shows the age, gender, ethnicity, caregivers, and diagnostic groupings for the 109 participants.
Demographic, Symptom, Family, and Service Predictors of Rehospitalizarion
For this clinically heterogeneous sample, DSM-IV principal diagnoses at admission were categorized into five major diagnostic classes (mood, disruptive, anxiety, psychotic, and pervasive developmental disorder [PDD]). Participants were then grouped into the mutually exclusive single or comorbid principal diagnoses shown under “Diagnosis.” Disruptive disorders (i.e., oppositional defiant and conduct disorder) alone or combined with another disorder were the most prevalent diagnostic class.
In the year after discharge, 37 of 109 participants reported rehospitalization; the Kaplan-Meier 1-year estimate of rehospitalization risk, taking into account both reported hospitalizations and censored obserations, was 0.37. The great majority of readmisions (81%) occurred in the first 3 months after discharge.
Predictors of Rehospitalization
shows the univariate associations of demographic, diagnostic, symptomatic, family, and treatment-related predictors with readmission risk. One-year rehospitalization risk estimates, based on Kaplan-Meier survival functions, are shown for each level of categorical variables. For continuous measures, the table contains overall hazard ratios (per unit increase) on the line showing the mean and SD. The table also presents categorical hazard ratios and regression analyses, using tertiles of continuous measures or, when more appropriate for a variable’s distribution, groups above and below its median.
For diagnosis, readmission risk was compared among the mutually exclusive groups of subjects formed by principal diagnoses and the major comorbid combinations. The analysis excluded groups with five and fewer subjects since a reliable estimate of their hazard ratios was not obtainable. With “Mood only” as the reference category, the disruptive group’s hazard ratio was significant. A simplified model that combined the disruptive only and disruptive plus mood groups (total N = 69) to estimate risk relative to the other groups combined (n = 40) also yielded a significant hazard ratio (2.98), but with a much narrower confidence interval (1.31–6.72) than the disruptive only group, so this division was used in the multivariate analysis below.
Among symptom rating scales, conduct problem ratings had a significantly elevated hazard ratio. In categorical analysis, the middle and high subgroups showed similar increases in risk over the low subgroup.
Among family measures, parental involvement yielded a significantly lower hazard ratio. Children in the high-involvement subgroup were significantly less likely to be readmitted than those in the low-involvement subgroup. Although the hazard ratio for the middle subgroup relative to the low-involvement subgroup was also smaller, it was not statistically significant.
The hazard ratio for corporal punishment showed a nearly 2.5-fold increase in the high group relative to the low group. (The severity of parent-reported corporal punishment among study participants was generally not at the level for which reporting to authorities is indicated or mandated. Modifying disciplinary practices was emphasized in treatment, and progress was monitored by unit staff and conveyed to outpatient providers as appropriate.)
Higher parenting stress carried a reduced readmission risk. Further analyses exploring this result appear below.
Among the treatment factors, noninvolvement in outpatient individual child or family psychotherapy was associated with a lower readmission risk relative to those who had these treatments consistently in the first 3 months after discharge. Because children with PDD, who had a lower readmission rate, may be referred less often for psychotherapy, this analysis was repeated excluding those with PDD, and the same result was obtained.
shows the hazard ratios for a combined variable model that included predictors with univariate hazard ratios with p of .10 and less. The overall model’s likelihood-ratio χ2 8 was 43.54, p < .0001. Hazard ratios for the four statistically significant variables in this combined model (p < .01; conduct problems, parental involvement, corporal punishment, and parenting stress) were re-estimated to yield the final model shown in (likelihood ratio χ2 4 = 40.02, p < .0001).
