[EM] Nader and safe states: a statistical reanalysis of 2000 exit polls

[Jameson Quinn]

As a voting reform activist, I am very interested in the question of voting
strategy. Specifically: how do voters respond to strategic incentives? Some
voting systems, such as range voting, have extra advantages for an
electorate full of voters who ignore strategic incentives; others, such as SODA
voting (Simple Optionally-Delegated
Approval)<http://wiki.electorama.com/wiki/SODA_voting_(Simple_Optionally-Delegated_Approval)>
have
results that are more robust to strategy.

Plurality voting, the most-commonly used voting system, is also the
absolute worst in terms of its extreme strategy incentives. Voting for
anybody but one of the two frontrunners is strategically stupid. And it's
clear that voters understand that; third-party candidates consistently get
far less support on voting day than they do in polls.

But the question is, how sophisticated is voter strategy? The 2000
presidential election provided a natural experiment. Voting for Nader in a
"safe state" was strategically safe; but, as the Florida results
dramatically demonstrated, doing so in a "swing state" was not, if you
cared at all whether Bush or Gore won. So a game theory model of rational
voters would predict that, for a pair of similar voters between a safe and
a swing state, the one in the safe state would be more likely to vote for
Nader.

I've just done a statistical analysis of 2000 exit polls¹ of 13,224 voters
nationwide. Simply taking Nader's overall vote share between safe and swing
states, and doing a naive correction for basic demographic factors, there
is no significant difference. But when I do a more sophisticated
demographic correction using matching, I find that a given safe state voter
was about 1.3 times as likely to vote for Nader, with a weakly significant
p value (0.0497).

To me, that seems like a surprisingly low factor. And probably, in
hindsight, most 2000 voters would agree. A well-known result is that, in a
poll months after the election, the number of people who admitted voting
for Nader was about half it was in the exit poll, suggesting that the other
half Nader voters regretted their decision. So this new analysis may be
cause for the "Nader haters" to grit their teeth once again at the
strategic naivete of Nader voters.

But the larger implications for voting system design are more encouraging.
This doesn't mean that voters won't be strategic; as I said earlier, the
evidence for widespread voter strategy in plurality elections is
overwhelming. Rather, it's that levels voter strategy will tend to respond
to long-term and pervasive incentives, rather than short-term and local
ones. That's good news for voting systems; it means that specific voting
pathologies like the chicken dilemma, in which two near-clones battling a
single opponent leads to possible strategy problems, are probably not a
huge issue.

-----

Here are the gory details of my analysis:

I searched google for lists of swing states from just prior to the 2000
election, and found 5 such lists. I defined "swing state" as anything that
showed up on 2 or more of those lists. That meant: FL, MI, MO, NM, OH, OR,
PA, WA, WI. I defined "safe state" as any state that ended up with a 2-way
margin of 7 or more points. Thus, there were a few states that were in
neither category, and were excluded from my analysis: AK, IA, ME, MN, NE,
NH, TN, WV. So, safe state voters were considered "controls" and swing
state voters were considered as "treated".

One of the questions on the exit poll was "political philosophy"
(liberal/moderate/conservative). I calculated the average political
philosophy of each precinct in the exit poll, and excluded 4 1 treated and
3 control precincts as having no plausible ideological match in the other
group.

I then did matching to improve demographic balance. I did exact matching on
race and church attendance, which were two of the poorest-balanced
variables; and propensity score matching on age, city size, income,
marriage status, education, religion, and precinct philosophy. Since
missing data was an issue for all of these variables, I had to include a
missing indicator for all of them in my propensity score model. This gave
me 3661 matches. Balance for the propensity model variables was improved,
but since I was trying to match so many variables at once, there was still
significant lack of balance; so I resolved to correct for all these
variables again when measuring my outcome. It's also worth noting that the
swing states were significantly whiter overall than the safe states, so
black and asian controls in particular were unlikely to be selected as
matches.

All of the above was done in the "design phase"; that is, without looking
at the outcome variable (Nader votes).

Once I'd done that, I did a logistic regression for Nader votes. The
variables allowed in the regression were all of the ones used in the
propensity match, plus two more questions: "How would you feel if Bush
wins" (excited/optimistic/concerned/scared) and "How do you want the
winner's policy to compare to Clinton's" (more liberal/same/more
conservative). These last questions were not matched for balance, but the
full spectrum was adequately represented in both treatment and control. The
estimate effect of the treatment indicator was 0.243, with the
aforementioned p value of 0.4967.

¹ Voter News Service. Voter News Service General Election Exit Polls,
2000. ICPSR03527-v2. Ann Arbor, MI: Inter-university Consortium for
Political and Social Research [distributor], 2004.
doi:10.3886/ICPSR03527.v2
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.electorama.com/pipermail/election-methods-electorama.com/attachments/20131203/9a0cdc1a/attachment-0003.htm>