Chapter 9 202 decrease in the number of similar patients rendered the algorithm of Wang et al.21 unable to complete the simulations in a considerable number of cases; at the extreme µe = −0.5 these methods did not generate output in 65% of all simulations. We removed the line from the figures when the algorithm error rate exceeded 5%.This result highlights a small advantage of weightingbased schemes over stratification-based approaches. In the ProPP, the weights are simply set to 0 when external patients have very different covariate values than the patients in the trial, implicitly discarding part of the data but allowing the analysis to continue. The rest of Figure 5 shows increased RMSE and type I error inflation compared with Figure 4. The ProPP performs favorably compared with the MPP and Wang’s suggested methods for a mild discrepancy e.g., µe ∈ (−0.25;025) between the covariate distributions. All in all, in the absence of latent classes, the ProPP (i) fares reasonably well for small differences between covariates and (ii) accounts for larger distortions when covariate distributions overlap decreasingly. 0.0 0.1 0.2 0.3 -0.50 -0.25 0.00 0.25 0.50 μe Type I error rate A No Mixture 1. Equal sample size:N0 =Ne 0.03 0.04 0.05 0.06 -0.50 -0.25 0.00 0.25 0.50 μe RMSE B No Mixture 1. Equal sample size:N0 =Ne Method Ignore Pooling Pro P M P Wang 10% Wang 20% Figure 5: Comparison of different estimation methods in terms of type I error (left) and root mean squared error (RMSE) (right) when there is no latent class structure in covariates (Setting 3). Estimates have been removed if > 5% of the computations did not produces estimates.
RkJQdWJsaXNoZXIy MTk4NDMw