Web Appendix to: Prospective cohort studies of newly marketed medications: Using covariate data to design large-scale studies August 5, 2013 Derivation of bias The derivation of bias due to inappropriate model extrapolation is due directly to the derivations presented in Drake & McQuarrie1. Specifically, we assume that the true underlying outcome model is given by log{πΈ(ππ |ππ , ππ )} = π½0 + π½π ππ + π½π ππ + π½π» ππ ππ so that the true average conditional rate ratio (RR) treatment effect on the log scale is given by the mean of π½π + π½π» πΜ where πΜ is the mean of X in the population; for estimation purposes we substitute with the mean in the sample. We assume that the treatment effect parameter is estimated via a log-linear model that does not account for treatment effect heterogeneity: log{πΈ(ππ |ππ , ππ )} = π½Μ0 + π½Μπ ππ + π½Μπ ππ Therefore, the bias in π½Μπ for π½π + π½π» πΜ is Bias(π½Μπ ) = πΈ(π½Μπ ) − (π½π + π½π» πΜ ) Drake & McQuarrie provide a calculation for the bias due to an omitted confounder (in our case, the omitted interaction ππ ππ ) in a generalized linear model: Drake C. and McQuarrie A. A note on the bias due to omitted confounders. Biometrika. 1995; 82(3): 633—8. 1 Bias(π½Μπ ) = πΈ(π½Μπ ) − π½π ≈ π½π» [πΈ(ππ|π = 1) − πΈ(ππ|π = 0) − {πΈ(π|π = 1) − πΈ(π|π = 0)} × πβ′ (π½0 + π½π )cov(π, ππ|π = 1) + πβ′ (π½0 )cov(π, ππ|π = 0) ] πβ′ (π½0 + π½π )var(π|π = 1) + πβ′ (π½0 )var(π|π = 0) where π = Prβ‘(π = 1), π = 1 − π, and β′ (π½0 + π½π ) indicates the first derivative of the inverse link function, evaluated at π½0 + π½π . Note that we use different notation than what was originally used by Drake & McQuarrie; their π, πΌ, π½1, and π½2 correspond to our π½0, π½π , π½π , and π½π» . In addition, their π1 and π2 correspond to our X and XT, and we have adjusted the above equation to reflect our use of T in {0,1} rather than {-1,1} to indicate treatment assignment. To arrive at the equation for bias presented in the main body of the paper, we simplify the above expression and subtract off the remaining term for the true average treatment effect: Bias(π½Μπ ) = πΈ(π½Μπ ) − (π½π + π½π» πΜ ) = [πΈ(π½Μπ ) − π½π ] − π½π» πΜ = π½π» [πΜ 1 − 0 − {πΜ 1 − πΜ 0 } × = π½π» [πΜ 1 − πΜ − (πΜ 1 − πΜ 0 ) ππ π½0 +π½π πΜ12 + ππ π½0 ∗ 0 ] − π½π» πΜ ππ π½0+π½π πΜ12 + πππ½0 πΜ02 ππ π½π πΜ12 ] πππ½π πΜ12 + (1 − π)πΜ02 Simulation study To evaluate the validity of this approximation in the context of ommitted interaction effects, we performed a small simulation study. In this study, we set π½0 = −2 and randomly selected 20 true parameter values for each of the other parameters to define 20 datagenerating scenarios: π½π ~ππππ(0, .25) π½π ~ππππ(0, .1) π½π» ~ππππ(−.15, .15) For each of the 20 randomly-selected data-generating scenarios, we simulated 500 datasets with the following variables: π~π΅πππππ’πππ(. 5) ππ=1 ~π(mean = 2,β‘sd = β‘ .5);β‘ππ=0 ~πΈπ₯π(rate = 2) π~ππππ π ππ(exp{π½0 + π½π ππ + π½π ππ + π½π» ππ ππ }) Treatment effect was estimated as the coefficient on treatment in each dataset using a generalized log-linear Poisson model: log{πΈ(ππ |ππ , ππ )} = π½Μ0 + π½Μπ ππ + π½Μπ ππ . The true bias for each data-generating scenario was found by taking the difference between the estimated treatment effect and the true log RR,β‘(π½π + π½π» πΜ ), and averaging across the 500 simulated datasets. Estimated bias was calculated in each dataset using the approximation above, and then averaged across the simulated datasets within each scenario. Appendix Figure 1 shows the comparison of the true bias in each simulation scenario and the average estimated bias, which indicated a strong correlation between true bias and estimated bias with some random variation. Appendix Figure 1: The true bias in each simulation scenario on the x-axis versus the average estimated bias for each scenario on the y-axis. Details of Example Study We implemented methods in a cohort study of the short-term effects of Cox-2 inhibitor (celecoxib, rofecoxib, or valdecoxib) use on gastrointestinal (GI) toxicity and myocardial infarctions (MI) versus ns-NSAIDs. Our study population was pulled retrospectively from patients enrolled in Medicare and either the Pharmaceutical Assistance Contract for the Elderly (PACE) provided by the state of Pennsylvania or the Pharmaceutical Assistance to the Aged and Disabled (PAAD) provided by New Jersey. However, we mimicked a prospective pilot phase study design by evaluating patients sequentially, as they initiated treatment, beginning in the early post-marketing period of each Cox-2 inhibitor. Celecoxib was approved by the United States Food and Drug Administration (FDA) on December 31, 1998. Rofecoxib was approved on May 20, 1999, and was pulled from the market by the manufacturer on September 30, 2004. Valdecoxib was approved on November 20, 2001, and was pulled from the market on April 7, 2005. We included new oral NSAID users age 65 years and older that filled a prescription for an oral preparation of a ns-NSAID or selective Cox-2 inhibitor between January 1, 1999, and December 31, 2005. A new user is defined as a patient with no use of the index drug in the 365 days prior to their index fill. Therefore, patients initiating celecoxib could have prior use of other study drugs. In order to assess past NSAID use and covariate status, patients were required to have continuous Medicare and PACE/PAAD enrollment during the 365 days prior to treatment initiation and to display continuous health system service use, defined as at least one prescription drug claim and one healthcare claim in each of two 180-day periods preceding NSAID initiation. Exposure was classified as ns-NSAID, celecoxib, rofecoxib, or valdecoxib, based on the index prescription. Covariates were created to capture known risk factors of NSAID-associated gastrotoxicity and acute MI and were assessed based on healthcare and prescription claims in the 365 days prior to the index prescription. Additional Figures Appendix Figure 2: Example covariate distributions that display balance and overlap. The covariate distribution for treated and untreated patients is plotted in black and red, respectively. In the left panel, the dashed lines indicate the region of overlap (the area between the lowest value among the exposed and the highest value among the untexposed). Appendix Figure 3: Estimated percent of patients lying in the region of overlap on the propensity score after restricting the cohort to patients with no prior NSAID use, separately for patients at each interim analysis and within each state. Nonselective-NSAID patients are the reference in all analyses.