Causation

Campbell’s Perspective

Donald Campbell originally emphasized making inferences about the direction of the causal (treatment) effect. Campbell and colleagues emphasized identifying all plausible alternative explanations (termed threats to internal validity) that could potentially account for the observed relationship between treatment and outcome. They compiled an exhaustive list of common threats. Prominent threats in longitudinal research include (a) history (specific events occur during the study and affect the [Page 324]outcome—e.g., great recession), (b) maturation (natural within-participant processes affect outcome—e.g., cognitive decline in elderly), (c) testing (prior measurement affects subsequent scores), (d) instrumentation (changes in the nature of the measuring instrument), and (e) statistical regression (participants selected because of high or low scores on an initial measure will be closer to the mean when measured again).

An additional threat that applies to studies that compare treatment groups or natural groups is (f) selection, in which participants in one group may differ from those in the other group at baseline. The first five threats may interact with selection (the sixth threat). For example, children in two schools (one of low SES, one of high SES) may have different baseline rates of gain in achievement, even in the absence of any treatment.

Only those threats to internal validity predicted to lead to the same pattern of results as the treatment need to be considered. For example, a study showing improvement in memory in the elderly following treatment would not need to place a priority on addressing maturation effects given memory’s typical downward trajectory in elderly populations. Campbell proposed adding design elements to the study that permit researchers to rule out as many threats to internal validity as possible.

Three examples from Campbell’s extensive list include matching participants on key baseline covariates to address selection, using repeated pretests measured over time to address maturation, and starting cohorts at distinctly different times to address the possibility that a common history effect could occur in all cohorts. The pattern of obtained results is compared with the pattern of results predicted if the treatment is effective versus the pattern of results predicted if the threat(s) to internal validity are operating.

For Campbell and colleagues, randomized experiments provide the strongest warrant for causal inference. In randomization, each participant has an equal chance of being in the treatment and the control groups. Participants in the treatment and control groups are equated on average on all possible measured and unmeasured baseline variables because a chance process (e.g., flipping a coin) determines treatment group assignment. Randomization theoretically rules out all the threats to validity mentioned earlier. Regression discontinuity designs in which participants are assigned to treatment versus control on the basis of a quantitative measure (typically of need, risk, or merit) can also provide strong causal inferences.

...