Background: Because heath care claims are financial records, they do no not capture free medication samples. Further, because sample doses are typically first doses, distribution of sample medications impedes ability to identify new users in claims. Prevalent users can introduce selection bias in claims studies.
Objectives: Develop a tool to assess the impact of free samples on effect estimates in claims studies by comparing the observed risk with the true risk for various scenarios.
Methods: A decision tree was constructed with six terminal nodes representing mutually exclusive combinations of the following events: sample receipt, outcome status (dichotomous) following sample receipt, and whether the patient filled a second dose. Probabilities were assigned to each node and patient totals were calculated, beginning with the total target population. A Microsoft Excel macro based on the decision tree was developed as a simulation tool. In our base scenario, 90% of study drug and 80% of comparator new users received a first dose as a free sample, adverse outcome risk was 0.365% in the treated population, and the study excluded patients who experienced the outcome prior to their index date. For both drugs, half of the sample users who did not experience an adverse outcome went on to fill a prescription for the drug, and 10% of sample users who experienced the outcome filled a prescription for the drug. We examined differences between true and observed risks and risk ratios (RR).
Results: When the true risk was 0.365%, depletion of susceptible patients due to sample receipt reduced the observed risk to 0.067%, underestimating the true risk by 82%. When the true RR was 2.0, the observed RR was 1.09, and when the true RR was 1.0, the observed RR was 0.55. The observed risk in the comparator group was underestimated by 45% in both cases. Extensions of the base scenario found that a higher prevalence of sample use in the treated than comparator population decreased RRs, while a higher prevalence of sample use in the comparator population increased RRs. For example, if sample use was 60% in the treated population and 80% in the comparator population, then observed RRs were 1.71 (true RR=1.0) and 3.43 (true RR=2.0). Additionally, the larger the proportion of sample users in the target population, the larger the bias, and this bias is exacerbated when sample recipients with events between sample use and first prescription fill are included in the study as incident new users.
Conclusions: Free medication samples can substantially bias estimates of risk and risk ratios in claims-based safety studies. This simulation tool can be used to estimate bias in new user studies arising from first doses not appearing in claims.
