Inspection of the bivariate correlations among the negative affect and sleep variables (both PROMIS and legacy measures) revealed strong relations (and potential multicollinearity). For the negative affect variables, the correlations at baseline ranged from .63 to .85; for the sleep variables, the range was .56 to .80. Therefore, we combined these 2 sets of variables to create 2 composite variables by adding the total scores of each measure. For the PROMIS measures, T scores were used in producing the composite variables. The composite score of negative affect was computed by adding the total scores of PROMIS anger, PROMIS anxiety, PROMIS depression, and the CES-D scale. The composite score of sleep disturbance was computed by adding the total scores of PROMIS sleep disturbance, PROMIS sleep-related impairment, and the MOS Sleep Scale. Significant predictors for each regression analysis and the corresponding variance explained by the aggregated predictors (R2) are shown in .
Regression analysis summary using aggregated predictor variables
The clinical and demographic variables selected as potential predictors did, in fact, account for significant proportions of variance in 14 of the 15 regression analyses. Aggregate R2 values ranged from 7% to 50%, with a median of 34%. Negative affect and sleep disturbance played the most prominent roles, with each of these variables making a significant contribution in 9 of the regressions. Education was included in 4 of the regression equations as the next most common predictor. At baseline, negative affect was more influential than sleep disturbance, predicting significant proportions of the variance for the RMDQ, PROMIS pain behavior, and PROMIS pain interference, whereas sleep disturbance made a contribution only to PROMIS pain behavior.
At the follow-up evaluations, the impact of negative affect remained strong, and sleep disturbance became a more important influence. At 1-month follow-up, sleep disturbance was the only significant predictor for the global ratings of improvement in back pain and leg pain. The standardized coefficients for sleep disturbance were −.41 and −.34 for back pain and leg pain, respectively. Both negative affect and sleep disturbance were significant predictors for the RMDQ (standardized coefficients were .40 and .24, respectively), whereas negative affect was also the only significant predictor for PROMIS pain behavior and PROMIS pain interference (standardized coefficients were .56 and .62, respectively).
At 3-month follow-up, sleep disturbance was a significant predictor for all 5 pain outcome variables. Its standardized coefficients were −.44, −.38, .31, .60, and .41 for back pain, leg pain, RMDQ, PROMIS pain behavior, and PROMIS pain interference, respectively. Negative affect again predicted outcomes on the RMDQ (standardized coefficient=.23) and PROMIS pain interference (standardized coefficient=.31). The proportions of variance explained by sleep disturbance and negative affect (along with other significant variables) were even greater at 3-month follow-up than 1-month follow-up. See for variance explained by each predictor at each time point.
Prior to testing for mediation, we confirmed that there were significant changes in the 5 dependent variables (PROMIS pain behavior, PROMIS pain interference, RMDQ, back pain, and leg pain) between pre-ESI and the 3-month follow-up evaluation. For all of the variables, there were statistically significant differences (P<.001) between pre-ESI status and outcomes at 3 months, supporting the appropriateness of the mediational tests.
Given the strong concurrent associations between negative affect and sleep disturbance and the 5 outcome variables for pain and disability, we also examined how these predictors performed as mediators (at 1-mo follow-up) in the pain trajectories from baseline to 3 month follow-up (see ). The correlations between the pain outcome variables at baseline and 3-month follow-up were .10, .15, .47, .42, and .46 for the global ratings of back pain, global ratings of leg pain, RMDQ, PROMIS pain behavior, and PROMIS pain interference, respectively. The correlation between negative affect and sleep disturbance at 1-month follow-up was .70.
When we tested the mediational paths for the global ratings of back and leg pain, none of the paths through negative affect and sleep disturbance were significant. For the RMDQ, only the path from negative affect at 1-month follow-up to outcome at 3-month follow-up was significant (β=.07, P<.05) (see b2 in ). In addition, the mediational paths through negative affect (a2: β=7.8, P<.001; b2: β=.11, P<.01; their cross product, a2 × b2: β=.87, P<.01) were significant for the 3-month outcomes on PROMIS pain behavior. The same mediational paths through negative affect (a2: β=3.16, P<.001; b2: β=.12, P<.01; their cross product, a2 × b2: β=.37, P<.01) were significant for the 3-month outcomes on PROMIS pain interference. There was no evidence of mediation by sleep disturbance on these 2 outcome variables. Thus, the evidence for mediation was mixed—it varied depending on the outcome variable—but when it did appear, it was stronger for negative affect than for sleep disturbance.