presents the core predictor variables identified in the WTCD developmental data set. As seen, these variables span the pre- and the postdisaster periods and bridge different clinical domains, including the social, environmental and psychological factors. The general population profiles for three studies used in our analyses are shown in . Clearly, the WTCD developmental sample appears to differ from the pain and trauma studies in terms of race/ethnicity and level of education. These study samples, however, appear more comparable with respect to lifetime trauma exposure and the prevalence of PTSD (). Nevertheless, χ2 tests indicated that differences between study samples shown were all statistically significant (P < .05).
Study profiles of populations used in New York PTSD risk score study
summarizes the results for each study by different groups of predictor variables. As suggested, the SSSP and the PCPS were selected because they were in general use and also included in the PTSD symptoms scales used in the studies evaluated. As can be seen, these screeners alone are good predictors of PTSD. For example, in the WTCD bootstrapped evaluation sample, the SSSP had a specificity of 88.8% and a sensitivity of 92.5% (AUC=0.907); the PCPS had a specificity of 82.2% and a sensitivity of 93.7% (AUC=0.880). Among the pain patients, the results for the SSSP were a specificity of 95.6% and a sensitivity of 94.3% (AUC=0.949); the PCPS also had a specificity of 95.6% and a sensitivity of 94.3% (AUC=0.949). When we used these screeners in the trauma study, the results for the SSSP was a specificity of 85.9% and a sensitivity of 65.4% (AUC=0.757). However, the PCPS performed better among trauma patients, with a specificity of 80.9% and a sensitivity of 96.2% (AUC=0.885).
Prediction results: WTCD, pain and trauma studies using different prediction models
In the WTCD development sample, adding the psychosocial predictors to the model with the SSSP resulted in a significant improvement (AUC=945, P < .0001). This is also true for the PCPS (AUC=0.943, P < .0001). This level of improvement was not observed in the pain study, either for the SSSP (AUC=0.965, P=.0725) or for the PCPS (AUC=0.970, P < .0516), after psychosocial variables were added to the model. By comparison, in the trauma study, the results were significant when psychosocial variables were added both for the SSSP (AUC=0.874, P < .0001) and the PCPS (AUC=0.929, P=.0063). However, adding demographic variables did not improve the predictions in the WTCD study (AUC=0.949, P=.0862 for the SSSP model; AUC=0.945, P=.2497 for the PCPS model). The results were similar for the pain study (AUC=0.966, P=.7779 for the SSSP model; AUC=0.972, P=.3673 for the PCPS model) and for the trauma study (AUC=0.883, P=.2337 for the SSSP model; AUC=0.928, P=.8929 for the PCPS model).
Because the prevalence of PTSD in our study samples was relatively low (7.3%–11.6%), the predictive value of a positive test (PV+) was generally less than 50%, while the predictive value of a negative test (PV−) was typically 99% (). However, we note that given our prediction models, if our study populations had a PTSD prevalence of ~20%, statistical simulations using Pepi software suggested that the positive predictive value of a positive test would be about 80% to 90%, a substantial improvement.
Following the assessments using a combination of predictors, we also examined the predictive value of the psychosocial risk factors alone (). As can be seen, the use of the psychosocial predictors alone (i.e., sleep disturbance, depression symptoms, trauma history, not having a regular source of care) also worked well in predicting PTSD. In , the AUC results for the WTCD, pain and trauma studies range from 0.873 to 0.946, suggesting good predictive results using only these variables. However, adding demographic factors to the latter prediction models only improved the prediction results for the WTCD study, with AUC=0.906 and P=.004, but not for the pain or the trauma studies ().
Prediction results: WTCD, pain and trauma studies using psychosocial risk factor and demographic models only
presents PTSD scoring results (i.e., the final derived weights) for the PCPS used to generate the classification results shown in for these PCPS models. We present the PCPS results because these appeared to produce the best overall predictions and the PCPS has fewer scale items than the SSSP (four items vs. seven items). As seen, a positive score on the PCPS (i.e., three or more positive items) is given a base score of 100, and the psychosocial and demographic items are given weights (or scores) relative to this score, This scoring is based on the logistic regression analyses, whereby the b coefficients in these logistic regression models are converted to standard weights with the aid of a nomogram, as discussed (). The last row of shows the cutoff score for a PTSD classification based on these weights: 100 for the PCPS used alone, 158 for the PCPS+psychosocial predictors and 194 for PCPS+psychosocial+demographic factors. Cutoff scores are also shown for psychosocial predictors alone (score=152) and for psychosocial+demographic factors (score=159). Thus, to be classified as a PTSD case using the full NY PTSD Risk Score model simply requires a positive score on the PCPS (score=100), exposure to two lifetime traumatic events (score=26), current sleep disturbance (score=46) and having no regular source of medical care (score=23), which sum to a total score of 195.
New York PTSD risk scores for PCPS, psychosocial risk factors and demographic factors
Fig. 1 Nomogram for PCPS with psychosocial predictors included. Fig. 1 shows scale metrics used in score development with the aid of a nomogram, whereby final regression coefficients are converted to additive risk scores relative the PCPS using a linear predictor (more ...)
Finally, we assessed these results for men and women separately and found that while the overall results were generally the same, the PTSD screeners tended to contribute more to the prediction models for the women, while the depression symptoms contributed more to the prediction model for the men. We plan to analyze these findings further by gender to determine if separate prediction models are justified in a future study.