Society for Social Work and Research

Background and Purpose: Many important research questions in social work research do not lend themselves to experimental designs. When researchers have observational data or randomized control trials are compromised by non-compliance, causal inference is still possible with the appropriate design. This symposium will present three papers that illustrate the use of research designs that incorporate instrumental variables. Instrumental variables are an important part of a tool-kit for causal inference especially when researchers are faced with a) heterogeneous treatment effects and b) non-compliance in randomized control trials.

Methods: Each of the three papers incorporates an instrumental variable in a regression framework. Two of the papers use instrumental variable techniques in natural -experimental settings. The first paper constructs an instrument by exploiting two sources of exogenous variation in family income to estimate causal effects on childhood obesity. The second paper uses the effective randomization of child protection investigators to cases to estimate causal effects of foster care placement on juvenile delinquency and utilization of emergency healthcare. The third paper uses the random assignment variable itself as an instrument for the receipt of services in two randomized control trials where there was substantial cross-over between treatment and control groups. There will be three points of emphasis in the methods portion of each presentation: a) defining the instrumental variable, b) explaining the assumptions underlying the instrumental variable approach, and c) presenting the mechanics of the technique (Wald estimates and 2-stage least squares).

Results: The instrumental variable approach will address bias due to omitted variables given the underlying assumptions. The presentation of results will emphasize the appropriate interpretation of the instrumental variable estimate and its limitations. Two of the papers will emphasize the appropriate interpretation of a local average treatment effect, namely that the causal effect applies only to individuals whose behavior was influenced by the instrument. The third paper will compare the instrumental variable estimate used to compensate for the compromised randomized control trial with estimates from the standard intent-to-treat analysis.

Conclusions: The symposium will provide attendees with a clear understanding of the assumptions underlying an instrumental variable approach and the appropriate interpretation of the results. The instrumental variables that are illustrated in these papers can apply to a wide range of research questions that are critical to social work research.