HTLH 207 Week 12 - Selection Bias

Selection bias

Type of systematic error and occurs when the measure of association calculated from the sample population is meaningfully different from the measure of effect which would have been estimated if we had data from all eligible subjects in the source population

4 possible sources of selection bias

1. Loss to follow-up

2. Flawed selection processes

3. Participation in the study

4. Participant non-response

Measures of association will only suffer from selection bias if…

The source(s) of selection bias is/are differential, meaning related to both exposure and outcome

The ways that selection bias can influence results are specific to what factor?

The study design

What do uppercase vs. lowercase letters mean in a 2×2 table?

• Uppercase letters: source population

• Lowercase letters: sample population

Differential loss to follow-up can result in selection bias in a longitudinal study if one or both of these factors apply

1. Different proportion of exposed person with the outcome are lost to follow-up compared to those without the outcome (a/A does not equal c/C)

2. Different proportion of unexposed persons with the outcome lost to follow-up compared to those without the outcome (b/B does not equal d/D)

Suppose we observe exactly the same amount of loss to follow-up in each group - is there bias? Why or why not?

• No bias since loss to follow-up was proportional in both the exposed and unexposed groups

• Loss to follow-up was neither associated with the exposure, nor with the outcome

Supposed we observe exactly half the loss to follow-up in the exposed group than in the unexposed groups - is there bias? Why or why not?

• No bias since loss to follow-up was proportional in both the exposed and unexposed group

• Loss to follow-up was associated with exposure but not associated with the outcome (losses among E+O+ were the same as among E+O- and losses among E-O+ were the same was along E-O-)

Does loss to follow-up always lead to bias?

No - this means simply knowing that loss to follow-up occurred does not guarantee that the results have been impacted by selection bias

Selection bias due to loss to follow-up in longitudinal studies occurs when ______________

the losses are differential with respect to the exposure and outcome

Selection bias due to loss to follow-up in longitudinal studies occurs when one or both of the following occur?

1. A different proportion of people in cells A and C (E+O+ vs. E+O-)

2. A different proportion of people in cells B and D (E-O+ vs. E-O-)

3 primary reasons for selection bias in case-control studies

1. Manner in which controls are selected

2. Manner in which cases are selected

3. Differential participation

How can the manner in which cases are selected cause selection bias in case-control studies?

Survival bias → incident vs. prevalent cases

What should controls provide in a case-control study?

Should provide an estimate of the frequency of the exposure in the source population

What occurs when the numerator of the biased OR is too small and the denominator is too big with a positive association?

OR that is smaller than it should be

Survival bias can result in selection bias in a case-control study if one or both of the following occur

1. Exposed cases have differential survival compared to unexposed cases in the sample population relative to the source population (a/A does not equal b/B)

2. Exposed controls have differential survival compared to unexposed controls in the sample population relative to the source population (c/C does not equal d/D)

What is the purpose of enrolling incident cases rather than prevalent cases?

To avoid survival bias

Differential participation can result in selection bias in a case-control study if one or both of the following occurs

1. Different proportion of exposed cases participate compared to unexposed cases in the sample population relative to the source population (a/A does not equal b/B)

2. Different proportion of exposed controls participate compared to unexposed controls in the sample population relative to the source population (c/C does not equal d/D)

Differential participation can result in selection bias in a cross-sectional study if one or both of the following occurs

1. Different proportion of exposed person with the outcome participate compared to exposed person without the outcome in the sample population relative to the source population (a/A does not equal c/C)

2. Different proportion of unexposed persons with the outcome participate compared to unexposed persons without the outcome in the sample population relative to the source population (b/B does not equal d/D)

Survival bias can result in selection bias in a cross-sectional study if one or both of the following occurs

1. Exposed persons with the outcome have differential survival compared to exposed person without the outcome in the sample population relative to the source population (a/A does not equal c/C)

2. Unexposed persons with the outcome have differential survival compared to unexposed persons without the outcome in the sample population relative to the source population (b/B does not equal d/D)

What 2 factors should be maximized, and what factor should be minimized to address selection bias in the design phase?

• Maximum follow-up in longitudinal study designs (trials and cohort studies)

• Maximize participation, of particular importance in case-control and cross-sectional studies

• Minimize participant non-response in all study designs

What type of cases should be selected in the design phase to address selection bias?

Incident cases

What sampling strategies should be used to address selection bias in the design stage? Why?

Use representative sampling strategies to generate a sample population that mirrors the characteristics of the source population

2 types of representative sampling strategies

1. Random/probability sampling

2. Non-random/non-probability sampling

Random sampling (probability sampling)

Individuals in the source population have a predetermined chance of being selected into the sample population

What does random sampling require?

Well-enumerated source population

3 common types of random sampling

1. Simple

2. Stratified

3. Cluster

Non-random (non-probability) sampling

individuals in the source population do not have a predetermined chance of being selected into the sample population

Pro and con of non-random sampling compared to random sampling

• Pro: easier to implement

• Con: has a higher likelihood of selection bias than probability sampling

3 common types of non-random sampling

1. Convenience

2. Voluntary response

3. Snowball

Simple random sampling

Individuals are selected from the source population for inclusion into the sample population at random, meaning all individuals from source population have the same probability of being selected

Stratified random sampling

Divide the source population into subgroups (strata) that differ by important characteristics then conduct random sampling within those strata to ensure that every subgroup is represented in the sample population

Cluster sampling

Divide the source population into subgroups and randomly select entire subgroups into the sample population

Convenience sampling

The sample population includes the individuals who are the most readily available to researchers

Voluntary response sampling

The sample population is comprised of participants who volunteered themselves

Snowball sampling

Uses current study participants to recruit additional participants

2 steps of how to address selection bias in the data analysis phase

1. We can use selection probabilities and data from the sample population to approximate the size of the source population

2. We can then estimate the measure of effect that would have been observed in the source population

4 probabilities that can be calculated from a 2x2 table data from a longitudinal or cross-sectional study

• ⍺ = a/A

• Β = b/B

• 𝛾 = c/C

• δ = d/D

Measure of effect

True association between the exposure and outcome in the source population

Main source of bias in controlled trials and cohort studies

Differential loss of follow-up

Addressing selection bias in the study design phase of controlled trials and cohort studies

Use strategies to maximize follow-up

2 main sources of bias in case-control studies

1. Survival bias

2. Differential participation

Addressing selection bias in the study design phase in case-control studies (2 ways)

1. Select incident cases

2. Use strategies to maximize participation

2 main sources of bias in cross-sectional studies

1. Survival bias

2. Differential participation

Addressing selection bias in the study design phase in case-control studies (2 ways)

1. Best to avoid using this design when survival bias might be a concern

2. Use strategies to maximize participation

2 main sources of bias in any design

1. Differential selection processes

2. Differential participant non-response

Addressing selection bias in the study design phase of any study design (2 ways)

1. Use representative sampling procedures if possible

2. Use strategies to minimize participant non-response

Addressing selection bias in the data analysis phase of any study design

Weighting data from the sample population using:

1. Selection probabilities

2. Inverse probability weighting

3. Imputation

Inverse probability weighting

Estimates causal effects by creating a “pseudo-population” where treatment assignment is independent of measure confounders

Imputation

Statistical process of replacing missing data points with estimated values

Selection bias impacts what type of validity?

Internal validity