Multivariate Model of Rehospitalization Risk
Parenting stress showed a marked increase in χ2 in the combined model over its univariate analysis. This suggested that understanding the association between higher parenting stress and reduced hospitalization risk should take into account the other variables for which the combined model adjusted. displays the joint influence on rehospitalization risk of parenting stress with conduct problems (top), with parental involvement (middle), and with corporal punishment (bottom). In each graph, the survival function closest to the x-axis (showing the highest rehospitalization risk) represents the group that combines low parenting stress with high values of the other risk factor. Put another way, high parenting stress attenuated somewhat the risks associated with conduct problems, parental underinvolvement, and corporal punishment. Further analyses supported this interpretation. For conduct problems, the hazard ratio for high parenting stress relative to low parenting stress was significantly smaller in the group with high conduct problems (0.39, 95% confidence interval 0.17–0.92, χ2 1 = 4.65, p < .03), but not in the group with low conduct problems (95% confidence interval 0.00 to >100). For parent involvement, the hazard ratio for high parenting stress relative to low parenting stress was significantly smaller in the low-involvement group (0.38, 95% confidence interval 0.16–0.93; χ2 1 = 4.47, p = .03), but not in the high-involvement group (0.56, 95% confidence interval 0.07–4.67). For corporal punishment, the hazard ratio for high parenting stress relative to low parenting stress was significantly smaller in the high corporal punishment group (0.48, 95% confidence interval 0.23–0.98, χ2 1 = 4.01, p < .05), but not in the low corporal punishment group (0.99, 95% confidence interval 0.96–1.03). Further analyses tested the inclusion in the model of interaction terms for parenting stress with the other variables, and all were nonsignificant.
Fig. 1 Joint effects of parenting stress with conduct problems (A), with parental involvement (B), and with corporal punishment (C) on rehospitalization risk. The high parenting stress subgroup represented the highest tertile (). The lowest conduct problem (more ...)
Rehospitalization Risk Index
Signal detection methods examined the utility of a simplified index of readmission risk. This index was the number of risk factors present for each child based on predictors in the final model (i.e., middle/high conduct problems, low parental involvement, high corporal punishment, low/middle parenting stress [ and ]) and therefore ranged from 0 to 4. Among those not censored before 1 year (N = 84), 80% of rehospitalized children had index scores of 3 or higher (sensitivity), a characteristic shared with 21% of the children who were not rehospitalized (i.e., specificity = 0.79). A score of 3 or higher carried a 75% risk of rehospitalization (positive predictive value), while 81% of those scoring 2 and lower were not rehospitalized (negative predictive value). All seven children with scores of 4 were rehospitalized. Total area under the curve was 0.83 (95% confidence interval 0.74–0.91).
Hospital Recommendations and Postdischarge Services
Cox regression for the service-related predictors in did not support the hypothesis that less involvement with aftercare services would increase the rehospitalization risk. Nonetheless, further details of children’s care in relation to inpatient recommendations may aid in depicting the service context for our findings.
The large majority of children admitted (both participants and nonparticipants) had prior psychiatric care. All but six children (5.5%) received outpatient care at the time of their admission. At admission, 90.8% of the sample had pharmacological treatment.
Parents reported high rates of aftercare services for the study sample. Inpatient staff included outpatient psychotherapy in the discharge plans of 91.4% of the children. shows the rates of individual and family therapies and of pharmacological treatment at the 3-month follow-up, covering the period with the highest readmission risk. For the two psychosocial treatments, the table distinguishes between those who had continuous therapy, those who had none, and those who discontinued. Children without outpatient individual or family psychosocial treatment were less likely to be rehospitalized, perhaps reflecting the calibration of aftercare to perceived risk. The same finding held for children who received neither. All the children not in outpatient psychotherapy, however, received pharmacotherapy.
Inpatient staff recommended pharmacotherapy for 99% of the sample. Parents reported medication treatment at any time between discharge and the 3-month assessment for 98% of the sample, while fewer (86%), received pharmacotherapy at the 3-month assessment point. Those who discontinued medication had a readmission rate (24%) nonsignificantly lower than those who did not (36%).
Inpatient staff referred 21% of the sample for day treatment and 20% for home-based services; at 3 months after discharge, these services were implemented for 65% and 55%, respectively, of those referred, reflecting their limited availability. A number of children who were not referred by inpatient staff to day treatment or home-based services obtained them after discharge nonetheless (7% and 15% respectively). The rehospitalization risk did not differ significantly between those referred for day treatment but not receiving it (50%) and those who were in day treatment (60%). Similarly, 33% of those referred for home-based services but not receiving them were rehospitalized, which does not differ significantly from the 45% readmission rate for those who did obtain home-based care.
Further Cox regressions for the influence of these services on readmission risk that adjusted for preadmission conduct problem ratings were not significant.