Defining and Measuring Treatment Effects

How Economists Use Statistical Modelling to Manage Selection Bias

college graduation
Ron Koeberer/Aurora/Getty Images

The term treatment effect is defined as the average causal effect of a variable on an outcome variable that is of scientific or economic interest. The term first gained traction in the field of medical research where is originated. Since its inception, the term has broadened and has begun to be used more generally as in economic research.

Treatment Effects in Economic Research

Perhaps one of the most famous examples of treatment effect research in economics is that of a training program or advanced education. At the lowest level, economists have been interested in comparing the earnings or wages of two primary groups: one who participated in the training program and one who did not. An empirical study of treatment effects generally begins with these types of straightforward comparisons. But in practice, such comparisons have the great potential to lead researchers to misleading conclusions of causal effects, which brings us to the primary problem in treatment effects research.

Classic Treatment Effects Problems and Selection Bias

In the language of scientific experimentation, a treatment is something done to a person that might have an effect. In the absence of randomized, controlled experiments, discerning the effect of a "treatment" like a college education or a job training program on income can be clouded by the fact that the person made the choice to be treated. This is known in the scientific research community as selection bias and, it is one of the ​principle problems in the estimation of treatment effects.

The problem of selection bias essentially comes down to the chance that "treated" individuals may differ from "non-treated" individuals for reasons other than the treatment itself. As such, the outcomes such treatment would actually a combined result of the person's propensity to choose the treatment and the effects of the treatment itself. Measuring the treatment's true effect while screening out the effects of selection bias is the classic treatment effects problem.

How Economists Handle Selection Bias

In order to measure true treatment effects, economists have certain methods available to them. A standard method is to regress the outcome on other predictors that do not vary with time as well as whether the person took the treatment or not. Using the previous "edition treatment" example introduced above, an economist may apply a regression of wages not only on years-of-education but also on test scores meant to measure abilities or motivation. The researcher may come to find that both years-of-education and test scores are positively correlated with subsequent wages, so when interpreting the findings the coefficient found on years of education has been partly cleansed of the factors predicting which people would have chosen to have more education.

Building upon the use of regressions in treatment effects research, economists may turn to what is known as the potential outcomes framework, which was originally introduced by statisticians. Potential outcomes models use essentially the same methods as switching regression models, but potential outcomes models are not tied to a linear regression framework as are switching regressions. A more advanced method based upon these modeling techniques is the Heckman two-step.