• Measurement of threats to internal validity:a helps diagnose the presence of specific threats to the inference that A caused B, such as whether units actively sought out additional treatments outside the experiment.

Comparison groups (selecting comparisons that are “less nonequivalent” or that bracket the treatment group at the pretest[s])

  • Single nonequivalent groups: compared with studies without control groups, using a nonequivalent control group helps identify many plausible threats to validity.

  • Multiple nonequivalent groups: serve several functions; for instance, groups are selected that are as similar as possible to the treated group, but at least one outperforms it initially and at least one underperforms it, thus bracketing the treated group.

  • Cohorts: comparison groups chosen from the same institute in a different cycle (e.g., sibling controls in families or last year’s students in schools).

  • Internal (versus external) controls: plausibly chosen from within the same population (e.g., within the same school rather than from a different school).

Treatment (manipulations of the treatment to demonstrate that treatment variability affects outcome variability)

  • Removed treatments: shows that an effect diminishes if treatment is removed.

  • Repeated treatments: reintroduces treatments after they have been removed from some group—common in laboratory science.

  • Switched replications: treatment and control group roles are reversed so that one group is the control while the other receives treatment, but the controls receive treatment later, whereas the original treatment group receives no further treatment or has treatment removed.

  • Reversed treatments: provides a conceptually similar treatment that reverses an effect—for example, reducing access for some students to a computer being studied by increasing access for others.

  • Dosage variation (treatment partitioning): demonstrates that an outcome responds systematically to different levels of treatment.


a Internal validity denotes the level of certainty of the causal relationship between an intervention and the observed outcomes.

SOURCE: Reprinted, with permission. Shadish and Cook, 1999. Copyright © 1999 by the Institute of Mathematical Statistics.

Other Perspectives on Causality

Other perspectives on causal inference exist. In economics, Granger (1988; Granger and Newbold, 1977; see also Bollen, 1989, in sociology and Kenny, 1979, in psychology) argues that three conditions are necessary to infer that one variable X causes changes in another variable Y:

  • Association—The two variables X and Y must be associated (nonlinear association is permitted).

  • Temporal precedence—X must precede Y in time.

  • Nonspuriousness—X contains unique information about Y that is not available elsewhere. Otherwise stated, with all other causes partialed out, X still predicts Y.

The National Academies of Sciences, Engineering, and Medicine
500 Fifth St. N.W. | Washington, D.C. 20001

Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement