of exposure cannot be determined with precision. Multilevel regression models are increasingly being used in epidemiology and public health studies to examine contextual influences in a variety of areas. He used John Snow’s discovery that removing the handle of a pump reduced the incidence of cholera in the neighborhood (by limiting access to tainted water) to emphasize the importance of exploring all potential factors. This finding led to efforts to disentangle more proximal causes for disease—which in turn led to the breakthrough discovery of germ theory and the effects of bacteria in sewage.
Oakes also found little innovation in media research methodologies in their approaches to inferring cause in observational studies. The axiom that correlation does not imply cause remains true, but the potential of surveys and correlational studies to illuminate cause, even in the absence of an explanatory theory or mechanism, is dismissed too readily, he argued. Other research fields are making significant strides in finding innovative ways to infer causation. Counterfactual causality, originally identified by philosopher David Hume as the “but for” condition—the case in which situation x would not be true but for intervention y—has been used in contemporary research to account for alternate outcomes that result from differing exposure to particular influences. This line of analysis compares scenarios in which two conditions with similar features produce very different outcomes as a result of the presence or absence of intervening factors.
This model helps to compensate for the basic challenge of causal inferences—that it is never possible to observe the exposure of one individual to two different sets of influences in the same period of time. A substitute for the unobservable condition—the individual exposed to influences other than those to which he or she was in fact exposed—must be found. Bias creeps in when the substitutes are not good controls because they are not sufficiently similar to the research subjects. Randomized groups of sufficient size provide one means of ensuring that the control group is similar to the targeted group, but in observational studies, in which no intervention is done, there is no opportunity for randomization.
To approximate the benefits of randomization when it is not possible, a variety of new statistical techniques can be used to analyze the data from observational studies to identify causal relationships. One such method, developed by Rosenbaum and Rubin (1983), is called propensity score analysis. This approach begins by developing a statistical model to predict exposure independent of the outcome in question. In a study of the impact of exposure to media on children’s diets, for example, logistic regression could be used to estimate the probability—called the propensity score—that a child is exposed to food advertising on TV as a function of parental monitoring and other factors. The probability of exposure, known as the propensity score, is then used as a control variable in an analysis of the relationship between diet and food advertising. Under certain assumptions, it has been shown that the relationship between the outcome (diet) and exposure (advertising) that can still be observed after controlling for other factors using the propensity score can be attributed to food advertising. Like any other statistical method, however, the validity of the inference depends on the assumptions, and propensity score analysis results can be biased if the assumptions are not correct (Rubin, 1997).
Oakes stressed that propensity score analysis is a relatively new approach that may be particularly well suited to the kinds of questions that media researchers tackle.