Internal Validity

Internal validity refers to the confidence with which researchers can make causal inferences from the results of a particular empirical study. In their influential paper on “Experimental and Quasi-Experimental Designs for Research,” Campbell and Stanley (1963) characterized internal validity as the sine qua non of experimental research. This is because the very purpose of conducting an experiment is to test causal relationships between an INDEPENDENT VARIABLE and a DEPENDENT VARIABLE. Hence, it is appropriate that the primary criterion for evaluating the results of an experiment is whether valid causal conclusions can be drawn from its results. However, the concept of internal validity can also be applied to findings from correlational research anytime that causal inferences about the relationship between two variables are being drawn.

It is important to clarify that internal validity does not mean that a particular independent variable is the only cause of variation in the dependent variable—only that it has some independent causal role. For instance, variations among individuals in their attitudes toward a social program can be caused by many factors, including variations in personal history, self-interest, ideology, and so on. But if the mean attitude of a group of people that has been exposed to a particular persuasive message is more favorable than that of a [Page 503]group of people that did not receive the message, the question of interest is whether the message played a role in causing that difference. We can draw such a conclusion only if no other relevant causal factors were correlated with whether an individual received the message or not. In other words, there are no plausible alternative explanations for the covariation between attitude and receipt of the message. The existence of correlated rival factors is usually referred to as a CONFOUNDING of the experimental treatment because the potential effects of the variable under investigation cannot be separated from the effects of these other potential causal variables.

In this example, the most obvious challenge to causal inference would be any kind of self-selection to message conditions. If participants in the research were free to decide whether or not they would receive the persuasive message, then it is very likely that those who already had more favorable attitudes would end up in the message-receipt group. In that case, the difference in attitudes between the groups after the message was given may have been predetermined and had nothing to do with the message itself. This does not necessarily mean that the message did not have a causal influence, only that we have a reasonable alternative that makes this conclusion uncertain. It is for this reason that RANDOM ASSIGNMENT is a criterial feature of good experimental design. Assigning participants randomly to different levels of the independent variable rules out self-selection as a threat to internal validity.