problems of polarization

Here is a form of study I have seen in various guises over the years: The authors are interested in the question of whether people in a society are becoming more “polarized” in their attitudes in some kind of global sense. So what do they do?

1. Find data that has a long repeated series on a variety of attitudes.
2. Select those items that have been continued in these data over a long period.
3. Look at whether responses to these items are more divergent now, or more divergent for some group, than they used to be.

For me, here’s the problem: questions only initially appear and stay on public opinion surveys so long as disagreement is sustained in the population. Imagine if we had really long time series on public opinion. Do you believe women should be allowed to vote? Do you believe alcohol should be outlawed? Do you think our state should secede? Do you believe fugitive slaves should be returned to their masters? Or, obversely: Do you think gay couples should be allowed to get married? Do you think people should be allowed to smoke in public places?

So it seems like the study is strongly biased against finding evidence of polarization from the get-go, especially given that I think it’s less likely for questions without disagreement to be asked in the first place then for them to be re-asked for comparative purposes later on. More speculatively, I would not be surprised if there is also a bias toward finding evidence of more highly educated people being leaders of public opinion, because I would bet that questions with low overall disagreement are more likely to make it onto surveys if there is more disagreement among elites than there is among the masses.

Author: jeremy

I am the Ethel and John Lindgren Professor of Sociology and a Faculty Fellow in the Institute for Policy Research at Northwestern University.

8 thoughts on “problems of polarization”

  1. Jeremy, I agree that trying to measure “polarization” as an abstract global concept over a long historical period could run into a number of selection problems. However, there are a number of good studies on the polarization over specific issues.

    When OJ Simpson was first accused of killing his wife, opinion polls about the case had relatively little polarization along racial lines. By the time of the trial, opinions were highly polarized. In his book on the case, Darnell Hunt argues the racial projects by people seeking news coverage, and the reporters themselves, could explain the increasing polarization in public opinion.

    There are also a number of political scientists doing good work on shifts in attitudes over recent wars, and whether the increasing polarization in media consumption (i.e. people shifting to Fox News or MSNBC instead of less partisan outlets) could affect the level of polarization in public opinion. It’s a part of Baum and Groeling’s book “War Stories,” which looks at both news coverage and public opinion. One of the key questions of the book is if/when both the news and public opinion will converge on being close to the events happening on the ground, particularly if the war is going differently than the president claims.

    My sense is that polarization will have an even stronger correlation with media consumption than education, but the causality is murky at best. Do people with polarized opinions prefer partisan media? Does partisan media help polarize people? The answer to both questions appears to be yes, so we can’t clearly say A causes B.


  2. I’m probably revealing my ignorance here, but what is polarization? Is it the increasing intensity of opinions on an item (fewer “somewhat”s, more “strongly”s)? Or is it the clearer clustering of opinions on several items (more people on each side taking the “party line”)?


  3. Jeremy, another approach is to study behavior of political elites and see how often they cooperate. This is Poole and Rosenthal’s approach. They pool voting data from Congress, and IIRC, there is increased clustering. That approach doesn’t suffer from omitted quest biases. It’s just a measure of how often people cooperate over time.


  4. I always liked the Dimaggio et al. solution: using variance and kurtosis across multiple survey items as measures of polarization because they are not necessarily sensitive to question wording.


  5. The flip side is that items that completely harmonize over time are the obvious candidates for deletion when surveys get too long or budgets get reduced. The GSS item FEWORK, for example, was dropped after 1998, presumably because by the late 1990s very few respondents expressed the opinion that women should not be allowed to work outside the home. This form of selection will introduce a bias *against* observing convergence, because completely converging items are also excluded from the time series.

    I’d guess that item selection bias, the problem Jeremy notes, is stronger than item removal bias in some topical domains (e.g., political attitudes, cultural or lifestyle practices) but weaker in others (e.g., gender or racial attitudes).


    1. I was thinking the same thing. I started writing a simulation and although I didn’t finish I realized that the obvious decision rules would probably lead to artifactual consensus. that is, if the t0 instrument chose the most controversial issues at t0 and subsequent waves replaced these as the old items approached consensus. (for instance, would anyone put a question on inter-racial marriage on a survey today if they were starting from scratch?). this would lead to artifactual bias towards consensus, not polarization, to the extent that one limits an analysis only to long-running items.


Leave a Reply

Please log in using one of these methods to post your comment: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: