View: How exit polls have repeatedly failed to predict the result accurately

By Atanu Biswas

Ever since the first exit poll was conducted in a local election in Kentucky, US, in 1967, it has been a history of controversy and entertainment for exit polls.

The sorry state of exit polls is illustrated again and again in India where many of the popular exit poll results fail to predict the final result with reasonable accuracy. No wonder people tend to treat them as pure entertainment.


There is enough doubt on the techniques and sample sizes of many polling organisations. Even if the sampling is random and representative of the population, and the sample sizes are adequate, other kinds of errors exist, including ‘non-response biases’.

With two contestants, A and B, suppose A’s support is 55%, and B’s is 45%. Assume that 40% of A’s supporters and 60% of B’s supporters participated in the survey. Among a sample population of 20,000, that means about 11,000 and 9,000 supporters of A and B, respectively. So, 4,400 supporters of A and 5,400 supporters of B took part in the survey, thereby predicting a victory for B.

This was the case, as believed by pollsters like Warren Mitofsky — the ‘inventor’ of exit polls — with Democrat candidate John Kerry in the exit polls of the US presidential election of 2004. Although Kerry was predicted to win by Mitofsky’s exit poll, George WBush won.

A small-scale survey prior to the exit poll could be attempted to estimate the percentages of the supporters of A and B who would be interested to participate in the poll. A properly designed and executed ‘pre-exit poll survey’ may actually yield ‘closer ‘numbers.

In addition, respondents often lie to pollsters, sometimes due to a ‘social desirability bias’ — they pretend to make the ‘politically correct’ choice to the pollster, without admitting their support to some controversial leader or party. They may even be afraid of admitting their support due to local political conditions, preferring to come across as riding the bandwagon.

For instance, 28% Latinos in the US supported Donald Trump, though the Republican candidate had said some fairly disparaging things about the community in his campaign.

The statistical technique, ‘randomised response sampling’, can be used to estimate the actual proportions supporting different parties.

With two parties, this may be implemented with the question, ‘Did you vote for Party A?’ along with a second question, ‘Is your birthday between January and June?’ the latter having a 50% chance of a ‘yes’ response, and the unlikelihood of the respondent lying about it.

The pollster can now estimate the proportion of voters who voted for a particular party more accurately.

If 1,000 respondents are interviewed, and 480 ‘yes’ responses are obtained for the first question, about half of them — nearly 250 —replied ‘yes’ to the second question about being born between January and June. So, there were 230-odd remaining ‘yes’ replies for the first question. This means that about 230/500 — 46% — voted for Party A.

In a multi-party set up like India’s, suitable modifications of this technique can be implemented. So, why not implement such techniques in exit polls? Understandably, it demands more effort and time, both for collecting and analysing the data, delaying the announcement of exit poll results by, at least, a few hours.

But, isn’t that better than announcing possible erroneous results that tend to be increasingly hilarious over time?

(The writer is Professor, Indian Statistical Institute, Kolkata)