

12.3

All Other Specific Injuries, NOC

12.5

All Other Specific Injuries, NOC

5.9

All Other Specific Injuries, NOC

5.7

Sprain or Tear

8.3

All Other Cumulative Injuries

2.4

Inflammation

1.7

All Other Cumulative Injuries

5.6

Inflammation

2.0

Dislocation

1.4

Inflammation

2.8

Dislocation

1.5

Other**

2.2

Multiple Physical Injuries Only

2.8

Other**

0.9

Missing = 3

Dislocation

1.4

Missing = 1

Hernia

1.4

Handling Linen Cart (n = 452)

Moving Linen/Supplies (n = 28)

Collecting/Disposing Trash (n = 86)

Strain or Tear

73.2

Strain or Tear

82.1

Strain or Tear

80.2

Sprain or Tear

14.6

Sprain or Tear

10.7

Sprain or Tear

9.3

All Other Specific Injuries, NOC

7.5

All Other Specific Injuries, NOC

3.6

All Other Specific Injuries, NOC

3.5

Inflammation

2.2

Dislocation

3.6

Inflammation

3.5

Dislocation

0.7

All Other Cumulative Injuries

2.3

Hernia

0.7

Hernia

1.2

Other***

1.1

Missing = 1

Moving Furniture (n = 61)

Strain or Tear

78.7

Sprain or Tear

9.8

All Other Specific Injuries, NOC

6.6

Carpal Tunnel Syndrome

1.6

Dislocation

1.6

Multiple Physical Injuries Only

1.6

*Percentages may not add up to 100% due to rounding.

**All remaining categories, each of which account for <1% of injuries.

***All remaining categories, each of which account for <0.5% of injuries.

Cause of Injury Variable

For the Task at Time of Injury dataset overall, by Cause of Injury, six subcategories account for

95.7% of all MSD claims: “Strain or Injury By, NOC” (28.1%); “Pushing or Pulling” (24.8%);

“Lifting” (17.8%); “Twisting” (8.7%); “Reaching” (8.3%) and “Repetitive Motion” (8%). The

remaining claims are represented by six other cause of injury categories with small percentages

(<1%), with the exception of “Holding or Carrying” that accounts for 2.3% of all MSD claims.

As with the other injury variables, the distribution of hotel room cleaner MSD claims for Cause

of Injury varies by the Task at Time of Injury category. The tasks with “Pushing or Pulling”

ranking first for cause of injury were handling linen cart (61.5%), moving furniture (45.9%) and

vacuuming (34.7%), followed by making beds (23.3%), moving linen/supplies (14.3%) and

collecting/disposing trash (10.5%) where it ranked third as cause of injury. Lifting was the most

frequent cause of injury for collecting/disposing trash (48.8%), moving linen/supplies (32.1%)

and making beds (26.5%), followed by moving furniture (19.7%) and handling linen carts

(12.8%). Cleaning surfaces was the task with the highest proportion of “Strain or Injury By,

NOC” (45.3%), followed by making beds (26.2%), moving linen/supplies (21.4%) and moving

furniture (18%) where it ranked second. Cleaning surfaces and vacuuming were the tasks at time

of injury with repetitive motion ranked second, 15.7% and 12.5%, respectively. Twisting ranked

fourth for the following tasks: moving linen/supplies (14.3%), cleaning surfaces and making

beds (both 9.9%), collecting/disposing trash (7%) and handling linen carts (4.9%). Bedmaking

had the first three causes of injury with equal prevalence: “Lifting” (26.5%), “Strain or Injury,

By NOC” (26.2%) and “Pushing or Pulling” (23.3%).

Table 4.14 Cause of Injury by Hotel Room Cleaner Tasks, California Workers’ Compensation MSD Claims, 2012-2016 (n=2,628)

Cleaning Surfaces (n = 784)

Making Beds (n = 1,140)

Vacuuming (n = 72)

Strain or Injury by, NOC

45.3

Lifting

26.5

Pushing or Pulling

34.7

Repetitive Motion

15.7

Strain or Injury By, NOC

26.2

Repetitive Motion

12.5

Reaching

14.7

Pushing or Pulling

23.3

Strain or Injury by, NOC

12.5

Twisting

9.9

Twisting

9.9

Holding or Carrying

11.1

Pushing or Pulling

5.2

Reaching

6.0

Using Tool or Machinery

11.1

Lifting

5.0

Repetitive Motion

5.3

Lifting

6.9

Other**

4.2

Other**

2.8

Twisting

5.6

Missing = 1

Missing = 3

Reaching

4.2

Other**

1.4

Handling Linen Cart (n = 452)

Moving Linen/Supplies (n = 28)

Collecting/Disposing Trash (n = 86)

Pushing or Pulling

61.5

Lifting

32.1

Lifting

48.8

Lifting

12.8

Strain or Injury by, NOC

21.4

Strain or Injury by, NOC

11.6

Strain or Injury by, NOC

10.6

Pushing or Pulling

14.3

Pushing or Pulling

10.5

Twisting

4.9

Twisting

14.3

Twisting

7.0

Reaching

4.4

Holding or Carrying

7.1

Holding or Carrying

5.8

Holding or Carrying

3.1

Reaching

7.1

Reaching

5.8

Other**

2.7

Repetitive Motion

3.6

Wielding or Throwing

5.8

Missing = 1

Repetitive Motion

3.5

Moving Furniture (n = 61)

Pushing or Pulling

45.9

Lifting

19.7

Strain or Injury by, NOC

18.0

Repetitive Motion

8.2

Reaching

4.9

Other**

3.3

100

*Percentages may not add up to 100% due to rounding. **All

remaining categories, each of which account for <2% of injuries.

