Background/Purpose: Methods for causal analysis in randomized clinical trials (RCT) where functional outcomes may be truncated by death of participants have been developed but are not widely used in arthritis. To demonstrate one of these methods, based on a composite score, we applied it to WORK-IT, a parallel-arm RCT of a work disability prevention intervention for persons with musculoskeletal conditions (ClinicalTrials.gov NCT01387100.). In WORK-IT the primary outcome was the 2-year (104-week) value of the Output Job Demand subscale (OJD) of the Work Limitations Questionnaire. The OJD measures the percentage of time an employee has job output limited by health, (0% best – 100% worst). Unemployment at 104-weeks truncated the OJD for some participants and made these values missing. The rate of unemployment was significantly lower in the intervention arm (11/11744 vs. 25/12535 in controls [Events/Person-Week], p=0.03 Logrank test), however the OJD was not different by arm (Intervention: 25.9±1.8, Control: 28.0±1.9 [Mean±SE], p=0.44 t-test).

Methods:  We derived a composite score (CS) defined as CS=#(Follow-up weeks employed)+(100-OJD). A participant employed at 104 weeks with no OJD limitation would have the highest CS=204; and participants unemployed at 104 weeks would have CS<104, depending on when they became unemployed. An intent-to-treat (ITT) analysis with multiple imputation was used to obtain OJD values for participants who withdrew or provided incomplete responses. A complete case analysis excluding these subjects was also performed. The CS combines different outcomes, so a Wilcoxon test-based estimate of the probability that an intervention subject had a better outcome than a control subject P(I>C) quantified the difference between treatments, with values above 0.5 indicating intervention subjects had better outcomes.  The difference between arms was significant if a 95% bootstrap confidence interval (CI) for P(I>C) excluded 0.5. 

Results: We studied 143 intervention and 144 control participants, 36 were unemployed at 24 months and unable to provide OJD, but used in CS. The median CS value of 179 for the intervention group (see table) indicates that at least half the intervention subjects were employed at follow-up with no more than 25% time of limited output. In both analyses P(I>C) is above 0.5 indicating that intervention subjects have higher probability of a better outcome than control subjects, and since the lower bounds on the 95% CI exclude 0.5, both the complete case and ITT analysis find a statistically significant difference between the arms.

Conclusion: Use of composite scores allow consistent estimation of treatment effects in clinical trials where outcomes are truncated by a related event. This approach could be extended to other longitudinal studies where a measured outcome is truncated, such as knee pain evaluation in osteoarthritis being truncated by knee replacement.