To explain Simpson's paradox let's start with a hypothetical example. During a particular summer, an experiment was conducted to find out the preference between two types of beverages: soda and lemonade. The data was drawn from two locations: city and rural. In each location, the gender and the choice of drinks were collected. The results are summarized as follows:

The odds ratio given that location = city is about 1.1, showing that females are about 10% more likely to drink lemonade than males. Because the conditional odds ratio given that location = rural is also 1.1, the same conclusion can be drawn.

Next, combine the results from both locations and form the following 2 by 2 contingency table:

The odds ratio of 0.86 shows that females are about 14% less likely to drink lemonade than males, rather than 10% more likely as was shown earlier! This is an example of Simpson's paradox.

In general, Simpson's paradox illustrates that the effect of an omission of a categorical explanatory variable can have on the measure of association between a categorical explanatory variable and a categorical response variable .

In the example, given the location variable , the conditional odds ratios show that the gender variable and choice of drinks response variable have a positive association, with positive log-odds ratios. However, when the location variable is removed, the marginal association between and is negative, with a negative log-odds ratio.

One reason for this apparent paradox is due to the dissimilar populations between the city and the rural groups. In the rural area, the majority of the test subjects are female, whereas in the city area, the majority is male.

For an excellent explanation of Simpson's paradox, please refer to the book below.

Bibliography

A. Agresti, An Introduction to Categorical Data Analysis, Wiley & Sons, New York (1996).