101

Part of Body Variable

Upper Extremities is the most frequently reported Part of Body Category for task dataset claims

(42.2%), followed by Trunk (37.55%) and Lower Extremities (14.33%). Together, these three

Part of Body Categories account for 94.1% of task dataset claims, yet the distribution of claims

by individual part of body varies with each task category (Table 4.15). Lower back injuries were

the leading cause of injury for all tasks with moving furniture having the highest proportion

(47.5%), followed by moving linen/supplies (42.9%), making beds (35.7%) and handling linen

carts (32.5%) among the tasks with higher proportions. Shoulder injuries ranked second for

hotel room cleaners across all tasks with collecting/disposing trash (19.8%), moving linen/

supplies (17.9%), cleaning surfaces (17.15), making beds (16.8%) and vacuuming (16.7%) the

leading tasks at time of injury. Moving furniture, cleaning surfaces and making beds were tasks

where knee and wrist injuries ranked third and/or fourth for hotel room cleaners. Vacuuming

had the highest proportion of hand injuries (9.7%) where it ranked fourth as cause of injury,

followed by cleaning surfaces (5.6%) and making beds (3.6%), ranking fifth for both tasks.

Ankle injuries occurred at low frequency levels yet were common to handling linen carts (6%)

and vacuuming (4.2%). The remaining cause of injuries were distributed among “other”

categories, with “Other – Upper Extremities” the leading “other” cause of injury across all tasks.

Cleaning surfaces (18.9%) was the leading task at time of injury for “Other – Upper Extremities”

as cause of injury, followed by handling linen/carts (14.8%) and making beds (14.2%). Moving

linen/supplies and collecting/disposing trash were the tasks with the highest proportion of “Other

– Lower Extremities” injuries, with 10.7% and 10.5%, respectively. Handling linen carts (7.7%),

moving linen/supplies (7.1%), moving furniture (6.6%), making beds (6.3%) and collecting/

disposing trash accounted for the task at time of injury for “Other – Trunk.”

102

Table 4.15 Part of Body Injured by Hotel Room Cleaner Tasks, California Workers’ Compensation MSD Claims, 2012-2016 (n=2,628)

Cleaning Surfaces (n = 784)

Making Beds (n = 1,140)

Vacuuming (n = 72)

Lower Back Area

22.4

Lower Back Area

35.7

Lower Back Area

26.4

Shoulder(s)

17.1

Shoulder(s)

16.8

Shoulder(s)

16.7

Knee

13.0

Wrist

8.9

Multiple Body Parts**

12.5

Wrist

7.0

Knee

5.5

Hand (excl. Wrist and Fingers)

9.7

Hand (excl. Wrist and Fingers)

5.6

Hand (excl. Wrist and Fingers)

3.6

Wrist

5.6

Other - Upper Extremities

18.9

Other - Upper Extremities

14.2

Ankle

4.2

Other - Lower Extremities

6.8

Other - Trunk

6.3

Other - Upper Extremities

13.9

Other

9.0

Other - Lower Extremities

3.9

Other - Lower Extremities

7.0

Missing = 1

Other

Missing = 3

5.1

Other

4.2

Handling Linen Cart (n = 452)

Moving Linen/Supplies (n = 28)

Collecting/Disposing Trash (n = 86)

Lower Back Area

32.5

Lower Back Area

42.9

Lower Back Area

31.4

Shoulder(s)

12.8

Shoulder(s)

17.9

Shoulder(s)

19.8

Wrist

7.5

Other - Upper Extremities

7.1

Wrist

9.3

Knee

6.4

Other - Trunk

7.1

Lower Arm

7.0

Ankle

6.0

Other - Lower Extremities

10.7

Upper Back Area

4.6

Other - Upper Extremities

14.8

Other - Multiple Body Parts

14.3

Multiple Body Parts**

3.5

Other - Trunk

7.7

Other - Upper Extremities

8.1

Other - Lower Extremities

6.4

Other - Trunk

5.8

Other

5.7

Other - Lower Extremities

10.5

Missing = 1

Moving Furniture (n = 61)

Lower Back Area

47.5

Shoulder(s)

14.8

Knee

9.8

Wrist

6.6

Other - Upper Extremities

9.8

Other - Trunk

6.6

Other

4.9

103

*Percentages may not add up to 100% due to rounding. **Includes Body Systems and Multiple Body Systems.

