Neyman’s bias, also known as prevalence-incidence bias, occurs when studying the relationship between an exposure and an outcome using prevalence of the outcome instead of incidence in cases where prevalence is a biased estimator of incidence.
Prevalence is the proportion of individuals who have the outcome/disease at a given time.
Incidence (or risk) is the number of new disease cases that occur over a specified time period in a population at risk of having the disease.
And because not all studies follow participants over time (as in cross-sectional and case-control studies), we sometimes use prevalence as an estimator of incidence. This can work if the following 2 conditions are satisfied:
- the duration of the disease in the 2 groups (exposed and non-exposed) should be the same
- the prevalence of the disease in the 2 groups (exposed and non-exposed) should be very close
Neyman’s bias (prevalence-incidence bias) will occur if either of these 2 conditions is not satisfied.
The direction of this bias will depend on whether the exposure increases or decreases mortality and/or the prevalence of the disease.
Let’s go through an example to understand how this bias works in practice.
Example of Neyman’s bias
A meta-analysis based on 10 years of cohort studies found that smoking increases by 28% the “breast cancer-associated mortality” of breast cancer patients compared to non-smokers.
Based on these results, if we conduct a cross-sectional or case-control study, using the prevalence ratio to compare the smoking and non-smoking groups will be subject to Neyman’s prevalence-incidence bias.
In this particular example, we will be underestimating the relative risk.
How to avoid Neyman’s bias
When studying the effect of an exposure on an outcome (e.g. the effect of a risk factor on a disease), incident cases should be used instead of prevalent cases whenever possible.
As we saw earlier, Neyman’s bias occurs not only when the exposure affects mortality from the disease but also when the prevalence of the disease in the exposed group is different from that in the non-exposed group — the bigger the difference, the larger the effect of this bias will be.
I found no way for controlling Neyman’s bias through data analysis, that is recovering the true relative risk using the prevalence ratio.
But if you’re interested in a statistical way of assessing whether Neyman’s bias affects your study, I recommend you take a look at the 3 hypothesis tests proposed by Swanson et al. although I must warn you, this is a highly technical paper.
Szklo M, Nieto FJ. Epidemiology: Beyond the Basics. 4 edition. Burlington, Massachusetts: Jones & Bartlett Learning; 2018.