ASA Connect

 View Only
Expand all | Collapse all

margin of error in political polls

  • 1.  margin of error in political polls

    Posted 10-20-2019 17:51
    it seems that the declared margin of error in political polls is computed with respect to the hypothetical penetration of 50%. Because the margin of error for penetrations decreases with it moving away from 50% towards 100% or 0% symmetrically, the common man may misinterpret this number all pollsters show as a single number while the candidates' penetrations differ even very much. For the Democratic Party, for instance, some penetrations are around 1%, while some others are around 30%.
    I suggest that, in view of the high interest people use to place on political polls, particularly during election years, ASA should issue  a recommendation to pollsters in order to have a personalized margin of error. In this way, people would understand the basic fact the small penetrations have a small margin of error in comparison to higher ones. For instance, in a sample of 500 respondents, a 30% penetration would have a margin of error of 0.04 while a penetration of 1% would have a margin of 0.008! Poor penetrations are therefore more precise than higher ones. This  can be observed empirically looking at all polls together.  Assuming no bias in all polls, they can be considered independent replicates of the same poll where the only source of variations can be the time span where the poll has been taken and the various data collection inconveniences, like interviewer/interviewee interactions, data collection mode or similar. Therefore, poor performers have two bad news: not ony they perorm badly! In addition their poverty is almost certain.

    ------------------------------
    [Ulderico] [Santarelli]
    [Las Vegas][Nevada]
    ------------------------------


  • 2.  RE: margin of error in political polls

    Posted 10-21-2019 09:02
    This is an excellent and timely topic, and probably much wider in application and of more general interest than the author has suggested with his excellent and insightful recommendation.

    While I am admittedly a complete novice at the practical empirical aspects of sampling theory as applied to polls of human respondents, political polls in particular, I find the subject fascinating and very timely. Yet I also find myself disturbed as a statistician that the margin of error concept is so poorly understood by the public and poorly explained to them by the professional polling organizations we all rely on. This is indeed an opportunity for the ASA and those among this online community who have the technical skills to address the recommendation proposed by Santarelli and to discuss it's broader ramifications.

    Ulderico Santarelli's conclusion:
    "Therefore, poor performers have two bad news: not [only] they perorm badly! In addition their poverty is almost certain."
    is a profound insight which is poorly comprehended by laymen and the statistically literate alike because like many statistical outcomes, it appears counterintuitive. Perhaps some visual interpretations of this and related phenomenon would help.

    Tom

    Thomas D. Sandry, PhD
    Industrial Statistical Consultant, Retired


    Tom

    ------------------------------
    Thomas Sandry
    ------------------------------



  • 3.  RE: margin of error in political polls

    Posted 10-21-2019 11:48

    Pollsters don't read ASA statements; there is a different professional organization  that they belong to, though -- American Association for Public Opinion Research (AAPOR; http://www.aapor.org). I would encourage you to look at their Transparency Initiative requirements in reporting the poll results, as well as their statements on margins of error. To understand how the assumption of "no bias in polls" is being relaxed in that discipline, read how poll aggregators like fivethirtyeight.com work.

    P.S. There is no need to reinvent the circular-shaped means of propulsion. There are several dozen thousand person-years of research in survey methodology, sampling, polls, and public opinion. If you are really interested, take online classes at JPSM (http://jpsm.umd.edu), join AAPOR and subscribe to their webinars, come to their conference (mid to second half of May)... and then maybe in three or four years time we can compare notes.



    ------------------------------
    Stanislav Kolenikov
    Principal Scientist
    Abt Associates
    ------------------------------



  • 4.  RE: margin of error in political polls

    Posted 10-22-2019 12:19
    Does not the margin of sampling (as I supposed that's what we're talking about here) error assume that the obtained sample is a probability sample in which every element in the population of interest has a probability greater than zero of being selected?
    Is it not the case that nowadays the ability to obtain a (near) probability sample is, with few exceptions, a governmental privilege (e.g. the BLS's CPS)?
    Does it make sense to speak of "margin of sampling error" under these conditions?
    Error no doubt there is, but then how does one estimate its magnitude?
    (Disclosure: I am a member of AAPOR.)