104

Section 5: Discussion and Conclusions

Study Summary

The goal of this dissertation was to establish the occurrence of hotel room cleaner MSDs using

administrative data and in doing so, develop research methods that will aid researchers to

produce science-based evidence necessary for effective intervention development and program

implementation for hotel room cleaner MSD injury control. This goal was achieved through the

study aims by first identifying the hotel room cleaner occupation (Aim 1) at the hotel industry

study population level (n = 23,070) and documenting the occurrence of hotel room cleaner

MSDs using 12,125 California WC claims. Data analysis produced descriptive findings by cause

and nature of injury and part of body injured for this study sample (Aim 3A). Second, by

identifying the task at time of injury by coding tasks from text data using artificial intelligence

methods (Aim 2), leading causes, nature of injuries and part of body were analyzed by each task

(Aim 3B) for the 2,668 claims where task could be coded, providing insights into MSD claim

injuries by task. Though resource intensive, this study demonstrates the potential for this type of

analysis to examine WC data at the occupation and task level for future research.

Summary of Findings

Aim 1: Identification of Hotel Room Cleaner Occupation

It was possible to identify the occupation of hotel room cleaner using narrative text data from the

Occupation Description field. The occupation of hotel room cleaner is comprised of five job

105

titles that originate from the definition of a “Housekeeper” using Cal/OSHA 3345. Using these

job titles, a list of 2,216 occupation terms was reduced to almost 184 terms that were considered

equivalent to a hotel room cleaner occupation. An algorithm was created to triage those claims

that were considered to be “Maybe” (n = 141 terms) or “Questions” (n = 43 terms) when

answering the question, “Is this a hotel room cleaner?” By coding them separately, these claims

are accessible for future research as potential hotel room cleaner claims. A list maintenance

method was applied to code occupation for 23,070 hotel industry MSD claims for years

20072016. These methods are replicable and make it possible for researchers to access the

narrative text data in the Occupation Description field and identify the occupation of hotel room

cleaner in WCIS data. Of the 23,070 hotel industry claims, 52.6% (12,125) are hotel room

cleaners and an additional 12% (2,765) are “Maybe” and “Questions.” Of hotel room cleaners,

the two most frequent job titles were housekeeper (58.2%) and room attendant (39.5%), together

accounting for 98% of the hotel room cleaner claims. The study dataset was limited to females

only. This decision was based on averaging together these two highly frequent job titles which

resulted in 96% of claimants being female.

Aim 2: Identification of Task at Time of Injury

It was possible to identify the task at time of injury by applying machine learning processing

methods to the narrative text data from the Injury Incident Description field. Training data

(sample narrative task data for each task) was provided to the AI team who refined their model in

a series of steps. The Task at Time of Injury variable is comprised of seven task categories that

originate from the definition of “Housekeeping tasks” using Cal/OSHA 3345 and answer the

question “What is the task at time of injury?” Responses classified as “unknown” account for a

106

high proportion (43.6%) of the claims for years 2012-2016. Additional codes were created to

identify subcategories of “unknown” responses for future research purposes, given the high

prevalence of such claims. Quality control checks of the task data used these “unknown” codes

and found “98,” pain/injury/worker motion/diagnosis to be the most frequent. The overall

estimated accuracy of the coded task data is 84%; 89% of claims sampled were deemed to be

related to hotel room cleaning tasks. Based on guidance from the AI team, claims categorized as

having a <70% confidence level (of being classified in the correct task category) were excluded

from the task analytic data sample. Of the coded task data sample, 18.6% fell into this category,

reducing the task data sample size further. The Aim 3B task analytic data sample includes 2,628

claims (one claim removed due to a missing injury incident description) which represents 37.8%

of hotel room cleaner MSD claims for years 2012-2016 (n = 6,952).

Aim 3A: Female Hotel Room Cleaner Claimant Demographic and Injury Characteristics

For the Aim 3A analytic data sample, years 2007-2016, there were 12,125 claims. By Nature of

Injury, “Strain or Tear” injuries accounted for close to two-thirds of the sample, followed by

“Sprain or Tear” injuries (11.9%). By Cause of Injury, “Strain or Injury By, NOC” (27.7%),

“Pushing or Pulling” (18.4%), “Repetitive Motion” (16.6%) and “Lifting” (14.5%) were the

leading causes. By Part of Body Category, “Upper Extremities” was the most frequent category

(40.5%), followed by “Trunk” (33.2%) and “Lower Extremities” (13.4%). By individual Part of

Body, Low Back Area accounts for the highest proportion of claims (27.2%), followed by

“Shoulders” (14.3%). Taken together, “Multiple Body Parts,” “Wrist” and “Knee” account for

about one-quarter of remaining hotel room cleaner MSD claims.

107

Aim 3B: Injury Variables Stratified by Seven Task at Time of Injury Categories

The Female Hotel Room Cleaner Task at Time of Injury analytic data sample, 2012-2016, has

2,628 claims for years 2012-2016. Bedmaking (43.3%), cleaning surfaces (29.9%) and cart

handling (17.2%) were the leading tasks at time of injury. By Nature of Injury, “Strain or Tear”

injuries accounted for three-quarters of the task dataset overall. By task, the prevalence of “Strain

or Tear” injuries ranged from 65.3% for vacuuming to 82.1% for moving linen/supplies. “Sprain

or Tear” injuries ranked second for all tasks, except vacuuming where it ranked third.

