MI580 Lecture 6 - Causality, Bias, and Confounding

Types of Associations (4)

1. Necessary and Sufficient

2. Necessary, but not sufficient

3. Sufficient, but not necessary

4. Nether sufficient nor necessary

Necessary Cause

Must be present for the outcome to occur, but do not guarantee that the outcome will occur. 

Sufficient

A set of minimal conditions and events that inevitably produce disease. Once these conditions are there, you will get disease.

Sources of Error (3)

1. Chance

2. Bias

3. Confounding

Precision

All data points are consistent.

Accurate

All data points are reflective of the true parameter.

Internal Validity

Results of an observation are correct for the particular groups being studied, implies valid statistical association

External Validity

Some valid associations exist only within particular subgroups, but externally valid results have similar results across diverse populations

Random error

Chance variation in the sample (analogous to precision problem on individual measurements)

Bias

A systemic error in the collection or interpretation of data, results may not be valid or accurate. Results in an erroneous association.

Cannot be correct with statistical manipulation and increases with sample size.

Study Level Error

An inaccurate estimate of the correct value for the measure under study, correct value is unknown and all we can ever really know is the value of the estimate

Major Limitation of Observational Studies

Not randomized, so the control and experiment groups are not balanced

Observational Studies Not Inherently Balanced With Respect To (3)

1. Probability of exposure (e.g., drug or vaccine)

2. Likelihood of detection of outcome

3. Many other factors (i.e., important covariates and risk factors) that can influence why a person gets exposed and also be associated with the outcome of interest

Chance

Random error, lack of power, arbitrary nature of statistical significance threshold (e.g., p-value of < 0.05)

Confounding

A factor related to both the exposure and the outcome that may distort the observed association between the two.

Major Types of Bias in Observation Studies (2)

1. Selection Bias

2. Information Bias/Misclassification

Selection bias

Systematic error due to differences between those selected and those not selected for study/study group participation

Information bias/Misclassification

A flaw in measuring exposure or outcome that results in inaccuracy of information and misclassification

Can be non-differential or differential between the groups.

Nonresponse Selection Bias

Those who do not respond may be/are different than those who do

Volunteer Selection Bias

Select volunteers as exposed group and non-volunteers as non-exposed group in a study of screening effectiveness

Healthy Worker Selection Bias

Study health of workers in a workplace exposed to some occupational exposures comparing to health of general population

Publicity bias/Notoriety bias

Increased awareness through medical publication or media communication may increase physician’s likelihood of reporting a case or recording a diagnosis

How To Minimize Selection Bias in Case Control, Cohort, and Clinical Trials (3)

1. Define criteria of selection of diseased and non-diseased participants independent of exposures in a case-control study

2. Define criteria of selection of exposed and non-exposed participants independent of disease outcomes in a cohort study

3. Use randomized clinical trials

Information Bias

The means for getting information about the subjects is inadequate and results in some information about exposure and/or disease outcome being wrong

To be biased, the information must be more likely to be wrong in one group than another

Major Types of Information Bias (3)

1. Misclassification of outcome

2. Misclassification of exposure

3. Misclassification of any other variables in the analysis

Examples of Information Bias (3)

1. Interviewer Bias (Questioning Bias)

2. Recall Bias

3. Reporting Bias - selective suppression or revealing of information

“Wish” Bias

Subjects who develop a disease and seek to show that the disease is not their fault

Denial Bias

Subjects deny that their disease is due to known exposure (like in STD cases)

How to Decrease Information Bias (6)

1. Memory aids (e.g. calendar)

2. Validate responses

3. Blind interviewers and abstractors

4. Provide strict protocols and standardized training sessions

5. Standardize data collections forms

6. Carefully word questions and pre-test for understanding

Identify a Confounding Variable (3)

1. X is associated with Disease B

2. X is associated with Exposure A

3. X is not in the causal pathway between Exposure A and Disease B

Age is a confounder if … (2)

1. Age is a risk factor for the disease AND

2. Age is associated with (but not caused by) exposure

Why to Control for Confounding (3)

1. To obtain a more accurate estimate of the true association between the exposure and the disease

2. To understand how a third variable could potentially impact the association between the exposure and the disease

3. Important to understand for public health intervention

Confounding vs. Bias

Unlike bias, confounding does not involve error in the way measurements are collected or the way patients are selected, but in the interpretation of what may be an accurate measurement

Positive Confounding

The confounder produces an estimate that is more extreme (either harmful or protective) than the true association, overestimate

Negative Confounding

The confounder produces an estimate that is less extreme than the true association, underestimate

How to Control Confounding Through Design (3)

1. Randomization

2. Restriction

3. Matching

How to Control Confounding Through Analysis (2)

1. Stratification

2. Multivariate Analyses

Randomization

Patients randomly assigned treatment or placebo (or control intervention), a technique used in interventional studies (e.g. clinical trials) and not observational studies

Randomization Advantages (2)

1. Balances drug and control groups on known and unknown confounders

2. The larger the sample size, the better the randomization

Randomization Disadvantages (3)

1. Can be unethical or impractical

2. Patients aren’t always willing to be randomized

3. Patients may not adhere to the treatment to which they were randomized

Restriction

Study participation limited to individuals who fall within a specified category, can be done for any study design

Restriction Advantages (2)

1. Straightforward and inexpensive

2. Avoid diluting risk for certain diseases

Restriction Disadvantages (2)

1. Narrow restriction range can limit recruitment and ability to generalize results

2. Broad restriction may result in residual confounding

Frequency Matching

Frequency of the confounding variable is similar between cases and controls

Individual Matching

Each member of the case group has the same value of the confounding variable as each member of the control group

Matching Advantage

Very powerful method to control for confounders

Matching Disadvantages (3)

1. Time consuming and expensive

2. Not feasible when controlling for large # of confounders

3. Cannot study the variable on which matching was performed

Stratification

Evaluation of exposure/disease association within “homogeneous” categories of the confounder, can be used with any study design

Stratification Advantage

Very powerful method to control confounders

Stratification Disadvantages (3)

1. Not feasible when controlling for large # of confounders

2. Depending on how the strata are defined, there may be residual confounding

3. May not be feasible if large number of strata are required

Multivariable Analyses

A statistical technique that adjusts for multiple variables simultaneously, can be used with any study design

Multivariable Analyses Advantage (2)

1. Can control for many confounders at the same time

2. Powerful method whether outcome is categorical or continuous

Multivariable Analyses Disadvantages (2)

1. Sample size has to be large to accommodate controlling for multiple confounders

2. Depending on the statistical technique, the data must satisfy certain assumptions