The Difference between Provisional Votes and Counted Votes in November 2016 Exit Poll
Before examining the exit poll results for provisional voters and counted voters, it is worth noting that for the five sites we collected data on, the percent of provisional voters with respect to the total number of counted votes has a near perfect correlation with the percent of registered Party Members for those sites (see Table below). The Democrats had a correlation of 0.9677, while the correlation was slightly higher for both the Republicans in the opposite direction. The party percentages are not independent of each other, so we expect similar correlations.
The Winfield results are not included in this analysis due to the low number (13) of provisional ballot surveys from that site. Urban Wichita is included because the concern regarding paper ballots being contaminated with provisional ballot voters will, even if true, decrease the probability of finding statistically significant differences between the provisional ballot votes and the counted votes.
Contamination the other direction is a concern for SE and SW Wichita as they have higher rates of provisional ballot voters than other voters. Contamination either direction will dilute the probability of seeing a statistically significant difference, it will not increase the probability of a type I error, so conclusions of statistically significant differences will hold even if some erroneous mixing of the groups occurred due to respondent error.
These results are also independent of any latent response bias in the survey sample due to party affiliation. If there was a party bias to responding the exit poll, it could be presumed consistent relative to being found unregistered or without adequate ID, thus necessitating a provisional vote. Results are shown below.
|Site||Total Votes||Prov Votes||Prov Vote %||% Reg Rep||% Reg Dem||% Reg Lib||Exit Poll Prov||% Prov||% Exit Poll|
We can use the binomial test for the provisional versus counted ballots. They are two separate samples and one is not a subset of the other, which rules out the hypergeometric test used with the machine and paper ballots analysis. The binomial test was done for each candidate and judge response with the results shown below in Table 11.
The differences in vote share between the provisional and the counted voters in the exit poll will fit student’s t-distribution. We will determine if the provisional voters in our exit poll were statistically significantly different from registered voters with proper ID.
There isn’t sufficient data to warrant reporting results for the Independent candidate or the Green party candidate. The candidate differences are not independent within a race (they will sum to zero) but the results for the three candidate races are independent of each other.
Each Judge is an independent contest relative to the other four judges and the three candidate races.
A paired t-test was performed on the vote share ratios for the Republican, Democrat and Libertarian candidates in all three candidate races and another for the yes, no, and blank responses for the five judges to determine if there was any bias with respect to any particular party or retention vote.
|Judges – Yes||7.93%||5.67%||10.19%||<0.0001|
|Judges – No||1.03%||-1.66%||3.72%||0.4313|
|Judges – Blank||-8.96%||-10.64%||-7.29%||<0.0001|
Republicans show a statistically significant difference with provisional voters being between 1.81% and 8.41% (average 5.11%) less likely to vote for Republican candidates compared to voters whose ballots were counted that day.
Libertarians show a statistically significant difference using the t-test with an increase between 0.59% and 4.31% (average 5.11%) in vote share from the provisional voters.
Neither the Democrats nor the ‘Write in/Left Blank” responders showed a statistically significant difference across the four sites and three races between the provisional voters and the regular voters with the t-test.
Provisional votes for the Kansas Supreme Court Justices show a distinct pattern of provisional voters being nearly 8% (average 7.93%) less likely to vote yes for all judges than the counted voters and nearly 9% (average 8.96%) more likely to indicate that they did not vote in that contest. The uncounted provisional ballots would not have altered the outcome of the races studied.
*LCL and UCL refer to the Lower and Upper Limits of the 95% Confidence Interval. If one is positive and the other negative, then we can presume there is no statistically significant difference between the counted voters and the provisional voters.
One thought on “Provisional Voters Analysis”
Thank you Ms Clarkson!