By Cause of Injury in the task dataset overall, the leading three causes are “Strain or Injury By,

NOC” (28.1%); “Pushing or Pulling” (24.8%); and “Lifting” (17.8%). “Twisting,” “Reaching”

and “Repetitive Motion” each account for 8% of claims. For specific tasks at time of injury,

cause of injury varied. “Pushing or Pulling” ranked first as a cause for cart handling, moving

furniture and vacuuming and ranked third for bedmaking (23.3%), moving linen/supplies and

collecting/ disposing of trash. Lifting was the most frequent cause of injury for collecting/

disposing trash (48.8%), moving linen/supplies (32%) and making beds (26.5%), followed by

moving furniture (19.7%) and handling linen carts (12.8%). Cleaning surfaces was the task with

the highest proportion of “Strain or Injury By, NOC” (45%), followed by making beds (26%),

moving linen/supplies (21%) and moving furniture (18%) where it ranked second. Cleaning

surfaces (15.7%) and vacuuming (12.5%) were tasks where repetitive motion ranked second.

Although cause of injury varied by task, it is important to recognize that hotel room cleaning

involves all seven tasks and together can have a cumulative effect resulting in injuries.

108

Upper Extremities is the most frequently reported Part of Body Category for task dataset claims

(42%), followed by Trunk (37.5%) and Lower Extremities (14%). Lower back injuries were the

leading cause of injury for all tasks with moving furniture having the highest proportion (47.5%),

followed by moving linen/supplies (42.9%), making beds (35.7%) and handling linen carts

(32.5%). Shoulder injuries ranked second for hotel room cleaners across all tasks with

collecting/disposing trash (19.8%), moving linen/supplies (17.9%), cleaning surfaces (17.15),

making beds (16.8%) and vacuuming (16.7%) the leading tasks at time of injury. Moving

linen/supplies (10.7%) and collecting/ disposing trash (10.5%) were the tasks with the highest

proportion of “Other – Lower Extremities” injuries.

Significance of Findings

This is the first time WCIS data have been sorted using the Cal/OSHA 3345 definitions as a

framework for both occupation and injury incident (task) narrative text descriptions and that list

maintenance and machine learning processing methods have been applied to WCIS data. These

methods allow for a systematic, replicable and valid approach to transforming narrative text data

into quantitative data, useable for data analysis by future researchers.

These findings establish, using state-level administrative data, that hotel room cleaners

experience MSD injuries and furthermore, account for 2.3 times as many claims (52.6%) as

would be expected given the proportion of the hotel workforce represented by this occupation

(23%).13 There are an additional 12% of claims that are potentially hotel room cleaners, the

109

“Maybe” and “Questions” claims which could bring the hotel room cleaner proportion close

to two-thirds (64.6%). This finding is supported by previous researchers identifying higher

MSD injury rates for hotel housekeepers among hotel workers, high prevalence rates for

musculoskeletal pain using survey data, high risk for low back disorders using biomechanical

evaluations and a high incidence rate using the BLS’s Maids and Housekeeping Cleaners

occupation category, among all workers with MSDs.6,7,9,10,16,46 This is a significant finding

indicating a need to track MSD injuries (surveillance) occurring to hotel room cleaners

(occupation) and to implement interventions at the employer-level (hotel industry) and

government-level (enforcement by Cal/OSHA), carried out at the hotel level (worksite and

employee).

When analyzing 2007-2016 claims (n = 12,125) by Part of Body Category (POB CAT), instead

of individual Part of Body (POB), “Upper Extremities” (40.5%) ranked first followed by

“Trunk” (33.2%), highlighting the importance of upper extremity injuries overall. When

analyzed by POB, “Low Back Area” ranked first and “Shoulders” ranked second. Similar results

were reflected in the smaller task at time of injury dataset for years 2012-2016 and are supported

by previous research.10 By analyzing the claims at the POB level, it is possible to identify

individual upper extremity categories such as “Wrist” (7.2%) and four other upper extremity

POB totaling 12.9%., triggering biomechanical studies and interventions specific to work factors

affecting these parts of body. The POB CAT lens versus the POB lens, shifts the focus from part

of body level to risk factor level, pointing to the need for a comprehensive intervention program

to prevent upper extremity MSD injuries. Such a program needs to recognizes the use of

different upper extremity POB by task and the presence of corresponding biomechanical forces

110

and impact on muscle groups. These observations by POB CAT support concerns expressed

earlier in this dissertation for further research on upper extremity injuries in hotel room cleaners.

