Critique of two research articles

Michelle_Michy

Lecture3Non-experimentalobservationaldesigns.pdf

Home >Education homework help >Critique of two research articles

Non-experimental (observational) designs

Neil O’Connell PH2600 Research Methods

Learning outcomes By the end of the lecture students should be able to:

 Describe basic common observational study designs

 Describe the concept of correlation

 Consider some of the biases that we attempt to control for

It’s a complex world

 Clinical phenomena and human behaviour are complex

 Research seeks to unravel the complexity

 But experiments are often not possible or appropriate… because…

Many naturally occurring variables can’t be

manipulated

Many variables shouldn’t be manipulated (ethics)

The research question does not require manipulation of

variables (descriptive/ exploratory)

Experiments are simply not feasible

DESCRIBE THE CHARACTERISTICS

OF A CERTAIN GROUP

DRAW COMPARISONS

BETWEEN DIFFERENT GROUPS

ANALYSE RELATIONSHIPS

BETWEEN NATURALLY OCCURRING VARIABLES

(CORRELATIONAL DESIGN)

INVESTIGATE THE INFLUENCE OF

VARIABLES THAT CANNOT BE

EXPERIMENTALLY MANIPULATED

INVESTIGATE THE NATURAL COURSE OF A CONDITION

INVESTIGATE RISK FACTORS FOR A

CERTAIN OUTCOME

Researchers might want to….

Hierarchy of study design

Experimental & quasi-experimental studies

Systematic reviews (+/- meta- analysis)

Randomised Controlled Trial (RCTs)

Controlled clinical trials

Observational studies

Cohort

Case-control

Cross-sectional

Case-report, case-series Expert opinion etc.

Generate hypotheses

Establish causality

Descriptive

Describe the sample characteristics,

behaviours, conditions

Exploratory/ Analytical

Systematically investigate relationships

between variables

Hypothesis-driven

Correlational designs

Correlation designs

Analysis of relationships:

correlation research to understand:

1. Whether patterns exist within data.

2. Whether and how variables are related.

Correlation  Variables often co-vary

 A variable rises or lowers in relation to the behaviour of another variable

 These variables might be described as “related” or “associated” with one another

 We measure these associations statistically

The problem of causality

www.xkcd.com

Study Population

Cross-sectional

Prospective Retrospective

Longitudinal

Basic model

• aim is to measure (quantify) the relationship between exposure & outcome

• cannot determine whether causal link – need experimental design

EXPOSURE

Risk factor

Determinant

Treatment

OUTCOME

Disease

Mortality

Health

Cross-sectional studies

• Simplest form of observational study

• Describes population at one point in time ‘snapshot’

• Exploration of factors associated with a disease or condition (risk and/or protective factors)

• Can compare subgroups within a population

Typical research question “What proportion of the population have characteristic x?”

Typical cross sectional study questions

What are the characteristics of this

population?

What proportion of the population have characteristic X?

What variables are related to

characteristic X

Do people with characteristic X differ from those who don’t?

Important to consider the temporal association between exposure & outcomes

PAST PRESENT FUTURE

EXPOSURE + OUTCOMEEXPOSURE

X sectional study

TIME

Cross-sectional study: example Elliot et al. Lancet 1999, 354: 1248-52

Research question: what proportion of the general population in North East Scotland have chronic pain?

Sample: - random sample of adults registered with GP in Grampian region (~5000) - stratified sample by age & sex

Method: Postal self-completion questionnaire, chronic pain grade questionnaire , standard definitions

Results: 46% popn had chronic pain, major problem in general practice

Cross-sectional studies

Strengths

- quick to conduct, relatively cheap

- allows exploration of multiple risk factors

- useful to assess health needs of population

- useful first step: hypothesis generating – may

provide clues & justification for further

investigation

Cross-sectional studies

Weaknesses

- selection bias (selection of sample, responders etc)

- cannot prove cause – single point in time (snapshot)

Cohort studies

• Observes people over time - can be for weeks, months or years

• Exposures (agents/risk factors) & outcomes are recorded

 Purpose  To understand the natural history of a disease  To evaluate long-term patient health or quality of

life  Analyse the relationship between risk factors &

disease

Cohort studies Cohort = group of people who share common exposure (e.g. year of birth)

= Latin legion of soldiers

• E.G. start with people without the disease, document exposures, follow through time

• Includes those with & without risk factors

• Determines incidence (new cases) of disease

Research question Whether exposure to factor ‘x’ leads to disease ‘y’?

OUTCOME PRESENT

NO OUTCOME

OUTCOME PRESENT

NO OUTCOME

PO PU

LATIO N O F IN

TER EST

PRESENT FUTURE

MEASURE EXPOSURE (S) AT

BASELINE

EXPOSED

NOT EXPOSED

MEASURE OUTCOMEOF INTEREST

SAMPLE

HEALTHY GROUP

ACUTE DISEASE

WHO DEVELOPES DISEASE?

WHO RECOVERS?

WHO BECOMES CHRONIC?

CHRONIC DISEASE

WHO RECOVERS?

INCEPTION COHORT

‘Landmark’ cohort studies  UK Doctors Cohort Study (1950s) - 40,000 UK doctors

 USA Nurses Cohort Study (mid-1970s) >121,700 female nurses (Oral contraceptive and breast Ca)

 UK Oral Contraceptive Pill cohort study(1968 - present)

- 46,000 women , all cause mortality

 women >50 years (1 in 4 women in UK) HRT, Ca, other disease

UK Doctors Cohort Study

CAUSE‐SPECIFIC MORTALITY (LUNG CA)

ALIVE/ DEAD OF OTHER CAUSE

CAUSE‐SPECIFIC MORTALITY (LUNG CA)

ALIVE/ DEAD OF OTHER CAUSE

1950 2001

MEASURE EXPOSURE (S)

TOBACCO USE

NO TOBACCO

MEASURE OUTCOMEOF INTEREST

UK DOCTORS

Tobacco & mortality 1951: Smoking questionnaire from 40,000 British doctors (Doll & Hill)

Assessed: 1951, 1957, 1966, 1971, 1978, 1991, 2001....