    ------------------------------
    Dominic Lusinchi
    Independent researcher & consultant (retired)
    San Francisco, Calif.
    ------------------------------



  • 5.  RE: margin of error in political polls

    Posted 10-22-2019 13:09
    of course this is a major point. Sample design is a critical element of the complex porcess we call Survey. A probabilistic sample guarantees unbiased sample, on condition that the actual sample is executed perfectly without any non-response and any substitution. Probabilistic samples are mandatory at design stage. However, actual samples may differ at a great extent.
    I just quote the two characteristic biases in election polls
    1. social desirability bias, whereby respondents do not reveal their true choice because fearing to be blamed when a candidate is badly hit by the media. "Deplorables" who voted for Trump probably did not reveal their choice to pollsters. This has been a major factor for the distance between the election outcome and almost all the surveys in 2016
    2. commitment bias, whereby respondents, possibly keeping a weak interest, anyway respond to the survey with the top of mind candidate.
    In additon, when you read that they "interviewed likely voters" even the Universe becomes a problem. What does it mean having a Unverse subject to daily changes? And what is the Universe of "registered voters"? Weighting should be done with reference to the Universe in question. Do pollsters know the distribtution by age-sex of the Universe of registered voters or do they take the overall population marginals? 
    Last but not least, in saying that the candidate A is significantly ahead, one should use the a margin of error 1.41 (square root of  2) greater because the variance of a difference is the sum of the variances of the two penetrations in the comparison. 
    Ulderico.

    ------------------------------
    [Ulderico] [Santarelli]
    [Las Vegas][Nevada]
    ------------------------------



  • 6.  RE: margin of error in political polls

    Posted 10-22-2019 14:16
    To your point about "social desirability bias," the AAPOR committee that analyzed the 2016 polling results concluded that it does not appear to have contributed substantially to the polling error.
    Another point we should add to the mix, and one which was brought up by an ASA member back in August, under the heading "Election Polling and Security" (and which received no reaction) is: we assume that the polls are wrong and the vote count is right. Is that a safe assumption? In light of what happened in Florida in 2000 and in 2004 (also in Ohio and Pennsylvania), perhaps not.
    We have ASA members concerned about this issue. Philip Stark had been advocating "risk-limiting audits" for some years now.
    Just a thought...

    ------------------------------
    Dominic Lusinchi
    Independent researcher & consultant (retired)
    San Francisco, Calif.
    ------------------------------



  • 7.  RE: margin of error in political polls

    Posted 10-23-2019 12:47
    Hi Dominic,

    I do think it is a very safe assumption that the vote count is right.

    Voting in federal elections in the United States is administered independently by each state. The unit of voting administration is typically the county, although there are regions with non-existent or weak county governments, such as New England, where the unit of voting administration is the town and city. There are over 3,000 counties in the United States.

    I think the fact that thousands and thousands of election bodies act without centralized coordination, and that election officials at the local level come from different parties, really does mitigate against any systemic issues in vote counting. The only cause for concern I can see are extremely close elections where minor errors one way or the other swing the outcome. However, that's not really a design flaw. Every measurement comes with error, and unfortunately our body politic is become so divided that a few hundred votes can matter, when in the past they didn't.

    I really think these sort of concerns are coming from politics, particularly on the Left. Errors are going to happen and that's perfectly fine. Systems like first-past-the-post voting and the Electoral College actually help ensure that the final result won't be swayed by minor administrative problems (or outright fraud) at the local level. This is why I have confidence in the correctness of the voting results from our federal elections.

