Enhance Multistate Models With Clinically Meaningful Graphs

It is a laudable effort to reanalyze historic datasets with improved methods as do Ibrahim et al. [1] with a quarter century-old study on diabetes prevention [2]. They state that the new analysis reduced potential bias in the original results by accounting for competing risks and by respecting the interval-censored nature of the data collection. This commentary does not focus on the lack of evidence that bias was indeed avoided. It asks instead the question of whether the new analysis presented meaningful and clinically interesting results. Coming from a similar training background as the authors [1], I know the temptation to use time-proven and familiar software and to tweak the methods until the top-down, data-driven model fits. In the current analysis [1], NONMEM was applied to estimate rate parameters and covariate influences that govern the probabilities (P1, P2, P3, P4, and P5) for a subject to reside in any of the 5 possible states of a multistate model (figure 1 [1]). Over the decades, some of us pharmacometric modelers have formed a community that has become rather insensitive to the physiologic meaning of the parameters we estimate (while others underwent the arduous work to build physiologically meaningful models). Consequently, we are content with models that are “fit-for-purpose”. In that vein, the authors [1] show that the chosen final model fits the data (figures 2 and 4) and that 95% confidence intervals support the choice of significant covariates (figure 3). The question remains whether the results are meaningful and clinically interesting. The model set-up (equations 10 through 15 [1]) follows the Markovian assumption that a transition into a future state depends only on the current state and not on the time spent in the current state nor on transitions from previous states. As with any longitudinal clinical data, the time already spent in a certain non-absorbing state should influence the hazard rates of future transitions. Thus, the multistate model should, in my opinion, be analyzed as a semi-Markov process. I miss a presentation (as, e.g., in figure 5 of [3]) where all 5 state occupancy probabilities are displayed simultaneously at all times after the start of the study. I also think that a statistical hypothesis test should demonstrate how the intervention influences the time spent in certain passages of interest involving more than two neighboring states (as, e.g., in figure 6 of [3]). Finally, I wonder whether the model can actually predict in which state a patient with a certain set of apparently significant covariates (baseline BMI, HbA1c, insulin sensitivity, sex and age) will most likely be, for example, 10 years after being assigned to one of the two intervention groups. Such displays and tests would in my eyes be more meaningful and clinically interesting than a statement that “the developed model […] demonstrated that lifestyle changes significantly decreased the risk of both diabetes and death.” Perhaps the authors find the time to produce displays that show the reader the claimed reduction in risk. The author declares no conflicts of interest.