Across the seven hotel room cleaner task categories for the Female Hotel Room Cleaner Task at

Time of Injury, 2012-2016 study population (n = 2,628), “Pushing or Pulling,” “Lifting” and

“Strains or Injury By” were the leading causes of injury, respectively. This finding is supported

by similar findings from previous research and by studies in a systematic review of pushing and

pulling forces that found an association with musculoskeletal injuries to the upper extremities,

including shoulders which are reflected in the POB finings in the preceding paragraph.11,101

These leading causes of MSD injury support the justification for the promulgation of Cal/OSHA

3345 which recognizes ten risk factors for an MSD injury including “(6) pushing and

pulling”;

“(4) lifting or forceful whole body or hand exertions”; and others that contribute to strains, “(2)

prolonged or awkward static postures; (3) extreme reaches and repetitive reaches above shoulder

height”. These findings are supported by biomechanical studies that identify lifting rate and

forward bending as key risk factors for low back injuries in hotel room cleaners.9 Given that the

leading causes of hotel room cleaner MSD injuries are addressed by Cal/OSHA 3345, these

findings also justify the elements of the standard including performance of a worksite evaluation,

injury investigation, identification of solutions and the active participation of workers and their

union representative.102 The dissertation findings also point to the need to address these causes of

injury by task when developing safe work practices and interventions and when training on them

as “pushing and pulling” occurred at high frequency for cart handling and also occurred for

bedmaking, as one of the three causes that occurred at similar frequencies.

111

Most tasks had one predominantly leading cause of injury and then substantially lower second

and third causes, followed by a sharply diminished prevalence of other causes. Bedmaking was

the only task that the first three causes of injury had equal prevalence: “Lifting” (26.5%), “Strain

or Injury, By NOC” (26.2%) and “Pushing or Pulling” (23.3%). This finding is extremely

significant because it points to three leading causes, each different than the other from a

biomechanical perspective and each raising unique research questions in search of appropriate

solutions. For example, the use of both fitted sheets and bedmaking lift tools reduced mattress

lifting by 48% in a study of hotel room cleaners.45

The use of the Cal/OSHA 3345 standard as a framework for identifying, classifying and coding

narrative text data into occupation and task at time of injury categories, the creation of an

algorithm for filtering questionable hotel room cleaning occupation terms and the application of

list maintenance and machine learning process methods applied were successful. Occupation and

task data about hotel room cleaner MSD injuries are accessible and useable for quantitative data

analysis, a significant finding of this dissertation. Study findings help point to risk factors of and

solutions to hotel room cleaning MSD injury exposures, providing valuable information needed

for injury prevention, including some never having been reported in the literature to date.

Study Strengths

This study builds on the methods of previous researchers using WCIS data and results in more

expansive findings using different methods directly relevant to hotel room cleaner work

organization for occupation and task, and in certain analyses, using larger sample sizes.11,47 The

112

importance of this study is that some of the major methodological obstacles have been identified

and ways to reduce those obstacles have been created, paving the way for further development

and refinement.

The most significant strength of this dissertation is that with support from the AI team, I was

able to develop methods to overcome the obstacles presented by the narrative text data and using

these methods, workers’ compensation claims can be used to create surveillance systems to

monitor the impact of the Cal/OSHA 3345 hotel housekeeping MSD injury prevention standard

over time. In addition, these methods can be used to further the research by injury variables and

task that was begun here. Important information about an occupation with a high prevalence of

MSD claims was accessed and gives insights into causes and parts of body injured which may

trigger intervention studies.

This dissertation is innovative in several ways in how the methods were applied to the

occupation and task narrative task data. First, definitions from Cal/OSHA 3345 Hotel

Housekeeping Musculoskeletal Injury Prevention standard were used as the framework to sort

occupation terms by the hotel room cleaner job titles. Where there were questions about sorting

the occupation terms into hotel room cleaner job titles, an algorithm was created to triage

occupation terms deemed “Maybe” or “Questions” in separate categories and not included as

hotel room cleaners but accessible for future research. Second, to the best of my knowledge, this

is the first study of hotel room cleaner California WCIS claims applying list maintenance

methods to occupation terms and transforming narrative text injury incident description data

using machine learning process methods into structured task data that can be used for

113

quantitative analysis. These methods are systematic, replicable and valid (Appendix I). Another

study strength is that starting the data analysis at the hotel industry level, followed by the hotel

room cleaner level and then by the hotel room cleaner task level allowed for important study

findings from a macro-level to a micro-level:

• the over-representation of hotel room cleaners (52.6%) in the Female Hotel Industry

MSD Claims, 2007-2016 study population (n = 23,070) compared to accounting for

23% of a hotel’s workforce, a 130% higher representation;14

• “Upper Extremities” identified as the leading part of body category for the ten-year

Female Hotel Room Cleaner Occupation MSD Claims, 2007-2016 study population (n =

12,125);

• “Pushing or Pulling,” followed by “Lifting’, and “Strains or Injury By” as the leading

cause of injuries across seven hotel room cleaning tasks for the Female Hotel Room

Cleaner Task at time of Injury, 2012-2016 study population (n = 2,628); and

• Bedmaking as the only task where the first three causes of injury had equal prevalence:

“Lifting” (26.5%), “Strain or Injury, By NOC” (26.2%) and “Pushing or Pulling”

(23.3%).

A strength of this study is that by following SMEs advice on creating an “unknown” response

code, it is possible to address the limitations of excluding this data in the Aim 3A and 3B

analyses. Having already created codes for subcategories of the “unknown” responses and

having coded them by task at time of injury, it will be possible for future research to analyze

these claims using the same variables as in Aims 3A and Aim 3B.

