Modelling cumulative exposure for inference about drug effects in observational studies

Area of Science:

• Epidemiology
• Biostatistics
• Pharmacovigilance

Background:

• Observational studies often suffer from time-invariant allocation bias, complicating the accurate estimation of drug effects.
• This bias arises from unmeasured factors influencing both drug prescription and patient outcomes.
• Distinguishing true drug effects from allocation bias is crucial for reliable causal inference.

Purpose of the Study:

• To present and validate a statistical modeling approach to control for time-invariant allocation bias in drug exposure-outcome association studies.
• To differentiate the impact of allocation bias from the genuine effect of drug exposure.
• To apply this method to estimate the cardiovascular disease risk associated with statin use in Type 2 diabetes patients.

Main Methods:

• A joint modeling approach incorporating both ever-exposure versus never-exposure and cumulative exposure terms was developed.
• The parameter for ever-exposure quantifies time-invariant allocation bias.
• The parameter for cumulative exposure estimates the drug's effect after adjusting for this bias, using time-updated Cox regression models.

Main Results:

• Crude analysis showed a hazard ratio of 1.13 for ever-statin use.
• Joint modeling estimated the statin effect at a hazard ratio of 0.97 per year of exposure (95% CI 0.97-0.98).
• The ever-exposed term, interpreted as allocation bias, showed a hazard ratio of 1.20 (95% CI 1.16-1.23).

Conclusions:

• Joint modeling of ever-never and cumulative exposure effectively separates causal drug effects from allocation bias.
• This approach is valuable for studying multiple drug effects and improving causal inference in observational research.
• The method is particularly useful when stepwise effects on adverse events are improbable, such as in cancer risk studies.