Hi,
A short message on the polls of the 2024
US presidential election. I decide to follow the polls early on this
time since we already know who are the candidates. This will allow me to
better follow what is happening and the quality of the polls. I will
introduce two graphs, one on the global trends in vote intention and a
second one on the trends according to mode of administration and
sampling sources combinations.
Trends in voting intentions since January 1st
This
first graph portrays the trends in the proportion of support for each
main candidate on what is called the 2-party share, that is, the sum of
voting intentions for Trump and Biden. This is the only comparable
metric. It is important to note that pollsters do not always publish
comparable estimates. Some include the small candidates, other don't.
Some group together the small candidates and the "Don't know". Some ask
the question twice, once with all the candidates and the second time
with only Trump and Biden (we selected only the first question for
analysis). Most pollsters do not exclude the "don't know", a proportion
that differs substantially between pollsters. In short, it is essential
to compare using the 2-party share.
The graph illustrates
what happened since January 1st. As you can see, the gap between Trump
and Biden seems to have decreased. Trump is still ahead but by a very
slight margin. However, we see that some polls differ substantially from
the others: Some place Trump way ahead, other put Biden ahead of Trump.
Be reassured, these "outliers" do not substantially influence the
trends presented here given the method used (local regression). However,
is there a possible explanation?
Trends according to mode-source combinations
The
following picture shows the same data for two types of mode-source
combination. One type of combination is Web Opt-in panel only. The other
type is Multiple modes, a combination that includes at least partially
the use of random or quasi-random sampling. I have shown here
that mode-source combinations had a substantial impact on estimates in
the US 2020 election. The question is whether the impact is present from
the beginning. It was not that evident until recently.
There are three types of mode-source combinations, described in the article referred to above. The third type is "one mode, random or quasi-random". It includes telephone polls. There are not enough polls in this category for analysis right now. Therefore, these polls are not included. The Activote polls are not included either because we cannot categorize them (see below).
The graph shows the trends in support for Biden portrayed by the two types of mode-source combination. As during the campaign in 2020, web opt-in polls seem, on average, not to detect any movement, besides a small linear increase in support for Biden, while polls using multiple modes do. They show an increase followed by a recent decrease. Who is right?
The recent decrease
portrayed by polls using multiple modes may be partly due to some
specific practices. For example, the "outlier" putting Biden under 45%
recently is a Rasmussen Reports poll. I decided to check and realized
that the question on voting intention followed 10 different questions on
immigration policies and preferences, a topic that is quite polarising.
This practice may have biased the estimate in favor of Trump. However,
even if I drop this poll, the difference between the two combinations
are still present.
This being said, Rasmussen published its
questionnaire, a practice not always present. Whitout the questionnaire,
we can't know the rank of the question on vote intention and whether
the questions asked before it may bias the estimates. My understanding
is that the current practice is to ask about voting intention at the
beginning of the questionnaire and not ask questions that are polarising
before the vote intention question.
A second pollster used a methodology that is quite disturbing. Activote seems to rely only on people recording their vote intention on a free application available on phones and computers. The application is always open so that people can answer when they want. Activote publishes estimates when they feel they have enough answers (200 for state polls). The polls are conducted over 10 to 15 days. Activote does not say whether they check for people answering multiple times. Their estimates are quite stable, around 48-49% for all polls.
Conclusion
Biden and Trump are close, very close, everybody knows. But we also have to be aware that the way the polls are conducted may have an impact on the estimates that are published. In addition, when looking at polls, it is important to check all the information that we may have -- or not -- on the methodology, including the questionnaire. As a general rule, the less information you have on a poll, the less trusty the poll is. But sometimes, the information provided also leads to doubt about the numbers published. However, it still gives you some relevant information.