Analysis • Excess mortality in smokers • Reversal of effect in ex-smokers • Other diseases eg IHD, respiratory disease, CVD, bladder

cancer, etc.

BMJ paper, Doll et al. 2004:

cessation at age 30 gain 10 years life expectancy age 40 gain 9 yrs age 50 gain 6 yrs age 60 gain 3 yrs

Dose-response relationship

- evidence for causality (one of Bradford-Hill’s criteria)

Strengths of cohort design

• Can directly estimate risk and relative risk of disease in a population

• Can determine the natural history of a condition

• Can calculate incidence

• Can demonstrate temporal sequence

• Analysis of many different exposures/outcomes

• Less prone to bias than case-control and x-sectional

studies

ATTRITION BIAS

e.g. loss to follow up (not at random)

SELECTION BIAS

e.g. unrepresentative

population

MEASUREMENT OR CLASSIFICATION

BIAS

e.g. poor standardisation,

unblinded assessors, changes

in diagnostic criteria

Weaknesses of cohort studies

Case-control studies  Start with people with disease [outcome]

Two groups: - Cases those with disease - Controls those without disease

 Collect data on risk factors [exposures] of interest

 Compare odds of exposure in two groups

 Research question

“ Are those with disease “x” more likely, than those without disease x, to have been exposed to risk factor “y”?

Key features

• Observational design ~ hypothesis generating

• Identification of subjects with disease X (or other outcome of interest)

• Identification of suitable control group without disease X

• For both groups, past exposure is then determined

• Retrospective or historical in nature of enquiry

CASES (with disease)

CONTROLS (disease free)

RISK FACTOR

NO RISK FACTOR

RISK FACTOR

NO RISK FACTOR

PO PU

LAT IO

N O

F IN T ER

EST

PAST PRESENT

DIRECTION OF ENQUIRY (RETROSPECTIVE)

Historical example: case-control study

• 1960—1961 in Hamburg, obstetricians noted 27 infants born with severe malformations

• Rare condition, no cases observed between 1930–1958

• German study recruited mothers of affected infants & sample of mothers of healthy infants

• Detailed investigation of behaviours & practices during pregnancy; documentation of multiple exposures in both groups of women

Historical example: case-control study

• Identified higher use of drug thalidomide in cases

• Sleeping pill introduced in late 1950s, safer than barbiturates – did not induce coma

• Prescribed for pregnant women, anti-emetic taken between 27–40th week of pregnancy

• Drug had been used in 46 countries - led to overhaul of drug development & licensing, withdrawn 1961

• http://www.guardian.co.uk/society/2012/sep/01/thali domide-cover-up

Mothers of babies with birth

abnormalities (n=100)

Mothers of healthy babies (n=200)

Thalidomide

No Thalidomide

Thalidomide

No Thalidomide

SA M

P LE

O F M

O T

H E R

DIRECTION OF ENQUIRY (RETROSPECTIVE)

OUTCOME (DISEASE)EXPOSURE

Selection of cases & controls Most critical stage in a case-control study

Selection of cases - clearly define ‘cases’ - specify inclusion & exclusion criteria - where is the ‘source’ population? e.g. residents from geographical region, patients recruited from a hospital ward or clinic.

Selection of controls - same source as the cases - random selection of controls from source population - or matching (age, sex, SE status.....) - increase comparator group e.g. >2 controls : 1 case

Case-control contd.

Data collection & measurement

Collecting information about the past… - questionnaires / face to face interviews - GP or hospital records - medical examination - occupational records - or other previously collated data e.g. blood tests, biological markers

- May be missing or incomplete

Strengths of case-control design

• Very useful for investigating rare diseases or diseases with a long induction period e.g. some cancers

• Time and cost efficient – when compared to cohort studies

• Existing records /database can be used

• Permits investigation of multiple risk factors /exposures

Weaknesses: case control

SELECTION BIAS (poor case

definition, difficult finding

representative controls)

RECALL BIAS (incomplete,

unreliable recall of past events,

incomplete records, worse without

blinding)

OBSERVER BIAS (awareness of hypothesis, exposure or

outcome status)

CONFOUNDER BIAS (controls not matched on important

variables, or unknown variables

not included in analysis)

VERY PRONE TO SYSTEMATIC ERROR

EFFICIENT BUT VULNERABLE

Study designs

Always consider the timeline

temporal association between exposure & outcomes

past present future

Case control

Cohort

Cross sectional

exposure

outcome

Exposure &

outcome

Confounding:

The weakness of all non- randomised studies

Confounding may increase/decrease or totally manufacture any observed effect.

How to control for confounding:

 Prevent at design stage: ◦ - restrict, randomly select, match

 Deal with at analysis stage: ◦ - stratification, multivariate modelling ◦ but only for things you know about

TEMPORAL SEQUENCE?

STRENGTH OF ASSOCIATION?

CONSISTENCY OF

ASSOCIATION?

BIOLOGICAL GRADIENT

(DOSE RESPONSE)?

SPECIFICITY OF ASSOCIATION?

BIOLOGICAL PLAUSIBILITY?

COHERENT WITH EXISTING

KNOWLEDGE?

Experimental evidence (RCT?)

Causality: Bradford Hill Criteria