    Best regards,

    Greg

    ------------------------------
    Gregory Csikos, CPA, CFE, GStat
    www.csikoscpa.com | greg@csikoscpa.com

    1229 Chestnut Street PMB 221
    Philadelphia, PA 19107

    P: 215.341.4206
    F: 267.392.6310
    ------------------------------



  • 8.  RE: margin of error in political polls

    Posted 10-24-2019 16:23
    Thank you for this, Greg. I am sure that in the majority of cases, you are right. However, I read enough (although probably not sufficiently in depth) about what happened in 2000 and 2004 (although the issues are somewhat different in each case) to be somewhat concerned. Those who wrote about the issues related to those elections don't strike me as being motivated by a political agenda. That said, I am no expert here, and I do not want stir this discussion away from the original topic; so I will leave it there.
    Best regards -- Dominic

    ------------------------------
    Dominic Lusinchi
    Independent researcher & consultant (retired)
    San Francisco, Calif.
    ------------------------------



  • 9.  RE: margin of error in political polls

    Posted 10-21-2019 14:29
    Good point.

    Another of my pet peeves about media reports on political polls is this: Many news items report the *difference* between two candidates, e.g. "Warren is leading Biden by 4 %-points in the latest poll". That is, I believe, a more relevant report than a given candidate's top line support. But they still report the top line margin of error. It's not uncommon for reports to add "... outside the margin of error +/- 3%". Implying that Warren's lead is significant, when it is well within the margin of routine sampling variation -- since the margin of error on the *difference* between two candidates is roughly twice the MoE of a single candidate's support.

    ------------------------------
    Stefan Sharkansky
    Usefull Work, Inc.
    ------------------------------



  • 10.  RE: margin of error in political polls

    Posted 10-22-2019 19:03
    Question: is this a conversation about polling accuracy, or about how the media reporting of the accuracy of polls often so very poor? How to interpret polls, or how interpret media reports of statistical results? Both conversations are needed, but they aren't the same conversation.

    ------------------------------
    David J Corliss, PhD
    Director, Peace-Work www.peace-work.org
    davidjcorliss@peace-work.org
    ------------------------------



  • 11.  RE: margin of error in political polls

    Posted 10-23-2019 12:48
    This is about communication. I think that the fact that somebody having 1% penetration enjois a margin of error of 0.008 can't be retraced by the tv listener or paper reader when one communicates that the margin of error is 4%. Listeners, in my opinion, apply the same margin of error to all the people in the table and this is seriuosly misleading.

    ------------------------------
    [Ulderico] [Santarelli]
    [Las Vegas][Nevada]
    ------------------------------



  • 12.  RE: margin of error in political polls

    Posted 10-23-2019 10:06
    In the original post there were several references to penetration. One of the issues with any stats article or explanation is the use of terms without definition or explanation. As one not terribly experienced in polls but an experienced stats guy, I do not know exactly what penetration means. An example would be helpful.

    ------------------------------
    Michael Mout
    MIKS
    ------------------------------



  • 13.  RE: margin of error in political polls

    Posted 10-23-2019 11:23
    this is an old problem. Unlike Mathematics, just a few terms have a true definition in statistics. What I call penetration correponds to what usualy they call percentage or prevalence in epidemiology.  I don't like percentage becuase too generic. For instance, percentage is also what the Bank pays for your deposit. Therefore, penetration means the ratio between the frequency by some criteria, in Political Polls candidates' names, and the sample total.

    ------------------------------
    [Ulderico] [Santarelli]
    [Las Vegas][Nevada]
    ------------------------------



  • 14.  RE: margin of error in political polls

    Posted 10-23-2019 12:47
    We should be honest and get out of the political polling business.  Bias trumps sampling error.  Most polls get a response rate of under 20%, respondents tend to lie, and the sampling frames even under more ideal conditions are not reflective of the true target populations. For example, other than exit polls, you never know if a respondent will actually vote.  I taught Survey Sampling in the 1990s, and we talked then about the fatal flaws in getting approximately unbiased estimates.​

    There are plenty of good applications for survey sampling but political polls are not among them.

    Jon Shuster

    ------------------------------
    Jon Shuster
    ------------------------------