EP813- Selection Bias

Diagnostic Bias- How to avoid it

- Account for referral and diagnosis patterns for outcome in selection criteria for cases and controls - Restrict study to severe disease

Selection Bias

- Affects who makes it into the 2x2 table - Who from the general population is placed into boxes of the 2x2

Healthy Worker Effect- How to avoid it

- Compare workers in a specific job with workers in other jobs that differ in occupational exposure - Be cautious that reference workers do not have own characteristics that may affect outcome

Differential Participation(Self-Selection Bias) How to avoid it

- Conceal hypothesis - Obtain high participation rates among cases and controls - Can assess likelihood by comparing characteristics of non-participants and participants (if possible)

What are some effects of bias?

- Diminish or exaggerate the actual association - Create association when there actually is none Hide an association that actually exists

How to mitigate Diff Participation in CC studies

- Don't give subjects hypotheisis of the study. - Obtain high participation rates from both cases and controls...Higher participation, smaller amount of bias

Differential Loss to follow up occurs:

- Error resulting from attrition related to both exposure and disease - Differential selection out of study We know exposure status, but loss of contact prevents ascertainment of outcome - In reality, nearly impossible to determine if losses are differential

Diagnostic Bias- How it occurs

- Exposure to determinant itself leads to increased diagnostic investigations and increases probability that outcome is identified - Probability of outcome diagnosis (and thus case selection) increased if exposed

Healthy Worker Effect- How it happens?

- Occurs when general population is selected as reference to a working population - Morbidity/mortality rates among working population generally lower than among general population

When does selection bias occur?

- Occurs when the probability of selection into or out of a study is related to both exposure status and outcome status (differential selection)

Mechanisms of Selection Bias: cohort studies

- Self-selection bias (Volunteer bias) - Healthy Worker Effect - Differential Loss to Follow-up

Roles of controls

- To provide distribution of exposure in the source population

Control Selection Bias, how it occurs?

- Violation of principles of control selection leads to inappropriate control group - Controls more (or less) likely to be selected if they have the exposure of interest.

What are some sources of bias?

- Who? Investigators, personnel, participants - When? Any stage - study design, study conduct, data collection, analysis, and interpretation

Bias away from the null

- overestimate of the effect of the determinant on the outcome

Bias toward the null

- underestimate of the effect of the determinant on the outcome - lesser relationship than in actuality.

Key challenges in Case-Control Studies

-Correctly identify source population of cases - Figure out good control group from source population - Sample Control group to get accurate estimate of distribution of exposure in source population.

Self-Selection Bias

Definition: Nonrandomized cohort studies of efficacy of screening → persons who "volunteer" to be screened differ with respect to disease risk Example: Hypothetical cohort study of new screening test for increased risk of heart disease and coronary heart disease (CHD) mortality

Selection Bias Definition

Definition: systematic error related to subject selection (by investigator) or study participation (by subject)

If intention or participation differs by outcome status, then its _____bias.

Differential Selection

What does Differential Selection do?

Differential Selection does bias the measure of association.

Selection Bias Summary

Occurs when there is differential entry or exit of subjects, or availability of data Case-control studies are particularly vulnerable - also occurs in cohort studies Little can be done to resolve selection bias after the fact, so prevent it Carefully evaluate your results (and those of others) Given conditions of the study, could bias have occurred? Which types? Which direction is the distortion? Are consequences of the bias large enough to distort the measure of association in an important way? (qualitative, quantitative)

In cohort studies, this bias means that the probability of selecting cases and controls differs by

Outcome

How to avoid control selection bias?

Use identical selection criteria

What is particular problem encountered with selection bias?

When an illness presents in a manner such that an accurate DX is not always obtained--ie NSAIDS--User presented with abdominal pain might be more likely to get a full work up since NSAIDS's known association with gastric ulcer. Same situate with oral contraceptives and DVT.

Differential Participation (Self-Selection Bias) How it occurs

Willingness or ability of cases or controls to participate is related to exposure status/history

Can systematic errors be removed?

Yes. They can (in theory) be identified and controlled for.

What is bias?

