Bias amplification of unobserved confounding in pharmacoepidemiological studies using indication-based sampling

Abstract

Purpose: Estimating causal effects in observational pharmacoepidemiology is a challenging task, as it is often plagued by confounding by indication. Restricting the sample to those with an indication for drug use is a commonly performed procedure; indication-based sampling ensures that the exposed and unexposed are exchangeable on the indication-limiting the potential for confounding by indication. However, indication-based sampling has received little scrutiny, despite the hazards of exposure-related covariate control.

Methods: Using simulations of varying levels of confounding and applied examples we describe bias amplification under indication-based sampling.

Results: We demonstrate that indication-based sampling in the presence of unobserved confounding can give rise to bias amplification, a self-inflicted phenomenon where one inflates pre-existing bias through inappropriate covariate control. Additionally, we show that indication-based sampling generally leads to a greater net bias than alternative approaches, such as regression adjustment. Finally, we expand on how bias amplification should be reasoned about when distinct clinically relevant effects on the outcome among those with an indication exist (effect-heterogeneity).

Conclusion: We conclude that studies using indication-based sampling should have robust justification - and that it should by no means be considered unbiased to adopt such approaches. As such, we suggest that future observational studies stay wary of bias amplification when considering drug indications.

Keywords: bias amplification; causal inference; confounding; pharmacoepidemiology.

Figure 1.

Directed acyclic graphs showing drug indication as a perfect IV (A), as a near-IV (B), and as a near-IV with induced correlation with the unmeasured factor due to collider stratification on drug use (C).

Figure 2.

Directed acyclic graph showing a hypothetical real-world example of the relationship between familial hypercholesterolemia (Z), statin use (X), smoking (U), and lung cancer (Y), and the magnitude of their relationships (RR, relative risk), where the causal effect of statin use on lung cancer is of interest (but there is no such effect, RR=1).

Figure 3.. The estimated relative risk of X→Y from a crude analysis (no control), an analysis adjusted for Z (multivariable regression), and an analysis using indication-based sampling (crude analysis while selecting those with Z=1), over ever-increasing confounding by indication (greater Z→Y), by different magnitudes of the effect of the confounder on the outcome (greater U→Y).

Legend: Dotted line indicating the preference in terms of net bias between crude and indication-based analysis; on the left side of the dotted line the crude estimator is less biased and on the right side of the dotted line the indication-based analysis less biased. The shaded area indicates a 95% confidence interval. The red line indicates the true causal effect (RR 1). For an extended description of the simulation, see Appendix.

Figure 4.. The relative risk of X→Y from a crude analysis (no control), an analysis adjusted for Z (multivariable regression), and an analysis using indication-based sampling (crude analysis while selecting those with Z=1), over ever-increasing confounding from U (greater U→Y), separately whether in the absence or presence of confounding by indication (RR Z→Y: 1.25).

Legend: Dotted line indicating the preference in terms of net bias between crude and indication-based analysis. The shaded area indicates a 95% confidence interval. The red line indicates the true causal effect (RR 1). For an extended description of the simulation, see Appendix.

Figure 5.. The estimated relative risk of X→Y from a crude analysis (no control), an analysis adjusted for Z (multivariable regression), and an analysis using indication-based sampling (crude analysis while selecting those with Z=1), over an ever-increasing prevalence of Z and U and by whether confounding by indication is absent or present (RR of Z→Y = 1.25).

Legend: The shaded area indicates a 95% confidence interval. The red line indicates the true causal effect (RR 1). For an extended description of the simulation, see Appendix.