Exit polls, or post polls in the case of CSDS, do not just predict the winner in an election but also form the backbone of our understanding of how different groups voted and, hence, what worked politically and what did not. So when the exit polls differ significantly from each other, like they did for 2019 Lok Sabha elections, it affects our understanding of what worked politically, and what we know about politics and demographics.
Of all the exit polls published on May 19, only three have put out detailed data and are seen as relatively credible. For this analysis, we will look only at voteshare, and not seats, as it is more relevant to the central point. The voteshare projections come from conducting a representative sample survey, and then projecting what share of the sample intends to vote for each alliance. While India Today-Axis My India surveyed every constituency, CSDS and CVoter surveyed a sub-set of constituencies based on which they projected voteshare for each state.
Within those samples, pollsters also estimate what proportion of different sub-groups — first-time voters, or the poor, for instance — voted for a particular alliance. After the final results came out on 23 May, CSDS and CVoter adjusted, or re-weighted, their group-wise projections against the actual voteshare; since Axis has not yet done this, the numbers shown for them are their original estimates. These group-wise estimates form the basis for analysis about the most fundamental questions after an election: who voted for whom and why?
But on at least two key issues (more may emerge as the pollsters put more of their underlying data into the public domain), these three pollsters differ in such significant ways that they make it difficult to form a consensus about how different groups voted.
The first is the female vote. One of the major narratives of this election has been that women’s votes were crucial. Female turnout as a proportion of registered electors more or less equalled male turnout for the first time in history, and female voters outnumbered male voters in 13 states. (Nationally, though, male voters still significantly outnumber female voters.) If women voted differently from men in the 2019 elections, it could have had an impact on the results, and, more importantly, have implications for political parties’ messaging for the future.
The Axis and CVoter exit polls find that a larger share of female voters than male voters chose the BJP and its allies. While Axis finds no difference between male and female preference for the Congress and its allies, CVoter finds that a larger share of men than women preferred the UPA. While the difference in preferences may appear to be small, it could upend conventional wisdom; until now, CSDS polls have been the main source of scholarship on electoral issues, and have found that the BJP has had a historical disadvantage among women, and the Congress a historical advantage.
But here’s the problem: CSDS’s 2019 poll finds the opposite — a higher share of men than women still prefer the BJP and its allies, while a slightly higher share of women than men prefer the Congress. It essentially means that the trend from CSDS’ past National Election Studies holds true.
The second major area of disagreement between the three pollsters is on the Muslim vote. Here, while Axis and CSDS find similarity, and in line with historical trends, CVoter is a massive outlier.
Finally, on caste-wise voting patterns, CSDS data for the country as a whole is not yet available, but there is wide variation between Axis and CVoter on the OBC and upper caste vote.
How do we make sense of these differences? Unlike in the case of Today’s Chanakya in 2015, when the pollsters attributed their mistaken Bihar exit polls to an assistant making a coding error, these three pollsters are confident of their sampling, methodology and results. CVoter’s CEO Yashwant Deshmukh suggests a few possible reasons for his high Muslim numbers: Muslims in non-polarised states voting for the BJP, Muslim women voting for triple talaq, Shia voters choosing the BJP, and Pasmanda Muslims choosing the NDA alliance in Bihar.
Given that Axis came closest to estimating the overall voteshares correctly, should its group-wise voteshares simply be taken as closest to the truth? That’s one way of looking at it, but might not necessarily be the case, either. “Variation in estimated vote share of 2-3 per cent for party or alliance is well within the range of what one could expect from a sample survey. For analysis, this voteshare estimate is adjusted with the actual vote share by the process referred as weighting which is an internationally accepted statistical technique,” says Sanjay Kumar, director of the Lokniti programme at CSDS. If the group-wise sampling in the original sample was correct, which CSDS remains confident of, this adjustment is statistically passable.
The issue is of narratives. These seemingly small differences in groups voteshares will make a big difference to future narratives about how the 2019 election was won. Does the BJP still have to work on its appeal among women, or is that battle now over? Has Muslim fear been overplayed as a media narrative, which the Right-wing likes to suggest, or is the BJP still largely unacceptable to Muslims? When exit polls could not agree on the final results, the differences were not resolved on 23 May; these arguments will now continue for the next five years.
The author is a Chennai-based data journalist. Views are personal.