Similarly, having created codes for potential hotel room cleaner claims, using “Maybe” and

114

“Questions” code and having coded them by task at time of injury, it will be possible for future

research to analyze these claims using the same variables as in Aims 3A and Aim 3B to see how

these claims may differ from the rest of the dataset or on the contrary, identify the associated

occupation terms as belong to the hotel room cleaner category and increasing the proportion of

hotel room cleaners with MSD claims to 64% of the hotel industry WC dataset, close to triple the

expected proportion given the previously stated representation of hotel room cleaners in a hotel

workforce of 23%.14

Study researchers have recommended manually coding data with a 30% confidence level as MLP

modes are less successful coding such data.94 Although the dissertation timeline did not allow for

such manual coding, a strength of this study is that the AI team and I addressed this by excluding

the 30% confidence level claims from the dataset for analysis. An easy first step to improve the

quality of data in the task dataset and to increase the sample size of the task dataset, would be to

code these claims.

Study Limitations

Limits on the sample size of the Aim 3B task analytic data sample. The “Task at Time of Injury”

data sample size was limited by: a) the number of years in the study period; b) the proportion of

the claims classified into task categories that were in the <70% confidence level category and

were excluded from the task analytic data sample; and c) the proportion of the claims classified

into task categories that were coded “unknown” and were excluded from the task analytic data

sample. These limitations introduce bias for the latter two limitation. For claims with <70%

115

confidence level cannot be expected to occur equally across all task categories as the quality of

the data determines the likelihood of accurate classification resulting in more easily identifiable

narrative text, e.g., bedmaking will be classified and less easily identifiable narrative text will

have a lower confidence level, e.g., cleaning surfaces. In the task dataset this occurred with a

lower percentage of claims with “Cleaning Surfaces” included after removing the <70%

confidence interval claims. This results in a skewing of the representation of claims with

“Cleaning Surfaces” as a task at time of injury category, resulting in an overrepresentation of

bedmaking claims. For claims with “Unknown” response codes that fell into the category of

medical diagnoses, one could theorize that these are more severe injury cases since a medical

diagnosis has been identified and there is a bias that they are being excluded for task data

analysis. Both of the examples, point to the value of identifying the proportions of the dataset

represented by the different “unknown” categories and the possible recoding of the claims

manually, as suggested in the literature for more challenging narrative data.60,86

Another major limitation was that occupation and task were narrative text data fields requiring

extensive work to transform the data usable for data analysis. Developing methods, creating

training data and applying list maintenance and machine learning processing techniques (Section

3 Methods) was time-consuming, slowing down the research timeline. In addition to data

transformation, performing the corresponding quality control checks of the data classified into

task categories at each step was a lengthy process.

Limitations of extrapolating study findings to the overall hotel room cleaner population in

California. This is a self-selected population – workers who filed a workers’ compensation

claim. Moreover, there is substantial under-reporting of work-related injuries to employers by

116

hotel room cleaners and an even lower rate of filing workers’ compensation claims by this

occupation group, with one-third of claims rejected.5-7 Thus it is not known how generalizable

these findings are to the overall hotel room cleaner population in California.

Time and funding constraints on improvement of the machine learning process model. With more

resources, a more effective model may be developed resulting in more claims with a ≥70%

confidence level. This would increase the task sample size and present findings that reflect better

the actual claims data. Currently, there is an under-representation of the cleaning tasks due to that

task having a high percentage of claims with <70% confidence level. Manually coding claims

with a ≤30% confidence level is a recognized best practice and is an alternative to an improved

model.94 Also, the correct classification for these claims is easy to identify and recode.

Lessons Learned

Using lessons learned from carrying out this study, approaches to address the study limitations

for future analyses are detailed below.

Expand the number of years in the study period to 2007-2016, increase the analytic sample size

The purpose of limiting the task study period to 2012-2016 was so that the injury incident data

would reflect the most recent five years of hotel accommodation trends and work practices that

result in injury, e.g., thick duvets, high thread-count sheets and more pillows. Although the

reasoning for limiting the data to the most recent five years was valid, in practice, the injury

incident descriptions lacked the level of detail needed to detect recent hotel industry luxury

117

accommodations and work practices. Therefore, adding earlier years of data is recommended.

Estimating the sample size for ten years of data involves multiple steps. Before excluding the

claims with <70% confidence level and the claims coded as “unknown,” the 5-Year Female

Hotel Room Cleaner MSD Claims dataset includes 6,952 claims. After removing 3,013 claims

coded as “unknown” (43%), 3,939 claims remained. Of these, 1,310 (33%) claims with a <70%

confidence level of task classification were removed, leaving 2,629 claims. If we apply the same

steps above to the ten-year “Female Hotel Room Cleaner MSD Claims” study population

(n=12,125), we can estimate that 43% of claims would be classified as “unknowns” (n = 5,214)

and removed. Of the remaining 6,911 claims, there are 2,280 (33%) claims estimated to be

removed for having <70% confidence level. This leaves an estimated 4,631 claims for the Aim

3B task analytic data sample years 2007-2016, a 43% increase over the 2012-2016 sample size (n

= 2,629). This provides a substantial increase in sample size for data analysis for future research.

