We thank Dr Koch and colleagues and Dr Gout for their attention to our article1 and for giving us the opportunity to stress further the usefulness of appropriate statistical analyses for enhancing the quality of observational studies.
Systematic bias is minimized when a propensity score analysis is performed together with a sensitivity analysis. We recognize that the treatment and control groups in our study were imbalanced for all of the baseline covariates, which meant that it was necessary to use appropriate statistical methods to adjust the comparisons. Propensity score analysis, taking into consideration parameters of interest that would likely affect the outcome, can create balanced groups that have a similar likelihood of receiving a therapy and resemble randomized cohorts of patients.2–4
We agree that an important limitation of this approach is that propensity scores cannot adjust for variables that are not measured in a study (eg, concomitant diseases). Therefore, a sensitivity analysis5 was conducted to evaluate their possible impact on the study outcome. This analysis showed that even a small imbalance between treatment arms in a parameter that doubled the risk for reaching the end point would eliminate the statistical significance of findings for Expanded Disability Status Scale outcomes.1 However, the results for secondary progression end point were more robust. Therefore, statistical adjustment can indeed compensate for even major imbalances; furthermore, sensitivity analysis can also help define to what extent a potential residual imbalance could account for observed outcomes. With regard to the baseline time point we used for each group, the aim of a propensity score analysis is to balance the two groups on measured covariates at their baseline time point, so by its nature, differences between groups at baseline are taken into account using this approach. Furthermore, a secondary progressive disease course at baseline was an exclusion criterion for both groups, and the control group showed a baseline mean Expanded Disability Status Scale score significantly lower than the treated group (see Table 1 in our article1). Therefore, Koch and colleagues' concern that there may be bias toward patients with more progressive disease in the control group compared with the treated group appears unlikely.
The 20% of patients in the control group who had discontinued immunomodulatory or immunosuppressive treatment in the first 3 to 6 weeks because of adverse events were not included in the treatment group intention-to-treat analysis because they were not receiving these therapies at the beginning of the follow-up. Moreover, the treatment group included only patients treated with interferon-β for at least 1 year. For potential imbalance between groups because of differences in patient attitude toward interferon-β treatment (eg, voluntary refusal of treatment), the expression “in ways not clearly measurable” given by Koch and colleagues appropriately stresses the importance of our sensitivity analysis.
Regarding the lack of blinding, we agree this might be of serious concern. Accordingly, this issue was addressed in the Discussion,1 and it is a well-known, unavoidable bias in all observational studies.
We reported a global measure of patients' adherence to treatment, defined as the proportion of days covered by treatment (75.1%). The presence of differential loss to follow-up (ie, informative censoring) between the two arms was addressed1 by assessing the consistency of the hazard ratios censoring the analysis to shorter time frames (from 6 to 3 years). Moreover, it is not appropriate to report absolute numbers of patients at risk for each year because this would be misleading when referring to an adjusted analysis. Finally, concerning the outcomes (secondary progression and Expanded Disability Status Scale scores of 4 and 6), we stated1 that they were confirmed at 6 months, but also at the end of follow-up. However, we agree that it is better to say “confirmed” rather than “irreversible” disability.