LESSON   9

INFORMATION BIAS

9-1   Information Bias

·        Information bias  - systematic error due to incorrect information

·        Misclassification of exposure and/or disease can lead to a bias in the measure of effect.

Misclassifying Disease Status

Misclassifying Exposure Status

9-2   Information Bias (continued)

Misclassification Probabilities – Sensitivity and Specificity

In a perfect world, no one would be misclassified: example, disease status:

In the real world, there will be misclassification.

For misclassification of disease

Sensitivity-probability subject classified as diseased when truly diseased

Specificity-probability subject classified as not diseased when truly not diseased

Misclassification table for exposure status to define sensitivity and specificity parameters.

Consider the numbers in the following misclassification table for exposure.

1.     Sensitivity for misclassifying exposure?

2.     Specificity for misclassifying exposure?

Nondifferential Misclassification

True exposure and disease status for 2000 subjects in a case-control study of the relationship between diet and coronary heart disease (CHD):

·        Exposure: fruits and vegetables eaten categorized as low or high

·        Disease: presence/absence of CHD; assume no misclassification of disease

·        Assume participants overstate intake of fruits and vegetables

The following two tables describe how exposure is misclassified separately for both the CHD cases and the non-cases.

Example illustrates non-differential misclassification of exposure; occurs when the sensitivities and specificities do not vary with disease status.

This example illustrates general rule about nondifferential misclassification - bias is always towards the null.

Differential Misclassification

True exposure and disease status for same 2000 subjects described in the previous activity for case-control study of the relationship between diet and coronary heart disease:

Suppose:

·        No misclassification of disease

·        Subjects over-report their intake of fruits and vegetables

·        CHD cases not as likely to over-estimate intake of fruits and vegetables as controls

This example illustrates differential misclassification of exposure. Sensitivities and/or specificities for misclassifying exposure vary with disease status.

1.     Bias due to misclassifying exposure?

2.     What is the direction of bias?

In general, differential misclassification of either exposure or disease can lead to a bias either towards or away from the null.

Independent Misclassification of Both Exposure and Disease

Two events A and B are independent if the probability that both events occur is equal to the product of their individual probabilities:

In some studies it may make sense that exposure and disease status determinations are more or less independent.

Quantitative Assessment of Misclassification Bias

Study Questions (Q9.9)

 Question Number Observed Effect Corrected Effect Towards the null? Away from the null? a. 2.2 1.7 b. 2.5 3.8 c. 4.0 6.1 d. 4.1 1.2 e. 0.5 0.9 f. 0.8 0.9 g. 0.3 0.2 h. 0.7 0.1

 Observed OR Sensitivity? Specificity? Corrected OR 1.5 80% 80% 3.5 1.5 80% 90% 2.5 1.5 90% 80% 2.5 1.5 90% 90% 1.8

Correcting for Nondifferential Misclassification of Disease

Table provides observed data from hypothetical cohort study of gender and peptic ulcer disease:

·        Assume that gender (exposure variable) not misclassified

·        Diagnosing peptic ulcer involves some misclassification

Need reliable estimates of sensitivities and specificities for disease.  Apply extensive physical examination on sub-sample (200 subjects) for “gold standard” determination of peptic ulcer status:

1.     Sensitivity and specificity estimates for males?

2.     Sensitivity and specificity estimates for females?

3.     Misclassification nondifferential or differential?

Corrected cell frequencies:

5.     Direction of the bias?

Correcting for Nondifferential Misclassification of Exposure

·        Hypothetical case-control study

·        Relationship between history of peptic ulcer (PEU) and stomach cancer

·        Assume case-control status correct

·        History of peptic ulcer has some misclassification.

PEU=peptic ulcer disease

How to correct for misclassification of exposure?

Suppose a different study had evaluated the extent of misclassification:

Exposure Misclassification Table: (Separate Study)

 Gold Standard PEU Not PEU Total PEU’ 48 7 55 Not PEU’ 12 133 145 Total 60 140 200

Sensitivity?            48 / 60 = 80%

Specificity?           133 / 140 = 95%

Assume misclassification of exposure is nondifferential in this example.

Formula to correct for nondifferential misclassification of exposure:

Corrected cell frequencies:

OR in observed data = 3.0

OR in corrected data = 5.1

Bias is towards the null

Correcting for Nondifferential Misclassification of

Exposure and Disease

Hypothetical cohort data – swimming in polluted water and illness.

Illness self-reported; likely some misclassification

Exposure – pollution level of water likely to be subject to error

Assume sensitivity and specificity estimates available:

Assume nondifferential misclassification of exposure & disease; probabilities independent

The resulting corrected cell frequencies are shown here:

RR observed = 1.6

RR corrected = 4.1

Bias towards null

9-4   Information Bias (continued)

Correcting for Differential Misclassification of

Exposure and/or Disease

The formulae get complicated – for the course, just know that it is possible to correct for differential misclassification.  Do not worry about the Quiz 9.15 that asks for you to do this.

Diagnostic Testing and Its Relationship to Misclassification

An Example of Clinical Diagnosis