De-identification of the WCIS data by CA-DWC as part of providing data extraction

De-identification of the numeric data could be performed by the CA-DWC with guidance from

the researcher, e.g., request that a tenure variable be created using date of hire and date of injury

variables, and provide a sample to the researcher with the original variables and the new tenure

variable results. Of the two de-identification issues mentioned above, the de-identification of the

narrative text data was the most time consuming and also the most disconcerting because of the

failure to meet HIPPA’s Safe Harbor requirements to protect confidential information about the

employee.97 For the numeric data, multiple steps in SAS (Section 3 Methods) to create new

variables (e.g., tenure) and then remove original variables, e.g., date of injury, date of hire,

birthdate, zipcode. De-identification of the narrative text included removing gendered terms,

118

diagnoses, names, ages, and court case numbers at the most basic level. Additional steps were

taken to maintain the integrity of the data. Claims with “sheets” in the injury incident description

were removed before the rest of the dataset was de-identified for the term “she”. Next, claims

with “sheets” were de-identified separately and when returned to the dataset. This way, the word

“sheet” remained intact and did become “et,” allowing for future research on fitted sheets or

changes in tucking styles.

De-identification of the narrative text data should be performed before providing data extraction

to researchers. None of the data I listed under narrative text data is necessary to be obtained from

the narrative text injury incident description field. Diagnoses is collected in a dedicated question

on the FROI and is obtained through a separate request from the agency. Although de-identifying

the data would add additional costs to the data extraction charge, I believe most researchers

would consider it worth the extra expense and I would expect that the CA-DWC programmers

would be more qualified to perform the de-identification than the average researcher.

Improving the Narrative Text Occupation and Injury Incident Descriptions Data

The “Occupation Description” data can be improved substantially by adding a new variable,

“SOC Code,” as detailed in Table 5 and is supported by the literature.91

The Injury Incident Description data would be improved by the CA-DWC (the “agency”)

implementing a training and monitoring program for employers on how to complete the

questions detailed further under the next heading. The agency would screen industries with

higher rates of employers not filling out the Injury Incident Description field correctly. The

agency would train these employers on proper ways to provide the information and track the

119

employer’s forms for a certain number of years to improve their data. By targeting the hotel

industry and identifying what size employers are most frequently failing to complete the “Injury

Incident Description” field correctly, CA-DWC could have an impact on improving the quality

of the narrative text data at the hotel industry level and not just at the employer level. This is key,

because short of creating a list of tasks for each industry and creating a “Task at Time of Injury”

variable to use along with the narrative text injury incident information, researchers are

dependent on this text field for key information. The task framework from my dissertation is an

example of a systematic, replicable and valid method, limiting bias from subjective judgement.

Conclusions

This study demonstrates that WCIS data can serve as surveillance data, although with serious

challenges due to two key data fields comprised of narrative text: “Occupation Description” and

“Injury Incident Description.” This dissertation proves that California WC data, with the

application of certain research methods and MLP techniques, can be used to identify at risk

occupations and tasks that need to be addressed by workplace and government interventions.

Importantly, this study supports that hotel room cleaners are an occupation with increased

prevalence of MSD injuries, 2.3 times the expected proportion when compared to the percent of

the hotel workforce they represent. Specific tasks are broken down by cause of injury, providing

new insights on risk factors and job hazards in need of solutions and development of

interventions. This dissertation also demonstrates the multiple causes of injury of each of the

120

seven hotel room cleaning tasks, highlighting three key causes: straining, lifting and pushing or

pulling. These findings support the inclusion of these three key causes as risk factors in

Cal/OSHA 3345 and provides further justification overall for the promulgation of this standard.

Cal/OSHA 3345 requires hazard control programs and interventions by the hotel industry and

government agency enforcement by the California Department of Occupational Safety and

Health.1,103 This analysis gives insight to the multiple exposures experienced by hotel room

cleaners as all seven tasks comprise hotel room cleaning work and can have a cumulative impact

on injury causation.

The study findings are supported by peer-reviewed literature published over the past twenty

years, along with biomechanical studies and hotel room cleaner and union organizer testimony

for the past fifteen years about the hazards of hotel room cleaning and recommended solutions.

Fortunately for the California hotel room cleaner workforce, their participation and that of their

union representatives in identifying hotel room cleaning hazards and solutions is codified in

Cal/OSH 3345, the only state with such a regulation. This dissertation contributes: 1) new

methods to access key information for regulators and researchers to protect hotel room cleaners

from MSD injuries and 2) new findings to guide employers to comply with Cal/OSHA 3345 and

to assist Cal/OSHA in enforcing its comprehensive state-wide regulation.

Future Direction

WCIS data provides useful and important information about MSD injuries including injury

characteristics and task at time of injury. Improvements in how the data is collected by the CA-

DWC is needed. Study findings indicate the need for further intervention research

121

Policy Recommendations for Improving CA-WCIS Data for Prevention Research

Four main problematic issues were encountered in the WCIS data that need to be addressed:

The lack of standardized occupation codes for the claimant. The lack of occupation codes limits

the ability of researchers to access the data to for quantitative data analysis and do so in a

uniform and systematic way. Methods used to ameliorate the problem of narrative text data are

time-consuming, costly to the researcher in human and financial resources, can introduce errors,

are dependent on the subjective judgement of the researcher and limit the comparability of

findings by different researchers who apply unique frameworks for sorting and/or coding the

narrative text occupation data. Even though occupation terms from Cal/OSHA 3345’s definitions

provided a practical framework for sorting 2,000 occupation terms into about 200 hotel room

cleaner occupations and five related job titles, this method is an example of a state-by-state

solution. Currently, this type of research is limited to individuals who have the time and the

resources to perform the time-consuming tasks explained in Section 3 Methods.

