Page 77 - Contributed Paper Session (CPS) - Volume 7

CPS2033 Ronnie P.

One-sided misclassification of a binary
confounder and bias when estimating causal
effects
Ronnie Pingel
Department of Statistics, Uppsala University, Sweden

Abstract
Measurement errors in the confounders is often neglected when adjusting for
confounders in observational studies. This study is a contribution in that we
study the case of non-differential mis¬classification of a binary confounder.
For the case of linear models an expression is provided so that applied
researcher may study whether the causal effect is under-estimated or over-
estimated. Similar pattern occurs for log-linear models and risk ratios. The
results regarding logistic regression and the odds ratio is less clear.

Keywords
Average treatment effect; Non-differential; Differential; Measurement error;

1. Introduction
The aim of causal inference in observational studies is to study the effect
of a single variable (treatment, intervention exposure, etc.) on an outcome. In
order to estimate causal effects in observational studies, the researcher needs
to include all true confounders in the analysis to achieve an unbiased estimate.
This is known and acknowledged by all researchers trying to make any causal
claims based on such a study.
Well-known is also that it matters how the confounders are used in the
statistical analyses, and that the adjustment for confounding can be made
more robust by applying non-parametric or semi-parametric methods, e.g.
matching on covariates or propensity score-based methods (Waernbaum,
2012; Stuart, 2010).
What to a certain degree is perhaps less emphasized, at least by applied
researchers, is that observational studies of any design require accurate
measurements of the confounders included in the statistical analysis. Still,
measurement error is one of the main sources of bias (Greenland, 1983;
Rothman, 2008; Willett, 1989). Furthermore, attention have mostly been on
measurement error of the treatment or the outcome (Armstrong, 1994), and
not the confounding variables.
Thus, although measurement error is a common feature of empirical data,
especially when working with data from registers, the consequences of using
confounders with measurement error are often neglected (Brakenhoff, 2018).

64 | I S I W S C 2 0 1 9