Evaluation of Family Preservation and Reunification Programs: Interim Report. 8.2 Hazard Analyses of Placement


Hazard analyses permit the examination of the effects of multiple independent variables (in addition to experimental group membership) on rates of placement. They also provide somewhat more precise estimates of the effect of experimental group membership, since they control for the effects of the other variables in examining experimental-control group differences. In addition, they allow for the examination of "interactions" between other variables and experimental-control group membership, to see if the effects of experimental group membership differ for subgroups of the sample. We conducted Cox regression analyses of placement hazards using as predictor variables case characteristics available in the administrative data. Case characteristics in the administrative data are quite limited. Unlike the survival analyses which were conducted at the family level, hazard analyses were done at the child level because we wanted to include in them characteristics of children.

Kentucky. We examined the effects of the child's age, prior placement, and prior allegation of maltreatment, together with experimental group membership, on rates of placement. Only prior allegation was significant, increasing the hazard rate by 68 percent. Interactions of the three variables with experimental group membership were not significant.

New Jersey. New Jersey hazard analyses indicate that older age and prior placement increase the hazard rate significantly (p < .05; prior placement by 77% and each year of age by 8%). Minority status is nearly significant (p = .08), increasing the hazard rate by 37 percent. In an equation with these three variables, experimental group status shows a significant effect in increasing the hazard rate of placement (since this is a child level analysis, significance levels are downwardly biased due to clustering effects). We also examined the interaction of age, prior placement, and minority status with experimental group. The interactions for age and minority status were not significant (that is, the effect of experimental group did not differ by age or by minority status). The interaction with prior placement was close to significant in some of the equations we tried. Examining the differences in placement rates between the experimental and control groups by whether or not the child had had a prior placement indicates that among those with a prior placement, there is little difference in placement rates (23% for the family preservation group and 22% for the control group) while there is a significant difference for those without prior placement (13% for the experimental group and 7% for the control group). (It should be noted that this simple three-way cross tabulation does not account for varying lengths of the observation period, which is controlled in the hazard analysis.) This suggests that children without prior placement who receive family preservation are more likely to be placed than if they did not receive family preservation.

The results do not, however, suggest that family preservation is better for cases with more prior exposure to the child welfare system.8

Hazard analyses were also performed to examine the effect of county on placement. These analyses were conducted at the family level. Burlington county was chosen as the reference category, as it had the highest rate of placement. Thus, rates of placement in the other New Jersey counties are compared to the placement rates of Burlington. In addition to the county variables, experimental group and interactions of county with experimental group were entered into the regression equation. The hazard of placement for families was decreased by 42 percent for Camden county, 71 percent for Ocean county, 78 percent for Monmouth county, 61 percent for Essex county, 71 percent for Bergen county, and 75 percent for Passaic county. There were no significant effects of experimental group or of county-experimental group interactions. This indicates that even after removing county variation, there are no significant differences between the experimental and control groups, nor does the effect of experimental group vary by county.

Tennessee. We examined the effects of the child's age, race, prior placement, prior allegation within 30 days of random assignment, prior substantiated allegation within 30 days of random assignment, and experimental group membership on rates of placement. Similar to Kentucky and New Jersey, we also explored interactions between experimental group membership and child characteristics. No significant interactions emerged. Only prior substantiated allegation had a significant effect on the likelihood of placement subsequent to random assignment. In the analysis of the CORS administrative data, a substantiated allegation within the last 30 days prior to random assignment increased the hazard rate by 209 percent. When unpaid relative placements were included ("any evidence") prior substantiation increased the hazard by 173 percent.

(8)  When a variable representing the designation of a child at risk is introduced into the equation, it has highly significant effects in predicting placement.