That so-called research is nothing but a bunch of confounded correlations and mediocre deductions.~ Dr. Frasier Crane, fictional radio psychiatrist
We all know that correlation does not mean causation. Students in courses on research methods learn this early on. Their professors explain the concept of correlation and they provide computational formulas. They do not, however, provide a formula for causation. Why? Because there is none. A causal claim can only be made by a correlation within a context of other cues, such as temporal order (causes come before effects), the presence of physical force or movement (visible impact), or an intervention (human activity). And even then, causation remains a construction of the human mind, not a mathematical fact. Hume was right.
To compute a correlation, we only need measurements. We can measure the weight and the height of a group of people and find that the two sets of numbers are correlated. High measures of weight go with high measures of height. A person weighing over 200 pounds is more likely to be over 6 feet tall than is a person weighing less, but no one believes that weight causes height.
To establish causation, we can put individuals on a supersizing diet and make them heavier. This is an experimental intervention. It is specific. A fatty diet will not make a person taller. What the experiment does, if it is successful, is to introduce a new correlation, namely the correlation between food intake and weight. This correlation has causal appeal. The food intake comes before the weight gain, there appears to be a physical process (metabolism), and it is we who control the delivery of fat.
Experimentation demonstrates causation by intervention. There is an experimental group, in which participants receive a treatment (the consumption of fatty foods), and there is a control condition, in which they do not. By assigning participants randomly to one group or the other, the role of unwanted variables is minimized. When we observe a correlation between group (treatment vs. control) and weight (high vs. low), we conclude that fatty foods make people fat.
Now consider a particular type of situation that can seduce schooled minds into falsely perceiving causation. The situation involves three variables, A, B, and C. Suppose C is a desirable outcome, such as prosocial behavior. Prosocial behavior is, by definition, good for the group and many individuals in it; so we want more of it. Let's say we have also observed that variable B, self-esteem, is positively correlated with prosocial behavior. Compared with individuals with low self-esteem, individuals with high self-esteem engage in more helping behavior, they are more altruistic, they donate more of their time and resources to the common good, and so on.
Remembering the lessons from research methods 101, we realize that the correlation between self-esteem and prosocial behavior may not be causal. While it is possible that high self-esteem makes people more prosocial, it is also possible that prosocial activity raises self-esteem. It is further possible that there is a hidden common cause that affects both self-esteem and prosociality. Perhaps being a first-born is such a common cause or having gene G is.
To test the claim that self-esteem, B, causes prosociality, C, we need to raise B experimentally and observe the consequences for C. We design an experiment, in which the treatment consists of giving participants plenty of social approval. Let's say, these participants receive passes in a ball game more often than would be expected by chance. Participants in the control condition receive passes at the expected rate. If variable A (treatment vs. control) is correlated with variable B (self-esteem) at the end of the game, we can make the causal claim that social inclusion raises self-esteem.
What about prosociality? A casual interpretation of the data may suggest that prosociality also went up, but note that this was not confirmed with measurement. The experiment only showed that A (social inclusion) raised B (self-esteem). There is no reason to conclude yet - however tempting it might be - that C (prosociality) was also raised simply because it had already been known from earlier studies that C is correlated with B.
The problem, I think, is a general one. Often we have some information regarding how one variable - typically the scores on some test - are correlated with later outcomes that we really care about (e.g., job performance, life satisfaction). We can devise treatments, such as coaching or training programs, that improve test performance. For the individuals we study, the life outcomes may lie too far in the future to be measured as part of this study, but we know that in past surveys, these outcomes were positively correlated with scores on the test. Now we conclude that since we improved test scores, we have also improved the life outcomes. This is the intervention illusion. There is no substitute for measuring outcomes directly.
