- Non-Differential Misclassification - Bias Towards the Null Hypothesis
- Non-Differential Misclassification - Magnitude of Effect of Bias on OR
- Non-Differential Misclassification of Outcome Variables
- Non-Differential Misclassification of Outcome Variables in a Cohort or Cross-Sectional Study
- Differential Misclassification of Exposure and Outcome Variables
- Measurement Bias in Analytic Studies

- The exception to this comes when specificity is 100%, there is no bias incurred with imperfect sensitivity of the outcome measurement when risk ratios or prevalence ratios are being estimated.

- Whenever you start to measure something with less than 100% sensitivity and 100% specificity you begin to get biased measures of association towards the null.
- In fact, they will always be towards 1.0 unless the sensitivity and specificity are absolutely atrocious in which case you will begin to reverse direction.
- Because this bias is typically towards the null, it has been called a conservative bias.

- After all, the goal in our work is to get at the truth.
- Consider how much underestimation of effects must be occurring in research.
- Or, how many
*negative*studies are truly*positive?* - How much does non-differential misclassification add to the confusion in given fields where some studies have positive results and others are negative?

In a scenario where the true Odds Ratio is 4.0, if sensitivity is 90% and specificity is 85% and the prevalence of exposure in the controls is 20%,

- the observed OR is 2.6.

- results in an OR all the way down to 1.9.

- then the odds ratio will fall from 3.2 to 1.9.

- In the presence of 90% sensitivity and specificity (the last three rows), note how the bias increases the true prevalence of exposure in the controls falls.
- The observed OR falls from 3.0 to 2.8 to 2.2.

- There is, however, a special situation when it comes to
**misclassification of outcome in a cohort or cross-sectional study**. - If
**specificity is 100%**, then any degree of imperfect**sensitivity**will not have any impact on the risk ratio in a cohort study or the prevalence ratio in a cross sectional study. - Here are diagrams of outcome in a cohort study.
- Here is an example of 100% specificity.

Assume that we are looking at an odds ratio in a case-control study and that the true odds ratio is 3.9 and that the prevalence of exposure in the controls is 10%.

- If specificity of the exposure measurement is perfect, we will get entirely different patterns of bias depending upon the pattern of differential sensitivity in the exposure measurement.
- If measurement of the exposure in controls is less sensitive than in the cases, you can see how this will result in an overestimate of the association under study.
- In contrast, if the exposure measurement is more sensitive in the controls, this will lead to a underestimate of the odds ratio, here 2.2.