Error not due to chance which arises in a systematic fashion.

Differential Participation in case control study is related to ____

Exposure

In case control studies, this bias means that the probability of selecting cases and controls differs by

Exposure

Three types of Systematic Error

1. Selection bias 2. Information bias 3. Confounding

Two types of Error

1. Systematic Error 2. Random Error (Chance)

Primary object of Epi Research (2)

1. To obtain accurate measures of a. Frequency of illness b. Association between illness and exposure 2. Must consider possible alternative to explanations for observations during a. design b. analysis c. interpretation

Threats to valudity

Alternative Explainations for observed assoc.

Result of selection bias

Association observed among participants differs from association that would have been obtained among all eligible individuals.

As a Ratio is getting smaller and is an underestimation of the truth, the bias:

BIAS TOWARD THE NULL

How might self-selection bias affect study results?

Can be 2 mechanisms at work: Volunteers who are screened might differ because they are more active or more likely to participate (could be family history) Could be healthier or not healthier volunteers.

Which study (case control of cohort) is more prone to selection bias?

Case control studies

Fundamental goal of controls:

Controls provide same (similar) distribution of exposure as would have gottten from assessing the total population

Counter factual idea:

Counter factual ideal= ideally you'd have the same person at the same time and the same experience just manifests outcome. The healthy worker effect s a violation of the counter factual ideal

Principles of Valid control selection

Fundemental goal: COntrls

Ultimate goal of epi research

Observed Measured of Assoc = True Measure of Assoc.

What can reduce random error?

Larger sample size?

Selection Bias - Solutions?

Little to nothing can be done to fix selection bias once it has occurred Must avoid it at design phase Follow the principles of valid control selection The "would" criterion Independent of determinant Take steps to maximize participation/retention Account for diagnostic/referral patterns Obtain all relevant subject records

Missing Exposure Data as Selection Bias

Loss of exposure (and/or covariate) information - incomplete questionnaires, missing record information, etc. Subjects with missing exposure (and/or covariate) information "selected" out of analysis If missing exposure data differentially related to disease status bias

Healthy worker Effect explain

People in the general population might be less fit and unable to find work, therefore yielding a healthier working population

Diagnostic Bias Definition

Probability of disease diagnosis (and thus case selection) is deferentially related to exposure

Non-differential selection in Case Control Studies

Probability of participation of cases and controls does not differ by EXPOSURE status

Non-Differential Selection in Cohort Studies

Probability of participation/retention of exposed and unexposed does not differ by OUTCOME status.

What do controls do?

Provide information needed to calculate the relative size of exposed and unexposed denominators.

Null in RR/OR and RD

RR/OR= 1 RD= 0

Differential Participation (Self-Selection Bias)

Related to exposure: some participants might be more or less willing to participate depending upon their exposure status or history.

Rues of Control selection (2)

Rule #1- Controls from the same population as cases ('Would" criterion) Rule #2- Selection independent of exposure (same chance of being selected as control whether exposed or unexposed)

Inappropriate control selection definition

Selection of controls does not reflect the source population of cases and/or related to exposure status

Example of inappropriate control selection

Selection of controls does not reflect the source population of cases and/or related to exposure status...Preexisting conditions might cause a difference...Not representation of source population...Some of the surgeries might be related to the exposure itself. We don't know...They were not sampled independent of exposure

How to avoid differential loss to follow up

Take measures to promote retention Simplify participant requirements Maintain good records for tracing Incentives

Why are case control studies more prone to selection bias?

Timing. In case control studies, participation (initial participation in the study) happens after the outcome, so the outcome is more likely to affect selection. Both axes (exposure and outcome) have already happened. In prospective cohort studies, the outcome wouldn't have an impact on the persons initial entry to the study, However, in case control studies, people might be lost based on both outcome and exposure. In cohort, if participation only differs based on exposure, there is not a bias as long as the deceased participation inst related to the outcome.

T/F Some bias is present in all studies regardless of design

True

Can larger samples sized reduce systematic error? Why?

no. There are mechanisms at work, and a larger sample size doesn't change the stability in systematic error because there is a method at work