3
$\begingroup$

I have a data set with 1500 individuals who were surveyed about many things including demographics, attitudes and opinions about politics and economic status. The most important question in the survey was about vote intention on the next election for president. The question had the four candidate options, plus options for 'I don't know yet', and for 'I'm not going to vote'. Among the 4 candidates, 2 are strong, and 2 are very weak.

The question I'm trying to answer is: among the undecided (the ones who say I don't know yet), what are the characteristics of the voters who are more likely to vote for each of the candidates? For example, are male or female among the undecided more likely to vote for candidate A?

How can I do that in R?

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ pretty broad question but perhaps start here $\endgroup$
    – Anthony Damico
    Commented Jun 3, 2014 at 4:41
  • $\begingroup$ This is not a statistical question per se, methinks, but rather a substantive one. Statisticians may or may not give you the best advice. To really figure it out, you need to read political science, polling and public opinion studies that do something similar; identify statistical methods that they use (especially if they use these methods with a verifiable degree of success), and then come back to this site and ask "How can the Best Linearly Aggregating Hyberbole (BLAH) method be implemented in R?" $\endgroup$
    – StasK
    Commented Jun 10, 2014 at 21:05

2 Answers 2

Reset to default
5
+50
$\begingroup$

This is a pretty broad question, and hard to answer. Election prediction itself is a science, which requires a lot of experience.

The most important thing is to control for bias. Step out of your way, and control the heck for bias. It's not fun. It's as boring as it can get. But unless you have a representative sample, your results will be badly off. If you asked only Twitter users, the results will be completely useless.

With elections, even with a representative sample, it's apparently hard to get the prediction error below 1%. Because people don't answer truthfully. Because people are undecided. Because people change their mind suddenly. Because the brain says A, but the gut says B.

The prediction error you can't get rid consists of many factors. If you have a very good representative sample (carefully chosen telephone interviews) it will be lower than by doing exit polls (where you can only roughly control demographics by choosing representative locations). If your poll has many options, the error will be higher than if it is a A/B election. Here are some sources:

Improve this answer
$\endgroup$
4
  • $\begingroup$ Do you have sources for analyses which get just outside 1% or was that figure off-the-cuff? $\endgroup$
    – user44764
    Commented Jun 10, 2014 at 16:16
  • $\begingroup$ This is the error margin I remember from the news; probably during counting when they have some final data already available. But this obviously depend on your likelihood threshold. According to edisonresearch.com/exit_poll_faq.php "The margin of error for a 95% confidence interval is about +/- 3% for a typical characteristic from the national Exit Poll and +/-4% for a typical state Exit Poll." $\endgroup$
    – Has QUIT--Anony-Mousse
    Commented Jun 10, 2014 at 17:38
  • $\begingroup$ It depends a lot on the data that you have available. theguardian.com/commentisfree/2012/jun/06/… says "That's why the final (not preliminary) exit polls in Wisconsin had a theoretical margin of error of ±4 points, while a telephone poll of the same sample size would have a margin of error of only about ±2 points." (Because it's easier to get a representative sample via telephone, than by exit polls where results will cluster.) $\endgroup$
    – Has QUIT--Anony-Mousse
    Commented Jun 10, 2014 at 17:40
  • $\begingroup$ Thank you. It's worth noting that the best methods for voter classification (Murray et al. 2009 POQ) do not even get in the 80% range when considering vote validation. Thankfully the people who say they intend to vote but are lying are the same as people who do vote, so polls are still useful. $\endgroup$
    – user44764
    Commented Jun 10, 2014 at 17:42
0
$\begingroup$

To truly assess the characteristics of the undecided voters who are more likely to vote for each of the candidates, you would need a dataset with known outcomes. More specifically, if you had this current data with their actual votes in the next election appended to it, you be able to compare undecided voters who voted for candidate A vs. candidate B across variables such as gender, economic status, etc.

Since you do not possess these known outcomes (next election voting results), you could look into unsupervised learning techniques such as neural networks or clustering. My expertise does not lie in these areas, but I bet someone reading this can provide more sufficient detail! I hope this helps!

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.