Lack of information about the activity/task at time of injury. Questions 22 through 26 from Form

5020 (Appendix B) ask about the event/exposure, i.e., location, equipment/materials/chemicals,

activity and sequence of events. Answers to these questions, when taken together, can provide a

robust description of the activity/task at time of injury. Yet despite these questions, 43% of the

6,952 hotel room cleaner claims coded for task were coded as “unknown.” This study finding is

30% higher than findings that reported 33% unknown for hotel room cleaner injury descriptions

using WCIS data.47 Although the employer took the time to fill out the field, the information

122

provided was not pertinent and hence, not usable. Numbers and percentages are available for a

breakdown of “unknowns” by response codes for quality control samples.

The three main categories of injury incident description data coded as “unknown” are: a) use of

generic phrases, e.g., “while cleaning a room, employee hit head …..” or “during regular duties,

employee hurt back…..,”; b) stating “pain” or part of body; and c) stating the diagnosis, all of

which do not without provide any additional information.

Inclusion of identifiable information in the injury incident descriptions field. Despite the

requirements of HIPPA’s Safe Harbor clause (Section 3 Methods) to protect the employee’s

identity, three main categories of sensitive information about the claimant were encountered in

the injury incident description data (in order of frequency): a) gendered terms, e.g., “she,”

“her,”.; b) medical diagnoses; and c) names (worker/supervisor/ hotel/clinic), injury date, court

case numbers and employee age.97 Even the Employer's Report of Occupational Injury or Illness

Form 5020 (Appendix B) uses gendered terms in its own example for question 26.

Below are specific recommendations for each policy issue identified (Table 5).

Table 5. Policy Recommendations for Improving CA-WCIS Data for Prevention Research

Issue Specific Policy Recommendation

1.The lack of standardized occupation a. Utilize the Office of Management and Budget’s

codes for the claimant Standard Occupational Classification (SOC) system

codes to provide uniform, quantitative data. See

SOC code 37-2012, Maids and Housekeeping

Cleaners (Section 1 Introduction).

b. Maintain use of WCIS Data Element No. 60

Occupation Description (Appendix E). SOC-coded

data can provide first-tier information to identify

123

the occupation of Maids and Housekeeping

Cleaners. Narrative text field can provide secondtier

information, e.g., job titles that may represent

differences in work organization factors that impact

the injury experience, e.g., night cleaners (night

shift), housepersons (extra dirty rooms).

2. Lack of information about the a. DWC must educate employers on the

importance activity/task at time of injury of questions 22-26 and provide training

materials on how to complete those data fields, including examples of correct and

incorrect responses.

b. DWC needs to train employers on the correct

field to enter a POB as a response (question 19).

c. DWC needs to train employers on the correct

field to enter a medical diagnosis as a response

(question 19).

3. Inclusion of identifiable information DWC needs to educate employers: in the

injury incident descriptions field: a. on the content of the Safe Harbor clause;

b. on the importance of not including gendered

terms and medical diagnoses information in the

injury incident description field; and

c. by providing employers examples of correct and

incorrect responses.

Intervention Recommendations for Prevention of Hotel Room Cleaning Injuries

Three key intervention recommendation categories were identified:

Full Implementation of Cal/OSHA 3345 by Hotel Employers and Enforcement by Cal/OSHA

Cal/OSHA 3345 (Appendix H) is designed to prevent hotel room cleaning MSD injuries and

does so through key elements, including: an injury and illness prevention plan, a worksite

124

evaluation, identification of risk factors and solutions, injury reporting, communication program,

training of housekeepers and supervisors, posting worksite evaluation results and maintain

records and making them available to workers and their union representatives. Many of the

elements require the active participation of hotel housekeepers and their union representatives.

The first intervention I recommend to focus on is the widespread implementation of this

comprehensive MSD injury prevention standard by hotel employers, along with the active

enforcement of the standard by Cal/OSHA. Although additional interventions are needed specific

to hotel room cleaning tasks, there is sufficient information available for employers to comply

with Cal/OSHA 3345 standard, including resources in the standard’s non-mandatory Appendix A

(Appendix B), an intervention study on bedmaking and a recently published book chapter on

ergonomic hazards in the hospitality industry focusing on key job titles.45,104

Extensive Training at Key Intervals in a Hotel Room Cleaner’s Tenure

This study found a high percent of claims for newly-hired employees with ≤6months (18%)

tenure and for low-range experienced employees in the 2-5 years (21.3%) tenure groups. A

review of existing risk factors, followed-up by an inventory of interventions, and implementation

of trainings programs is needed to target these two employee groups to prevent MSD injuries

Students also viewed