Session: Frameworks for Causal Inference (Society for Social Work and Research 22nd Annual Conference - Achieving Equal Opportunity, Equity, and Justice)

Making causal inferences from statistical associations is critical to social work researchers' mission of testing practices and policies to improve social outcomes. Randomized control trials (RCTs) are the standard for causal inference across the social sciences. Certain quasi-experimental methods, such as propensity score matching are popular in social work research; others, like regression discontinuity are less so.

For all of these methods our ability to make causal inferences depends upon whether the data meet certain assumptions or conditions. RCTs, for example, yield causal inference under the assumption that unobserved and observed confounders of the treatment are balanced by randomization. For quasi-experimental research this assumption may be too strong but other, weaker assumptions may be acceptable. Rather than require researchers to develop assumptions from scratch each time a study is undertaken, widely accepted standards exist. These standards take the form of three major causal frameworks: the Campbell, Rubin, and Pearl frameworks.

Understanding the role that these frameworks play in supporting causal inference will help promote better design and scholarship. These frameworks can be used to strengthen design at the planning stage to foresee and prevent easily-solved problems. Further, social researchers are often not clear about their causal assumptions, often obscuring the conclusions of otherwise well-designed studies behind assumptions that may not be credible. A causal framework can help translate the real world into credible assumptions about date generation processes, and may even provide a structured approach to testing some of these assumptions.

Each of three frameworks will be described and compared in the context of applied uses. The framework developed by Donald Campbell and colleagues, which is the most widely-known framework in social work research, identifies four types of validity required to make causal inferences. Within this framework, plausible alternative hypotheses that raise doubts about a causal effect may take the form of one of nine threats to internal validity. Thoughtful researchers can use this framework to identify challenges in their study conditions and data and address them using design or methods.

The potential outcomes framework largely developed by Donald Rubin is also well-known among social work researchers. This framework provides a definition of the average treatment effect, as well as assumptions which must be met in order to estimate such an effect of experimental and quasi-experimental treatments. Using this framework a researcher can precisely state the challenges associated with estimating a causal effect in a quasi-experiment.

The directed acyclic graph framework developed by Judea Pearl is the newest of these frameworks but has become widespread across the social sciences and is become more widely used by social work researchers. This framework is a graphical tool for representing a researcher's assumptions regarding causal relationships among variables . A set of rules then clarify which variables need to be controlled for in order to estimate a causal effect, and which relationships can lend support to the causal estimate. Similarities and differences between each framework will be identified.