As human beings, we are rather obsessed with a very simple question: why?  We infer causality based on very limited evidence. For example, John goes against his better judgment and watches Netflix until 1:00 am.  He then gets in an argument with his boss then day at work.  He assumes that it is because he watched Netflix until 1:00 am that he had the argument with this boss. 

We often make decisions based on a limited amount of information.  We often rely on System 1 thinking, making decisions based on heuristics - that is, cognitive short-cuts. In this case, we use the availability heuristic, using easily recalled pairings of people and behaviours to make a judgment.  For example, people tend to think that a change in weather leads to arthritis pain. They can easily recall when the weather changed and this happened, but in a study by Redelmeier and Tversky (1996), they found no such correlation.  However, they did find a high belief in this correlation.

There are two types of illusory correlations: expectancy-based and distinctiveness-based illusory correlations. Expectancy-based illusory correlation occurs when we mistakenly see relationships due to our pre-existing expectations surrounding them. In other words, I notice that a librarian is quiet and I assume that is because librarians are quiet people or introverts.  This is a form of confirmation bias - I see what I expect to see and this strengthens my belief.  If I see something that doesn't match that expectation, I either forget about it or I rationalize that is an exception to the rule - something known as the discounting principle. 

Distinctiveness-based illusory correlations happen when a relationship is believed to exist between two variables due to focusing too much on information that stands out.  It is this type of correlation that Hamilton and Gifford investigated in their classic study.

Hamilton & Gifford (1976) carried out an experiment with 70 American undergraduates (35 males; 35 females). Participants were shown a series of slides, each with a statement about a member of one of two groups - simply called group A and B. There were twice as many people in group A (26) as group B (13), so group B was the minority group.  The participants were told that group B was smaller than group A before starting the experiment. Each statement was about one individual in one of the two groups; the statement was either positive or negative. Each group had the same proportion of positive and negative comments.

Participants were then asked to rank members of each group on a series of 20 traits - for example, popular, social, intelligent.  After completing this task, they were given a booklet in which they were given a statement and then asked whether the person who did this was from Group A or Group B. Finally, they were asked how many of the statements for each group had been "undesirable."

Half of the group changed the order of measuring the dependent variables in order to avoid interference effects.  So, the booklet was completed before the trait rankings.

The results are as follows:

• On the trait ratings, group A was ranked higher than group B for positive traits and lower for negative traits.
• In the booklet, participants correctly recalled more positive traits for group A (74%) than for group B (54%) and more negative traits for group B (65%) than for group A (55%)
• Participants overestimated the number of negative traits in the minority group, but this finding was not significantly significant.

Hamilton & Gifford argued that this was because the minority group was by nature smaller in number, their negative behaviours appeared more distinct and appear to be representative of the group. So, one minority male is caught stealing and it appears to be related to the fact that he is a minority.  This demonstrates why negative stereotypes may be more common for minority groups than for the majority.   Such research has led to the modern practice in many countries not to report the race or ethnicity of people who have been charged with a crime

In the same report (Hamilton and Gifford, 1976), the researchers discuss a variation of their original experiment.

This study was carried out with 70 American female undergraduates. The procedure was the same as the study above - but this time, the participants were not told that there were fewer people in group B before starting the experiment. Each statement was about one individual in one of the two groups; the statement was either positive or negative. Each group had the same proportion of positive and negative comments.

The researchers created two groups, A and B, for which there would be no pre-existing stereotypes.  This increased the level of internal validity.

It was, however, highly artificial - meaning that ecological validity was low.  We cannot know to what extent this predicts the development of stereotypes under natural conditions.  In real life, there is much more context to making stereotypes - this could include economic competition, legitimate or unwarranted fear of others, institutionalized racism or prejudice and/or actual experiences with members of the other group.

The study was a repeated measure design.  The IVs were the positive or negative statements - and the size of each group. The conditions were concurrent.  This means that there were not two distinct conditions run separately. This eliminates the variable of participant variability - all participants took part in all aspects of the study.

Although the findings do show a difference between the positive and negative traits attributed to each group, the findings were not all significant.

There have been practical applications of this study.  For example, it has been found that doctors tend to "over-remember" poor health practices in obese patients compared to other patients (Madey and Ondrus, 1999).  It suggests that stigmatized patients suffer more from an illusory correlation bias than do non-stigmatized patients.