SUPPLEMENTAL APPENDIX Relationship between Recovered and enrollment propensity
To assess the form of the relationship between Recovered and enrollment propensity at the site-level, we plot the proportion recovered adjusted for participant characteristics against site RCT enrollment proportion (enrollment propensity), super-imposing a smoothing spline fit by weighting observations by the number of participants at the site. Because small sites were more likely to obtain extreme values and there was evidence of nonlinearity for site RCT enrollment propensity ≤ 0.1 and ≥ 0.25, we used the binary variable defined by whether a site’s RCT enrollment propensity exceeded the median as the measure of enrollment propensity in our main analysis. In the sensitivity analysis reported in Table A2 we evaluate results using site enrollment propensity Winsorized (i.e., truncated) at 0.1 and 0.25 as the measure of enrollment propensity.
Figure A: Site proportional recovered versus site RCT enrollment propensity
Progression to Recovered
The eight-category clinical status measure used in STEP-BD observational and RCT arms was represented as a three level measure reflecting the mood symptom severity/course. Level 1 includes the four diagnostic states (depression, mania, mixed/cycling, hypomania) and continued symptoms; level 2 includes the roughening and recovering process states (transition points from and to recovered); level 3 is recovered. Studying the times until recovered confirms that
progression towards recovered typically occur in steps as opposed to large jumps. Notably, the process states (roughening, recovering) subgroup of the clinical status measure used in the model is a gate way from the initial diagnostic states to Recovered (Table A1). For example, on
only 2% of occasions individuals transition from a diagnostic state to Recovered compared to 28% from a process state. However, remission from Recovered was more discretionary with transition rates of 15% and 12% back to a process and a diagnostic state, respectively. Thus, process state – Recovered transitions are symmetric (21% and 15%) while diagnostic state – Recovered transitions are asymmetric (2% and 12%).
Table A1: Transitions in status of Clinical status
Current Clinical Status Number of Lag Clinical Status Diagnostic state Process state Recovered observations
Diagnostic state 0.70 0.28 0.02 5297
Process state 0.27 0.52 0.21 3093
Recovered 0.12 0.15 0.73 850
Construction of time scale
To gain insight into the optimal representation of time, we plot the adjusted proportion of recovered against week of follow-up for weeks 5-26 (Figure 1). As displayed in Figure 1, the periods of time <9, 9-11, 12-15, and 16+ weeks group time such that the Recovered rate is largely invariant within the period (9-11 weeks is a possible exception) but changes extensively between periods. The same contraction of the time scale applies equally-well to the RCT and observational arm participants. The reduced time-scale is parsimonious and amenable to using a linear trend for the population and individual time trends, facilitating easy interpretation of results.
Sensitivity Analysis
To check if our results were sensitive to the specification of enrollment propensity, in the following we repeat the analysis using enrollment propensity Winsorized at 10% and 25% as enrollment propensity. That is, any rates less than 10% were changed to 10% and any above 25%
were changed to 25%; these truncations were suggested by Figure A. The results from using this alternative measure are reported in Table A2.
In response to a 10-percentage-point change in proportion of participants enrolled in a RCT, the odds of Recovered in a given week was 1.68 (1.13, 2.49) in the observational arm and 1.19 (0.48, 2.94) (not significant) in the RCT. Thus, the results are minimally changed by using the continuously-valued Winsorized enrollment propensity variable.
Table A2: Estimated Odds Ratios of Predictors on Probability of Recovered when site RCT enrollment propensity is Winsorized (Truncated) at 10% and 25%.
Explanatory Variable Estimate 95% CI P (effect) Site enrollment proportion (Effect of a 10 percentage-point increase)
Truncated linear trend (OBS) 1.68 1.13 2.49 0.01 * Truncated linear trend (RCT) 1.19 0.48 2.94 0.70 Follow-up time
Linear trend (OBS) 2.98 2.48 3.58 0.00 ***
Linear trend (RCT) 5.56 3.87 7.98 0.00 ***
RCT participation 0.65 0.06 6.78 0.72
Psychosocial study participation 0.72 0.30 1.71 0.45 Baseline clinical status = Process 17.6 2.69 115.2 0.00 **
Past substance abuse 1.40 0.87 2.25 0.16
No insurance 0.63 0.33 1.21 0.17
Income 40000 or greater 1.71 1.07 2.74 0.03 *
Age 0.98 0.96 1.00 0.03 *
Male 0.86 0.54 1.36 0.52
Race/Ethnicity (base = white)
Black 1.20 0.37 3.86 0.76
Hispanic 1.87 0.65 5.40 0.25
Other 0.62 0.16 2.46 0.49
Education (base = Less than high school graduate)
High School Graduate 0.92 0.28 3.06 0.89
Some College 1.35 0.44 4.12 0.59
Greater than college 2.21 0.76 6.44 0.15
General Health Comorbidities
1 condition 2.33 0.78 6.93 0.13
2+ condition 1.19 0.69 2.07 0.54
Mental Health Comorbidties
1 condition 0.58 0.35 0.97 0.04 *
2+ condition 0.31 0.15 0.66 0.00 **
Baseline symptoms (Base = None or Mild)
Moderate 0.74 0.43 1.27 0.27
Severe 0.27 0.13 0.54 0.00 ***
Baseline Bipolar Disorder Course of Illness Index (% in category)
Non-specific 1.15 0.27 4.92 0.85
Known convincingly 1.65 0.44 6.25 0.46
Note: *** = p-value < 0.001; ** = 0.001 ≤ p-value < 0.01; * = 0.01 ≤ p-value < 0.05; . = 0.05 < p-value < 0.10.
RCT indicates that the patient participated in the RCT while OBS indicates they remained in the
observational arm. The p-values for the interaction of site enrollment proportion above median and linear time trend with RCT participation were p = 0.494 and p = 0.003, respectively. The truncated linear trend variables were multiplied by 10 so that a unit increase would represent a 10-percentage-point increase.