Evaluating possible confounding by prescriber in comparative effectiveness research.

Area of Science:

• Health Services Research
• Biostatistics
• Pharmacoeconomics

Background:

• Nonrandomized comparative effectiveness studies often overlook prescriber influence on treatment assignment.
• Prescriber preference is a significant factor that can introduce confounding bias.

Purpose of the Study:

• To evaluate the bias of different analytical approaches that incorporate prescriber information.
• To compare these methods against the standard approach that ignores the prescriber.

Main Methods:

• Monte Carlo simulation was used to assess bias in three prescriber-inclusive methods.
• Methods evaluated included instrumental variable (IV) analysis and stratification on prescriber.
• Propensity score models with prescriber random intercepts were also examined.

Main Results:

• Instrumental variable (IV) analyses were unbiased only when specific IV criteria were met (no unmeasured patient characteristic clustering within prescribers).
• In other scenarios, IV analyses demonstrated significant bias; stratification on prescriber effectively reduced confounding bias.
• Including a prescriber random intercept in propensity score models led to unpredictable bias changes and reversed confounding direction.

Conclusions:

• Caution is advised when using instrumental variable (IV) analysis in comparative effectiveness research, particularly with weak instruments.
• Stratification on the prescriber appears to be a more robust method for reducing confounding bias.
• Further research into prescriber-focused analytical strategies is warranted.