Confounder or Effect Modifier

Excelsior College PBH 321

Page 1

BIAS IN EPIDE MIOLO GY

THE MEANING AND CONTEXT OF BIAS

We briefly touched upon bias earlier in the course. The word bias probably has an intuitive meaning to you, and recall the definition from Module 3:

“Deviation of results or inferences from the truth, or processes leading to such deviation. Any trend in the collection, analysis, interpretation, publication, or review of data that can lead to conclusions that are systematically different from the truth.” (J.M. Last, A Dictionary of Epidemiology, 4th ed.)

Bias is a systematic error that results in an incorrect (invalid) estimate of the measure of association. Bias can therefore create the appearance of an association when there really is none (bias away from the null), or mask an association when there really is one (bias towards the null). Bias can arise in all study types: experimental, cohort, and case-control designs. It is primarily introduced by the investigator or study participants in the design and conduction of a study, and once it occurs, cannot be removed – but it can be evaluated during the analysis phase of a study. Bias is a direct threat to the validity of any epidemiologic study, and influences whether we can believe the observed association is a true association.

DIRECTION OF BIAS

We can describe the impact of bias on the measure of association in a few different ways. Remember, the null value is 1.0; the value that means there is no observed association between exposure and disease. If the true association is protective (less than 1.0), bias towards the null will make it seem less protective as the measure of association gets closer to the null value.

• Positive bias: the observed value is higher than the true value • Negative bias: the observed value is lower than the true value • Bias towards the null: the observed value is closer to 1.0 than the true value. • Bias away from the null: the observed value is farther away from 1.0 than the true value.

Bias towards null

Bias away from null

Bias away from null

Bias towards null

Protective effects Harmful effects

1.0 (no association)

Bias towards null

Bias away from null

Bias away from null

Bias towards null

Protective effects Harmful effects

1.0 (no association)

Excelsior College PBH 321

Page 2

TYPES OF BIAS

Two main types of bias are selection and information (observation) bias. You had an introduction to these biases in Module 4. 1. Selection Bias Selection bias is error that occurs if selection of subjects into a study is related to both exposure and disease. As a result, the measure of association differs from what would have been obtained if we could examine the entire population targeted for study. Selection bias is most likely to occur in case-control or retrospective cohort studies because exposure and outcome have occurred at time of study selection, and may influence the willingness of an individual to participate in a study. Selection bias generally arises in several ways:

• Selection of inappropriate control group (case-control study) • Participation bias: response or non-response may be related to exposure and/or disease • Surveillance/diagnostic bias: likelihood of detecting disease differs between exposed and unexposed,

such as if exposed individuals are more carefully examined Selection bias in a case control study Occurs when controls or cases are more (or less) likely to be included in study if they have been exposed -- that is, inclusion in study is not independent of exposure. As a result, the relationship between exposure and disease observed among study participants is different from relationship between exposure and disease in individuals who were eligible but not included. The odds ratio from a study that suffers from selection bias will incorrectly represent the relationship between exposure and disease in the overall study population.

Example: Do PAP smears prevent cervical cancer? Imagine a case control study where cases are diagnosed at a city hospital. Controls are randomly sampled from households within same city by the investigator canvassing the neighborhood on foot. Only controls that were at home at the time the researchers came around to recruit for the study were actually included in the study.

Source population Controls only

women who were home

Cases Controls Cases Controls

Exposed 100 150 Exposed 100 100

Unexposed 150 100 Unexposed 150 150

Total 250 250 Total 250 250

True OR=.44

Biased OR=1.0

Excelsior College PBH 321

Page 3

Interpretation of true OR: There is a 56% reduced risk of cervical cancer among women who had Pap smears as compared to women who did not. 40% of cases had PAP smears versus 60% of controls. Interpretation of biased OR: There is no association between PAP smears and the risk of cervical cancer. Here, 40% of cases and 40% of controls had PAP smears. There were ramifications from using only women who were at home during the day as controls – it resulted in a selection bias. Women at home were less likely to work and less likely to have regular checkups and PAP smears. They did not accurately represent the distribution of exposure (likelihood of getting a PAP smear) in the study population that produced the cases, and so they gave a biased estimate of the association. Selection bias in a cohort study In a cohort study, selection bias occurs when selection of exposed and unexposed subjects is not independent of outcome (and so can only occur in retrospective cohort study).

