Establishing long-term efficacy in chronic disease: use of recursive partitioning and propensity score adjustment to estimate outcome in MS

Abstract

Context: Establishing the long-term benefit of therapy in chronic diseases has been challenging. Long-term studies require non-randomized designs and, thus, are often confounded by biases. For example, although disease-modifying therapy in MS has a convincing benefit on several short-term outcome-measures in randomized trials, its impact on long-term function remains uncertain.

Objective: Data from the 16-year Long-Term Follow-up study of interferon-beta-1b is used to assess the relationship between drug-exposure and long-term disability in MS patients.

Design/setting: To mitigate the bias of outcome-dependent exposure variation in non-randomized long-term studies, drug-exposure was measured as the medication-possession-ratio, adjusted up or down according to multiple different weighting-schemes based on MS severity and MS duration at treatment initiation. A recursive-partitioning algorithm assessed whether exposure (using any weighing scheme) affected long-term outcome. The optimal cut-point that was used to define "high" or "low" exposure-groups was chosen by the algorithm. Subsequent to verification of an exposure-impact that included all predictor variables, the two groups were compared using a weighted propensity-stratified analysis in order to mitigate any treatment-selection bias that may have been present. Finally, multiple sensitivity-analyses were undertaken using different definitions of long-term outcome and different assumptions about the data.

Main outcome measure: Long-Term Disability.

Results: In these analyses, the same weighting-scheme was consistently selected by the recursive-partitioning algorithm. This scheme reduced (down-weighted) the effectiveness of drug exposure as either disease duration or disability at treatment-onset increased. Applying this scheme and using propensity-stratification to further mitigate bias, high-exposure had a consistently better clinical outcome compared to low-exposure (Cox proportional hazard ratio = 0.30-0.42; p<0.0001).

Conclusions: Early initiation and sustained use of interferon-beta-1b has a beneficial impact on long-term outcome in MS. Our analysis strategy provides a methodological framework for bias-mitigation in the analysis of non-randomized clinical data.

Trial registration: Clinicaltrials.govNCT00206635.

Figure 1. The effect of MPR transformation on the bias introduced by informative-censoring of exposure (see text).

In panel A are the results of the RP analysis incorporating all of the baseline variables and the unweighted raw exposures (measured in years). In this analysis the exposure variable (in years) dominates all other variables with a p-value of 10⁻¹⁶. However, in panel B, where the same analysis is conducted using the unweighted-MPRs (in place of the unweighted “raw” exposures), the entire “spurious” treatment-effect disappears and the resulting tree is identical to that found when all predictor variables (but not treatment) are included in the RP-analysis. In both Panels, the Kaplan-Meier survival estimates are displayed below each of the identified subgroups (splits). X-axis is time in years. Y-axis is survival in % (1 = 100%).

Figure 2. Optimal split determined by the recursive partitioning algorithm considering all predictor variables ( Table 2 ) together with all possible weighted-MPR exposures to IFNβ-1b.

Two highly significant split-levels were identified by the algorithm based on EDSS at the start of therapy and weighted IFNβ-1b exposure during the LTF. The first split (as in the Supplemental Material; Appendix S1; Figure S6) occurred at EDSS = 2, whereas a second level split occurred only for the EDSS>2 branch and was based on DMT exposure. Importantly, the algorithm for this analysis selected precisely the same weighting-scheme (bTN4SN2) that was selected in the Supplemental Material (Appendix S1; Figure S7). Survival curves are displayed below the identified subgroups and survival is best in the EDSS≤2 and the high-exposure groups. The split-point for DMT exposure is slightly different than that identified in the Supplemental Material (Appendix S1; Figure S7) because, in this instance, the split-point was determined only from the subgroup of patients with an EDSS>2. Note: the number (0.028) cannot be interpreted in time units because it represents a mathematical transformation from the raw exposure in years. In both Panels, the Kaplan-Meier survival estimates are displayed below each of the identified subgroups (splits). X-axis is time in years. Y-axis is survival in % (1 = 100%).

Figure 3. Propensity adjusted Cox proportional hazard estimates for the effect of treatment on each of the “hard” negative-outcomes examined in the study.

The results for our principal analysis (i.e., for “any negative-outcome”) are shown on the far left. For each of these outcomes, there is approximately a 60% to 70% long-term benefit to therapy. Error bars indicate the 95% CI for each the different outcomes.