HSCI330 WEEK 5: SELECTION BIAS

What the the 4 selection ratios?

- alpha - beta - gamma - delta

Cohort studies, weaknesses related to selection bias

- loss to follow up - withdrawal from study - non-response

The difference of error between study population and sample is ________________? The difference of error between source population and study population is ________?

- random error - systematic error

A more spread out distribution graph has:

- smaller sample size - more random error - less precision

Bias towards the null will

- strengthen a conclusion that E and D are associated - weaken the conclusion that E and D are not associated

Bias away from the null

- weaken the conclusion that E and D are associated - strength the conclusion that E & D are not associated

Selective survival

a selection bias, where people survived the cohort study in time to the time of CS study

Selection Bias in Hospital Case-Control Studies (summary)

Berkson demonstrated that case-controls studies carried out using hospitalized patients are subject to a type of selection bias called Berkson's bias. Patients with two disease conditions or high-risk behaviors are more likely to be hospitalized than those with a single condition. Hospital patients in a study population tend to be over-represented when compared to a community population. Berkson's bias can either be towards or away from the null depending on how hospital cases are more or less over- represented than hospital controls when compared to the community.

Selection Bias Due to Inappropriate Choice of Controls

In case-control studies of STDs, the typical control group has been a random sample of persons without STDs from a clinic population. A criticism of such studies is that the correct source population consists of individuals from these clinics who have sexual partners with STDs. In such studies, unexposed controls are likely to be over-represented when compared to unexposed non-cases in the source (i.e., clinic) population. Selection bias is likely when controls are allowed to include persons who have sexual partners without STDs. One way to minimize selection bias would be to re-define the control group to be uninfected individuals known to have sexual partners with STD's.

Odds Ratio (selection probability or selection ratio)

OR = (a/b / cd) aka OR = ad/bc if OR: >1 (bias to the right, >0) =1 (no bias, =0) <1 (bias to left, <0)

Selection Bias in Cross-Sectional Studies (summary)

In cross-sectional studies, there are two distinct ways to consider selection bias, depending on the objective of the study. If the study objective is to survey an assumed stable population at a point or short interval of time, then selection bias may occur because of non-response and/or some other selective distortion of the study population. If the study objective is to determine whether there is an etiologic relationship between exposure and disease, then the primary source of selection bias is selective survival. Selective survival can occur if the probability of surviving long enough to be included in the cross-sectional study is different for the four cells of the source population cohort.

Quantitative Assessment of Selection Bias (summary)

Selection bias can be assessed using a mathematical expression involving the four selection probabilities that relate the target to the study populations. Bias in estimating the odds ratio = 0 if and only if the cross product ratio ( )/( ) = 1. The bias is either > 0, = 0, or < 0 depending on whether the cross product ratio of selection probabilities is > 1, = 1, or < 1. *see p205 for symbols*

Selection Bias in Different Study Designs (Summary)

Selection bias can occur from systematic error that results from the way subjects are selected into the study and remain for analysis. The primary reason for such bias usually differs with the type of study used. In case-control studies, the primary source of selection bias is the manner in which cases, controls, or both are selected. In cohort studies and clinical trials, the primary source of selection bias is loss to follow-up, withdrawal from the study, or non-response. In cross-sectional studies, the primary source of selection bias is what is called selective survival.

Example of Selection Bias in Case-Control Studies (Summary)

Selection bias concerns a distortion of study results that occurs because of the way subjects are selected into the study. In case-control studies, the primary concern is that selection of cases, controls, or both might be influenced by prior exposure status. In the 1970's, there was a lively published debate about possible selection bias among researchers studying whether use of estrogen, the exposure, as a hormone replacement leads to endometrial cancer. The argument supporting selection bias has not held up over time; current medical practice for hormone replacement therapy typically involves a combination of progesterone and estrogen rather than estrogen alone.

