Results of India general elections 2024 have thrown a surprise no one saw coming. NDA has secured a majority but BJP on its own has failed to secure the majority, unlike last two general elections.
No exit poll predicted this
scenario. As per exit polls, BJP was going to reach majority mark on its own
and NDA was going to win about 350 seats. But BJP has won 240 seats and NDA has
won 292 seats. The results seem to be beyond the confidence intervals projected
for the prediction. What does that say about exit polls?
There are
multiple possible answers to this question. I will rule out conspiracy answers
at the outset. I am not going ahead with argument that exit polls were staged
to help some agents. One interesting possibility that I might want to consider
is false answers from voters. Respondent’s response to exit poll enumerator can
be a strategic choice if respondent thinks that revealing what she truly voted
is not in her best interests. If she believes that many others would have voted
differently than how she has voted, she might respond that she too has done the
same, even when in anonymity of the booth she has behaved differently. But I think
most likely explanation is structural.
Exit polls
are based on sample which represents the underlying voter population. Essentially,
sample is drawn with the hope that sample properties will mimic the underlying
population properties closely. The large variation of sample properties from
later revealed population properties imply that drawn sample was not a good
representation of the population. So, first possibility is, exit polls are
based on wrong sampling.
I assume that those who do exit
polls try to pay attention to their sample as much as possible. The real
constraint is, we know, in some broad sense, about proportions of various
attributes in registered voter population but we do not know about same
proportions among registered voters who have voted. For example, there might be 30% voters above
age 60 among registered voters. But on the day of voting, 20% of the voters who
voted are of age above 60. But later information will not be available and
hence exit poll sample will have larger proportion of above 60 voters which can
lead to misleading estimation if voters above 60 vote considerably differently
than voters below 60.
The sampling problem is
accentuated by the fact that much of the proportional understanding, in terms
of gender, religion, social groups, still stems from 2011 population census. Though
broad national proportions might not have moved much from 2011, we cannot say
so at constituency level.
In short, drawing a
representative sample for each of the 545 constituencies to generate nationally
representative sample is empirically a very difficult task. So, exit poll
sample is very likely to be non-representative of population of voters who have
voted unless non-voting is prevalent at same proportion across voter attributes.
Also, if voter preferences are not varying much across voter attributes like
gender, religion, social group, age etc. then even such sample is likely to
generate estimates which are in proximity of the underlying voters who have
voted population.
So, exit polls will go wrong if non-voting
proportion is varying across different voter groups. Voter mobilization from
political parties can be correlated with voter groups. If voter mobilization
from local cadre of parties is different then it can result in varying non-voting
proportions. Other possibility is variation in voter preferences across voter
groups. Perhaps some voter groups had greater revision of preferences than
other groups before this election, leading to more differentiated range of preferences,
unlike previous elections where degree of differences of political preferences
across voter groups was lesser. In such cases, non-representative sample will
throw different results than underlying true voters.
One also wonders what purpose exit
polls are serving beyond providing excellent infotainment. One argument is they
aid the price discovery in the markets trading on future outcomes, like stock
market. Regime change can lead to policy change which can lead to differences
in future cash flows which will have bearing on valuations. Hence, those
interested in capturing value through speculation would like to know in advance
if there is going to be such change. In such cases, there should be a market of
electoral outcome forecasts, where speculators will demand forecasts and psephologist
will supply. May be such market already exits.
The exit polls are such a forecast, though with very limited time window between forecast and realization. And we saw how it played out in the stock markets. The point is not about this one instance. The point is it is bound to happen again. The demand and lure of forecasting is always going to be there, more so for event like national election. But considering the constraint of inadequate information on voter characteristics, exit polls will always remain mix of science (sampling) and prayer. We must treat them the same.