By Marcos Pérez Talia.
The polls have started to be published and with them the controversies. Just as Tereré Cómplice published several articles about them during the last electoral cycle, we will put the magnifying glass on them again.
Let’s start with the little known pollster: FASAC SRL. According to a report in the newspaper La Nación on February 20, it published a survey that gave a 13-point advantage to Peña over Alegre. Among the few antecedents of FASAC SRL, one notes that it projected a difference of 20.3 points in favor of Peña over Wiens in the internal elections of the ANR in December 2022, when the difference was 7.7 points. In addition, one can observe in his Twitter account tweets in favor of Peña (as the image below), indicating his preference and bias towards that candidacy.
Two other little-known pollsters that have been highlighted by La Nación’s article of February 20 are Oima Data and Multitarget. Oima Data has no track record and projected a difference of more than 20 points in favor of Peña against Alegre. Multitarget has a bad track record projecting a difference of almost 25 points between Nenecho and Nakayama in the 2021 municipal elections, when the difference ended up being 5.4%. In the ANR internal elections they projected a difference in favor of Peña of 20 points when it ended up being 7.7 points.
These three pollsters (FASAC SRL, Oima Data and Multitarget) are apparently new, the 2023 general election being the first in which they appear publicly. Their results are widely disseminated by the media of Grupo La Nación, owned by Sarah Cartes, sister of Horacio Cartes.
Among the inexperienced pollsters, with projections favorable to Alegre, is of Chiocchini y Asociados, which gives a margin of 3.6 points in favor of Alegre. This pollster also has no track record.
Now let’s move on to the group of pollsters with more experience. Among these are Ati Snead and GEO, which already published polls in several general and municipal elections since at least 2008. In addition, we include Ecodat, which appeared for the 2018 elections. Ati Snead gave in February 3 more than 20 points to Peña, Ecodat on February 20 10 points to Peña and GEO also on February 20 gives 3.7 points to Alegre. New Ati and GEO polls have already come out after these which we will analyze in a following article.
In this first round of surveys, GEO stands out with, at least, a better sample design.
Ecodat is the least experienced among the three and with a history of significant error in its projection. In 2018 it projected a difference between Abdo and Alegre of about 25 points, when the actual difference was 3.7 points.
GEO and Ati Snead are the most experienced. In the 2008 elections both projected Lugo’s victory. In the 2013 elections both Ati and Geo projected a narrow margin victory for Alegre over Cartes, a result far short of Cartes’ 8.9% victory. In the 2015 municipal elections Ati gave an 18-point advantage of Ferreiro over Samaniego in Asunción, while GEO gave almost 11 points. The difference was 10.5 points.
In these comparisons, GEO and Ati were quite similar, with GEO having a better approximation in 2008 and 2015.
Ati then was the closest in 2018, but in the 2021 municipal elections their poll gave 20 points in favor of Nenecho and the exit poll gave a difference of 13 points, when the final difference was only 5.4 points. In the ANR internal elections, his last poll projected a difference in favor of Peña of 16.4 points when the difference was 7.7%.
In conclusion, Ecodat has little experience with a very large error in the 2018 elections. Ati and GEO have a lot of experience, with mixed results, although GEO has better accuracy. While Ati had a good projection in 2018, both in the 2021 municipal elections and in the 2022 NRA internal elections, its projections were quite far from the results.
Let us now turn to another important point for analysis. Beyond history, we can analyze the “sample design” used by the pollsters, provided they have published the details.
The sample selected by each pollster should have certain characteristics consistent with voter registration and historical voter turnout data. For example, half of the pollsters should be male, since male voters tend to be about half. Also, Asunción and Central should be about 40% of the respondents, since that is the proportion in the voter list. Certain proportions should also be respected by age, political party preference, income range, urban/rural area, among others.
In the case of the Ecodat, Ati and GEO surveys we can analyze the sample by age range and compare with what the turnout was in the 2018 general election. GEO directly showed the percentages of their sample by age range. From the Ati and Ecodat polls we can infer the number of respondents in each age range, given that we have the published preferences by age range.
In the following figures we can see what the weight by age range of the 2018 election turnout and polls was:
Graph 1: Weights by age range. ECODAT
Source: own calculations based on TSJE and ECODAT.
Graph 2: Weights by age range. Ati Snead
Source: own creation based on TSJE and Ati Snead.
Graph 3: Weights by age range. GEO
Source: own creation based on TSJE and GEO.
Of the three, ECODAT is the pollster with the worst sample. Not surprisingly, it had a margin of error of about 20 points in the 2018 election. The Ati poll has a lot of young people and few over 60, relative to what happened in 2018. The GEO poll, on the other hand, is close to what it should be.
If a survey takes a bad sample, its results are not representative, and this is key. They are simply useless. It is the good choice of sample that allows a poll of about 1500 people to be representative of millions of votes.
Therefore, the Ecodat survey should be discarded as inconsistent. The Ati survey has serious problems that should be corrected in future projections. The GEO survey has a representative sample (at least in age range), so it seems to be well designed.
In this first round of surveys, GEO stands out with, at least, a better sample design.
Cover image: cambio24.com.co