Some Fine Points about Selection Bias in Cohort Studies (pg 199-200)

Selection bias in cohort studies may occur even with a fairly high overall response rate or with very little loss to follow-up. Consider a cohort study in which 95% of all subjects originally assembled into the cohort remain for analysis at the end of the study. That is, only 5% of subjects are lost to follow-up. If losses to follow-up are primarily found in exposed subjects who develop the disease, then despite the small amount of follow-up loss, the correct (i.e., target) risk ratio could be underestimated substantially. This is because, in the sample that is analyzed, the estimated risk for developing the disease in exposed subjects will be less than what it is in the source population, whereas the corresponding risk for unexposed subjects will accurately reflect the source population. There may be no selection bias despite small response rates or high loss to follow-up. Suppose only 10% of all initially selected subjects agree to participate in a study, but this 10% represents a true random sample of the source population. Then the resulting risk ratio estimate will be unbiased. The key issue here is whether risks for exposed and unexposed in the sample that is analyzed are disproportionately modified because of non-response or follow-up loss from the corresponding risks in the source population from which the initial sample was selected. We are essentially comparing two 2x2 tables here, one representing the source population and the other representing the sample: *see textbook for diagrams/charts*

Other Examples

Selection bias may occur because of the so-called "healthy worker effect". Workers tend to be healthier than those in the general population and may therefore have a more favorable outcome regardless of exposure status. Selection bias may result from using volunteers, who may have different characteristics from persons who do not volunteer. Clinic-based studies may lead to selection bias because patients from clinics tend to have more severe illness than persons in a population-based sample.

Unexposed controls are likely to be over-represented in the study when compared to unexposed non-cases in the source (i.e: only those at risk) population. T or F

T

Example of Selection Bias in Cohort Studies (Summary)

The primary sources of selection bias in cohort studies are loss-to-follow-up, withdrawal, and non-response. In cohort studies, the collection of subjects that remain in the final sample that is analyzed may no longer represent the source population from which the original cohort was selected. Selection bias will occur if loss to follow-up results in risk for disease in the exposed and/or unexposed groups that are different in the final sample than in the original cohort.

Selection Bias for Risk Ratio (summary)

The rare disease approximation suggests that assessing selection bias in the risk ratio can involve the same rule about the cross-product of selection ratios that applies to assessing selection bias in the odds ratio. If, however, the rare disease approximation does not hold in both source and study populations, then the presence or absence of bias in the odds ratio might not correspond to bias in the risk ratio calculated for the same data. There is no guarantee that the presence or absence of bias in the odds ratio will always correspond to the same degree of bias in the risk ratio.

Selection Ratios and Selection Probabilities (summary)

To quantify how selection bias can occur, we need to consider underlying parameters called selection ratios. There are four selection ratios to consider, one for each cell of the 2x2 table relating exposure status to disease status. A selection ratio gives the number of subjects from one of the four cells in the study population divided by the corresponding number of subjects in the source population. If the study population is a subset of the source population, a selection ratio is typically called a selection probability. A selection probability gives the likelihood that a person from one of the four cells in the source population will be a member of the study population.

Selection probabilities

calculated by dividing the number of categorical people in each cell of study population by (source population) tells you what % of each type of person (relative to their D and E statuses) in the source population is eligible to be selected in your study population -----------------I E I-NON-E _________________________________________ D I --------ALPHA (a)IBETA (b) __________________________________________ NO-D IGAMMA(c)IDELTA(d)

Cross sectional data contains...

prevalence data to make suggestions to incidence data

For cross-sectional studies, there are 2 ways to consider selection bias (weaknesses)

surveys of assumed stable population at a point or short interval of time: - selection bias may occur b/c of non-response and/or some other selective distortion of the study pop. - source pop. includes only those at risk determination of etiologic relationship between exposure and disease - selection bias may occur b/c of selective survival - source pop. includes those without disease at some earlier time, but who were at risk of developing disease by the time of CS study. IN other words, a cohort study begun at an earlier time with follow-up to the time of CS study

How do we use info. about suspected bias toward or away from the null to judge a research result?

suspicion of bias - weak or strength conclusions to make clear judgment about weakening or strengthening conclusions - majors sources of bias act in the same direction ( relative to null) if there are multiple sources of bias that act in different directions - can't conclude if strengthened or weakened

For case-control studies, the primary source of selection bias is...

the manner which cases, controls or both are selected