Bridging Disciplinary Boundaries (January 11 - 14, 2007)

Purpose: To understand the relationship between foster care and youth outcomes it is essential to disentangle the effects of care and pre-existing risk factors. As a group, foster youth have characteristics associated with negative outcomes independent of care. Researchers use several approaches to understand this relationship. Some researchers compare foster youth to the general population, but this does not account for existing characteristics. Other researchers use matching; however, a greater set of characteristics beyond those matched on may place youth at risk. Theoretically, the best way to examine the impact of care would be random assignment, but this is not possible. Given these limitations, new approaches are required to examine this relationship. One potential method, propensity scoring, allows the extraction of youth with high likelihood to be in care that have not had this experience. This study examined whether results were similar when using propensity scoring, traditional matching or comparisons to the general population to understand the relationship between foster care and youth outcomes.

Methods: Using the National Longitudinal Survey of Youth 1997, foster youth (n=171) were matched using several strategies. Propensity scoring used logistic regression to model the likelihood to be in care and find matched youth. Traditional matching was undertaken with three schemes: 1) income, 2) gender, race, and age, and 3) gender, race, age, parent's education, and having a stepparent. Statistics, including R-square values and measures of association from regression models, and comparisons of observable characteristics, were used to see if different schemes produced similarly well-matched groups. Groups were also compared on educational attainment, welfare, parenthood, marriage, and cohabitation using bivariate analysis.

Results: The propensity score logit model had a higher R-square (27%) and higher percent concordance (84%) than models produced using the other schemes (scheme 1: R2=1%, concordance=52%; scheme 2: R2=4%, concordance=61%; scheme 3: R2=10%, concordance=73%). When compared to a model without explanatory variables (general population scheme), the likelihood ratio test indicated the propensity model was a better fit than a model without parameters (p< .0001). Using observable characteristics to compare the groups showed the propensity score group was a better match for foster youth than the matched youth or youth in the general population. Using traditional methods, foster and comparison youth showed differences in race, education, poverty, and father contact, while youth matched with propensity scoring showed no differences. Lastly, while foster and propensity score youth showed similar outcomes in education, welfare, parenting, marriage, and cohabitation, the other schemes produced group differences. This suggests that these outcome differences related to existing risk factors rather than foster care placement.

Implications: This study demonstrates how propensity scoring provides a more well-matched set of comparison youth than traditional matching or matching to youth in the general population. This gets us closer to understanding the relationship between foster care, existing risk factors, and youth outcomes. Though promising, the use of propensity scoring in the child welfare context has several limitations. Potential applications for using propensity scoring for understanding the impact of foster care on youth outcomes will be discussed.