Saturday, February 23, 2008

Armenia Exit Polls 2008

The press reports that the British pollster Populus exit polls showed that 57.1 percent of the Armenian voters chose Serzh Sargsian, followed by Levon Ter-Petrosian (17.4 percent), Arthur Baghdasarian (14.6 percent), and Vahan Hovhannissian (5.98 percent); the remaining candidates received 1.6 percent. The interviews were conducted in 126 polling stations (clusters) in Armenia. These are fairly close, and within the poll's margin of error, to the official results where Sargsyan was declared the winner with 52.86% (863,000 votes).

It has been some time since I taught statistics and designed samples (not for polling), but the process of polling and extrapolating to the voting population is conceptually straight forward. First you design/select a sample with some criteria in mind to minimize sampling errors, and next apply appropriate weights to scale up the choice of those interviewed after they have voted to the size and demographics of precincts.

There are a number of problems with exit polls. For instance, the active partisan supporters of one candidate may be more likely than the supporters of other candidates to participate in exit poll interviews (selection bias). In addition, the young are more likely to participate in interviews. Furthermore, the age of the interviewers may matter as younger interviewers do not get high response rates from older voters. Of course, some of those interviewed may lie about the choice they have made (response bias).

Once the interview process is over, the next step is to weigh the sample to the size of the precincts and adjust for the non-responses say based on age and sex. One may rely here on the demographics of voters from previous elections in each precinct. While the size and demographics of the population are pretty well known (with few year lags), that of the voters is more problematic.

Exit polls are powerful indicators of corrupted elections. But without adjustments for sample selection and demographics, the results can be biased and at best meaningless. Consider the hypothetical case where most of those interviewed are old in precincts where the majority are young!

With the above in mind I was surprised to read the exit poll results from what ArmInfo reports as that conducted by the Alliance civil initiative, which I believe is organized by a large number of NGOs. Here is the full article:
Exit poll by Alliance civil initiative gives victory to Levon Ter-Petrossyan
2008-02-19 20:54:00
ArmInfo. Alliance civil initiative has conducted an exit poll at 100 polling stations. They have questioned 4,406 people. 3,550 of them (66%) agreed to answer.

The vice chairman of Alliance Gevork Melikyan says that ...
474 people (15.5%) voted for Artur Bagdassaryan,
40 (1.3%) for Artashes Gegamyan,
38 (1.2%) for Tigran Karapetyan,
8 (0.2%) for Aram Haroutyunyan,
195 (6.4%) for Vahan Hovhannissyan,
90 (2.9%) for Vazgen Manukyan,
9 (0.2%) for Arman Melikyan,
1084 (35.4%) for Serzh Sargsyan,
1152 (37.7%) for Levon Ter-Petrossyan.

Melikyan says that it was an independent survey. They planned to question 0.2% of all voters - 2,300 people.
This is strange. The sample design and selection seem odd. And what are we to make of the sample sizes of 2300, 4400, 3550, and the 3090 (sum of the people voting for individual candidates). Why are the sample estimates (1152, 1084, ...) reported without any adjustment for the size of the voting population (precincts), and seemingly not any of the demographics? What do the above figures tell us about the likely winner of the elections? Clearly, and unless ArmInfo got it wrong somehow, at best this is sloppy work. None of the foreign news services, except for Radio Free Europe (RFE/RL), reported its findings. Instead, they reported those of the reputable Populus.

No comments: