Unaccounted Variations Can Surreptitiously Spoil the Validity of “Good” Biostatistical Models

Abstract

Most biostatistical models have enormous underlying uncertainties despite being a good fit for the data. These uncertainties arise not just by variations due to sampling fluctuations in the results but also in estimating its components. The clinical application of these models at the individual level aggravates these variations because the models are obtained and tested for groups, not individuals. Such variations can cause a large imprecision but almost invariably remain unaccounted for due to a lack of awareness amongst the researchers. Add to this are variations due to restricting to a limited number of predictors for achieving parsimony, considering a simple model such as linear in place of, say, quadratic, measurement errors, non-random sampling, and other variations such as between observers and instruments. These also are mostly ignored at the time of developing and interpreting the results of a model. Thus, the models generally fail to give correct results in applications. We illustrate the enormity of the unaccounted variations in models with the help of two examples and suggest ways to minimize them.