Weighting Results for Undersampled R's in Exit Polls, Parallel Elections [Updated]

Weighting Results for Undersampling of Republicans in Exit Polls or Parallel Elections

Nov. 26, 2006, updated May 20, 2007
By Marj Creech, risenregan[at]earthlink[dot]net

I am currently working on an analysis of a four-precinct parallel election exit poll done on Nov 7, 06, demonstrating how this analysis can be used on more complicated data, including the large number of “Undeclared” voters in Ohio. Please contact me at honestelectionscow[at]gmail[dot]com for this new report.

The most common criticism of using parallel election results as data to reveal fraud in the official election, is that conservatives/Republicans will be under-sampled. Indeed that has been the case in the four official elections in which we have run parallel elections in Ohio, to a greater or lesser degree.

In the previous parallel election run in the May primaries, 2006, in three Westerville, Ohio precincts, all in one location, it was easy to tell who the Republicans, the Democrats, and the Independents were, both in the official election and the parallel election, since they were given ballots by party affiliation. The spoiler in the purity of party identification is that some Democrats registered as Republicans in that election in order to vote against Blackwell, by voting for Petro. But there is no reason to doubt that these party line crossers voted the same way in the parallel election, they asked for a Republican ballot to match how they voted officially.

It is a proven assumption, at least in Ohio, that progressives/liberals/Democrats participate in parallel elections in greater percentages than conservatives/Republicans of all those voting official on election day. In the primary mentioned above, we had 25% of all the Republicans who voted that day, voting in the parallel election, 30% of the Democrats, and 35% of the Independents. [need to check to get exact numbers!] The percentages seem to depend in part upon precinct composition: in majority Republican precincts we get higher percentages of Republicans (parallel participation/ total R official voters) participating than in majority Democratic precincts. And vice-versa. I have not analyzed these data in detail.

Another factor is how we, the parallel data gatherers, present ourselves. Voters are highly partisan at election time, and if they perceive that we are “left (or right) wing activists,” while they are the opposite, they are less likely to participate. We also tell them we are non-partisan, “not representing any political parties or candidates.” In the primarily Republican area such as the May primary in Westerville, we were in Republican territory, so we dressed in OSU colors (neutral at least) , and did not mention “machine insecurity” or fraud as a reason for collecting data, unless we were pressed. Several voters asked if we were there for the school bond issue, an issue that passed by over 80 percent, along mostly non-partisan lines.

In the recent Nov. 7, 2006, parallel election, in Clintonville, several voters responded to our request for their participation in the parallel election with, “ I don’t believe in polls,” or “We should just count the real ballots!” Also the National Republican Party has stated to the media that, “Exit polls are bad because they negatively affect Republican turn-out,” … “or under-report Republican votes.” Of course exit polls would not under-report Republican votes if there was not negative publicity to Republican constituents!

It is a fact that conservatives are less likely than progressives/liberals to believe that voting machine fraud is a problem (don’t have the citation, but I am thinking of the study done earlier this year on people’s belief in voting insecurity correlated to TV stations watched. CNN watchers were over 50% likely to believe in voting fraud, while Fox watchers were less than 1 percent!) And therefore if they believe that an exit poll is for the purpose of exposing machine insecurity, or even malfunction, they are less likely to participate. One can feel the reception change in primarily Democratic precincts from chilly to “Thank goodness you are doing this!” By our conscious effort to be non-partisan and promote “honest elections for ALL voters,” we eliminate some party bias to participation.

The rest of the bias can be corrected for, to a great extent. One must know the party affiliation of both all the parallel voters and all the official voters. From Board of Election records sent to us, we can look up the party affiliation of our parallel voters, most of whom signed in our parallel poll books . The BOE also will send to any citizen who requests it, the names, addresses, and party affiliation of all who voted in the official election , for the same precincts. Unfortunately, the Franklin County BOE has told us that this list of voters won’t be available till Dec. 15. Meanwhile, we have the list from the individual precincts of all voters through 4 pm and we are interpolating that data through 7:30 pm. We are making the assumption that voter party affiliation did not change much by percentage in the last 3.5 hours. We will check this assumption when we get the BOE data after Dec. 15.

But with the estimated breakdown by party of official voters and parallel voters , a comparison can be made of the percentage of official Democrat voters we got to participate versus the percentage of official Republican voters we got to participate. Then calculations can be done so that if the percentages are made to be equal, we can calculate how this would have affected the vote totals for a given race.

For example, if there are 100 voters in the official election registered as Democrats (call this number 100 OD—official voters Dem, and 80 registered as Republicans (80 OR), AND we get 50 Dems to vote parallel, but only 20 Repubs, of course the R. candidate is going to get fewer votes than if we had sampled the R’s in the same percentage. In this example we got 50% of the D’s to vote parallel, but only 25% of the R’s. To correct for this sampling bias we need to project how many votes the R candidate would get if we sampled the R’s ALSO at 50% (40 of our R voters out of 80) Since 50% is twice 25%, or double, we simply need to double the number of votes the R candidate got. This method assumes the R’s we would have sampled to make the percentage equal to that of the D’s, voted the same way as the R’s we DID sample. If R’s crossed over to vote for a Democrat candidate, such as many did in the Nov. 7 election, we are assuming that the same number did in the corrected R vote as we captured in the actual parallel R vote. We are also ignoring the Independent, or Undeclared Party voters, since it is much harder to guess which candidate they voted for.

In this same example, if our parallel election got 60 votes for Strickland, and 10 for Blackwell, AND we correct for the under-sampling of R’s (half as many percentage-wise as for D’s), we must multiply Blackwell’s votes by 2, to give us 20 votes for Blackwell, in our projection of how an equal sampling of R’s would affect the vote totals. Then the corrected percentage of votes for the governor’s race is 60/80, or 75% for Strickland, and 20/80 , or 25% for Blackwell. Uncorrected percentages were 60/ 70, or 86% for Strickland, and 10/70, or 14% for Blackwell. (The percentages will not add up to 100% if we consider the Independent voters.)

One can see by these numbers how under-sampling R’s can lead to inflated results for D candidates, if the under-sampling bias is not corrected for.

Here is the formula:

First determine the percentage of parallel D voters to total official election D voters. Call this PED/ OD.

Do the same for R voters. PER/OR

Call R candidate total PE votes RV. Call corrected RV “CRV”.
Call total PE Dem candidate votes DV.

Ignore Independents. (Independents, and/or Undeclared voters, are not ignored in the soon-to-be-released report on the General Election on Nov.7, 06 in Ohio, since there are so many.)


To get the corrected percentages:

CRV/ (CRV+ DV) = percentage of CRV

DV/ (DV+ CRV) = percentage of DV

This formula will also work in the event that D’s are under-sampled. That will happen when pigs fly, and/or when D’s lose faith in polling as a way to reveal machine/election fraud.