Example: A retrospective study of an occupational exposure and a disease in a factory setting. The exposed and unexposed groups are enrolled on the basis of prior employment records. The records are old, and many are lost, so the complete cohort working in the plant is not available for study. If people who did not develop disease and were exposed were more likely to have their records lost, then there will be an overestimate of association between the exposure and the disease. Assume each person is followed for the same length of time with no loss to follow-up (remember: since time is equal in both groups, you may calculate a risk ratio from the cumulative incidences, not a rate ratio where you incorporate follow-up time).

The left side of the table below shows the true relationship if all records were available. Let’s say that 200 records were lost, all of them among exposed individuals who did not get the disease. The result is a positive bias of the RR.

All records available 200 records lost, all

among E+, D-

Diseased Non-

diseased Total Diseased

Non- diseased

Total

Exposed 50 950 1000 Exposed 50 750 800

Unexposed 50 950 1000 Unexposed 50 950 1000

True RR=1.0 Biased RR=1.25

Selection bias can also occur in prospective cohort and experimental studies from differential loss to follow- up. For example, this would be the case if loss to follow-up over the study period is due to getting the disease (say, because subjects are too ill to continue participation) and occurs only among exposed or only among the unexposed populations (i.e., loss to follow-up is different by exposure status).

Excelsior College PBH 321

Page 4

Little or nothing can be done to fix selection bias once it has occurred. You need to avoid it when you design and conduct the study, for example, by using the same criteria for selecting cases and controls, obtaining all relevant subject records, obtaining high participation rates, and taking in account diagnostic and referral patterns of disease. 2. Information or Observation Bias Remember that besides the experimental design, epidemiological studies are observational in nature. Information bias is an error that arises from systematic differences in the way information on exposure or disease is obtained from the study groups. This bias results in participants who are incorrectly classified as either exposed or unexposed, or as diseased or not diseased. This type of bias occurs after the subjects have entered the study. There are several types of information bias: recall bias, interviewer bias, and misclassification. Recall bias occurs when people with disease remember or report exposures differently (more or less accurately) than those without disease. Recall bias can result in an over- or under-estimate of measure of association. We typically are most concerned about recall bias in a case-control study or a retrospective cohort design, where disease has already occurred and participants are asked to recall their previous exposure history.

Example: Recall bias in a case-control study. Illness affected the cases’ memory, so they were less likely to report their exposure than controls.

Truth Observed study data

Cases Controls Cases Controls

Exposed 40 20 Exposed 30 20

Unexposed 60 80 Unexposed 70 80

Total 100 100 Total 100 100

True OR = 2.7 Biased OR = 1.7

Because cases were less likely to report their exposure, the OR produced from this study is substantially smaller than the true OR. Ways to avoid recall bias in a case-control study:

• Use controls who are themselves sick • Use standardized questionnaires that obtain complete information • Blind subjects to study hypothesis

Interviewer bias is a systematic difference in soliciting, recording, and interpreting information. This can occur if exposure information is sought when the outcome is already known (such as in a case-control study), or when outcome information is sought when exposure is known (as in a cohort study). Knowledge of the subject’s exposure or disease status may affect how information is obtained. This type of bias can be avoided by blinding interviewers to the study hypothesis and disease or exposure status of subjects, and use of

Excelsior College PBH 321

Page 5

standardized questionnaires or standardized methods of outcome (or exposure) ascertainment, such as a clinical diagnostic test. Misclassification is a type of bias where a subject’s exposure or disease status is erroneously classified. Misclassification may be non-differential and differential. We will cover only the more common form: non- differential misclassification. Non-differential misclassification is inaccuracy with respect to disease classification that is independent of exposure, or, inaccuracies with respect to exposure are independent of disease. This type of misclassification will generally bias the measure of association towards the null. In other words, non-differential misclassification makes the groups appear more similar, and attenuates the true measure of association. When interpreting study results, ask yourself these questions …

• Given conditions of the study, could bias have occurred? Think about the study design and how information was collected.

• Is bias actually present? Are there any clues as to possible bias? • Are consequences of the bias large enough to distort the measure of association in an important way? • Which direction is the distortion? –is it towards the null or away from the null? How does this affect

our interpretation of the study’s results?