Friday 28 October 2016

Anomaly in Statistics

In reporting on the American Presidential election campaign the BBC publishes a 'poll of polls' calculated by taking the median of the results of the most recent five polls.

When I checked today's figures I thought there was a mistake.

The Poll of Polls showed Clinton ahead by 5 percentage points, while the individual polls showed leads of 0, 2, 3, 7 7  with median 3.

However taking medians of the figures showing total support for each candidate gives 49 for Clinton and 44 for Trump, with a lead of 5 for Clinton.

So the difference of the medians is not equal to the median of the differences. That may be a reason for treating the median with caution.

In the case of opinion polls based on samples of different sizes, I'd be inclined to use a weighted mean, weighted by the sample sizes. That calculation gave means of 47.47 for Clinton and 44.18 for Trump, a lead of about 3 points for Clinton.



No comments :