A renaissance for measurement error

Our best friend in epidemiology, it seems, is the confounder. The confounder preoccupies our thinking, we respect its omnipresence, and we are endlessly entertained by attempting to identify one in someone else's study. As epidemiologists we spend our days chasing the confounder like detectives, anticipating its disturbing appearance when designing a study, considering potential confounders in our analysis, and trying to illuminate unconsidered or residual confounders when the results of our study do not conform with the expected.

Other toys have also come to occupy our minds. Advanced and fancy analytical methods increasingly find their way into epidemiological analyses. They challenge the epidemiologist and impress the reader. Some real progress has been made with using more refined methods such as hierarchical models, structural causal models, and the improved graphical display of data.

But when we contemplate how to further improve our trade maybe we have to regress to our roots and reconsider one of our oldest acquaintances. One that we seem to have neglected over the years and which apparently has lost favour in the epidemiological community, namely measurement error.

It seems as if measurement error has been pushed into the role of the unwanted child whose existence we would rather deny. Maybe because measurement error is common, insipid and unsophisticated. Unlike the hidden confounder challenging our intellect, to discover measurement error is a ‘no-brainer’ —it simply lurks everywhere. Our epidemiological fingerprints are contaminated with measurement error. Everything we observe, we observe with error. Since observation is our business, we would probably rather deny that what we observe is imprecise and maybe even inaccurate, but the time has come to unveil the secret: measurement error is threatening our profession. The threat is even more serious since mostly it is difficult if not impossible to know …