Morning everyone!
Sujit, thanks a lot for this non-medical Epi post!!
Layth, Fathima and Bethany, thanks for your comments. I really enjoyed and learnt reading them!
Some thoughts from my side:
1. What epi study design is an election poll?
I agree with all. This seems a population-based cross sectional study with descriptive data (outcome: presidential voting intention).
I guess another cross sectional analytical study could address the correlation between sociodemographic characteristics and political position. In this case, we would need to deal with confounders, as you can have a political position and in some point you can decide to ´punish´ your candidate (voting another candidate) because of bad governance, or whatever... I guess that's why it is important the proportion of undecided voters in polls designs?
2. How would you improve polling for the next election?
I have added below some reflections, ideas about how to minimize bias and therefore how to improve polling.
Selection bias:
It seems that the main issue for pollsters is to construct a representative sample of the target population (US eligible voters) due to low poll respondent rates (6% compared to 50% in the 80s) and I would also add because of survey design issues (often linked to costs/budget issues). This (low response rates and survey costs) makes almost impossible to eliminate selection bias. So the key point would be how to reduce selection bias:
1. Increase response rates to reduce selection bias. I agreed with Fathima that it would be helpful to conduct perhaps a more qualitative study to explore ´who´ are those who do not respond polls, and to understand why people do not respond telephone surveys (in this article, there are some reasons
https://www.pewresearch.org/fact-tank/2019/02/27/what-our-transition-to-online-polling-means-for-decades-of-phone-survey-trends/ ) and under what circumstances they would respond election polls. Perhaps, the poll industry needs to change the telephone survey protocols or just replaces the data collection technique (instead of random telephone surveys with live interviewers, use online polls, as proposed by Pew Research Center, or text messages, proposed by others. Bethany mentioned different options as well!
2. Construct a representative sample to reduce selection bias. Underrepresentation of specific population groups is one of the major source of selection bias in all election polls. I think it is important to have up to date descriptive data on the eligible voters, stratified by age, sex, residency, socio-economic and education levels, ethnic groups, political approaches and also by types of voters (e.g. undecided or late-decided, refuse-to-vote, new voters).
With this information (and good budget) you could combine stratified (e.g. sociodemographic parameters) and multi-stage (e.g. regions, districts, etc.) sampling strategies to get a better representative sample. I would concentrate efforts in those regions with lowest response rates and among relevant groups (e.g. undecided, new voters). For example, I would wonder how would affect the response of young new eligible voters´ (the millennials) to the polls´ results.
As Layth mentions, what seems interesting is that some studies found that there is no association between response rate and accuracy (Pew Research Center). I would like to read these studies, as it seems that the response rate or sample size is crucial in all quantitative studies, isn´t it?
Information bias:
Another information/measurement bias could be related to the survey protocol/questionnaire? Obviously, you need to have some knowledge about the country election system (Congress, Senate), but simplifying if you ask binary questions (Biden vs Trump) and you ignore questions about the election of Senate candidates, you might be missing info of those voters who vote both Biden and Republican Senate candidates, and got some misleading information.
I would also like to mention in this post some lines about Social desirability bias, which is often a source of information bias in election polls. Social desirability bias refers to the tendency of research subjects to give socially desirable responses instead of choosing responses that are reflective of their true feelings. In the case of this election, different factors might lead to this type of bias. As mentioned in the article, “some combination of the coronavirus, Mr. Trump’s reaction to police brutality and his erratic behavior at the first debate had put Mr. Biden within reach of the most lopsided presidential win since Ronald Reagan’s in 1984”. This influential factors might have led people not to give true responses about their vote and as a result, the polls might underestimate Republicans´ voters in liberal areas, where supporting Trump can be less socially acceptable than in conservative areas.
It seems that media including NYT are re-evaluating how they portray polls in future coverage, giving them less prominent coverage. I think this is really interesting, because media has a massive influence not only on the politicians´ campaigns, but also on the voters´ opinions, attitudes and actions. As said before, voters may change their mind along the campaign process, so I agreed with Fathima about questioning the validity of the polls´ results, and that conducting polls closed to the election day would give more accurate results. However, I think it is not allowed by law campaigning and conduct polls close to the election day, at least in my country, as the polls results could influence the final election results.
From an epidemiology point of view, polls can never be perfect and it seems that there is an acceptable “range of historical polling errors” in US, so the point is how to reduce “random error” and systematic error... I guess the answer is with the help of stats experts (to improve inferences from sample to population) and good insights into the parameters of the population under study, respectively.
I must say that I am not epidemiologist. Actually, I am journalist with some years behind working in Politics, covering some political campaigns in my small Basque country and Spain, and “translating” some polls´ results/stats to the wider audience. I have ambivalent feelings about the need of election polls. I guess we can live without election polls, they are actually thought to cover different interests... however, I also think these polls may play a role in encouraging people to participate in the elections. In some countries there is a big problem related to the ´absent voters´, which leads to low participation rate in the elections, which means that governments/decisions may not be representing accurately the population. But this is another debate!! Think also that not all those who respond polls go later to vote!
Have a good day, guys!
