Midwit Theory: State of the Research

It has been hypothesized online that people of low and high intelligence tend to agree on some issues, while those of average intelligence have a different opinion.

To my knowledge, the only person who has tested this quantitatively before me and Emil took over the field was Noah Carl, who found some limited support for midwit theory favoring left wing beliefs in the GSS. Within this dataset, he found that the most socially liberal decline was always the 10th decline of cognitive ability, while the most fiscally conservative decline was usually around the upper average.

from here

from here

Personally, I don’t put much stock into these results, because the cognitive tests used (WORDSUM + interviewer rating) were of low quality and had low ceilings.

Two measures of cognitive ability are available in the 2012 wave of the ANES. The first is a 10-item vocabulary test in which the respondent must identify which of five phrases supplies the correct definition of a given word. It was administered to respondents once, during the pre-election interviews. In general, vocabulary tests load more strongly onto the crystallised factor of general intelligence than onto the fluid factor (Cattell, 1963). They tend to have high heritabilities and high g-loadings, relative to other subtests (Jensen, 2001). For a longer discussion of the measure’s validity, see Caplan and Miller (2010). The second measure of cognitive ability is a rating by the interviewer of the respondent’s apparent intelligence. In particular, the interviewer assesses whether the respondent’s intelligence appears to be ‘‘very low’’, ‘‘fairly low’’, ‘‘average’’, ‘‘fairly high’’ or ‘‘very high’’. These categories were re-coded from ‘1’ to ‘5’, respectively. Because assessments were made during both the preelection and post-election interviews, I utilise each respondent’s average rating. Encouragingly, the Pearson correlation between the two ratings is strong, namely r = .69 (p < 0.001, n = 1906)

Noah Carl doesn’t seem to put much stock into the results either:

There are of course several important limitations to this study. First, the measure of cognitive ability utilised can at best provide no more than a rough gauge of general intelligence: it was constructed from just two variables, neither of which evaluated respondents’ abstract reasoning ability through analytical tests. Second, as noted above, none of the effect sizes observed is particularly large. Third, the results correspond to a particular country in a particular time period; they cannot be assumed to hold universally. As Woodley (2011) argues, the association between cognitive ability and political beliefs may vary depending on prevailing cultural values and social norms.

That aside, there was a study presented at ISIR which researched this topic, but I didn’t see it because I was too busy talking to a friend.

II.

Me and Emil Kirkegaard tested midwit theory in our study on belief in government responsibility and belief in individual responsibility, and found zilch.

We used a good test (ASVAB) to measure IQ, but our sample sizes weren’t great (n = 1400), so if there was a slight trend, we wouldn’t have been able to detect it. That said, the observed results were extremely consistent with linear associations.

I also tested this theory in the NLSY79 using measurements of belief in gender roles (4 waves) and political affiliation (2008 wave) using the same cognitive test (ASVAB). I also found zilch - for the most part. Past the 115 IQ ability level, IQ and belief in conservative gender roles no longer correlated.

Emil tested this theory in some of his datasets and also found zilch. I’m not willing to post his results without his consent, so I can’t prove it, but this information did affect my position on whether the theory had any air to it.

So, I gave up on midwit theory. I have aggressively countersignalled it on my blog and on twitter since.

And what happened? Well, I ate dirt. I tried testing it again in the WLS, and found strong evidence for midwit theory, where high and low ability individuals both trend towards identifying with the Democratic party or as left wing (composite of conservative self-id and republican self-id).

linear vs nonlinear fit is posted here. passes significance (p < .001). Ignore the group variable.

passes significance within women (p < .001), but barely within men (p = .02).

The difference between the nonlinear fit and linear fit passess significance within women, but only barely within men. It turns out that this effect is largely specific to conservative beliefs, and when the residuals of republican support independent of conservatism are measured, there is no midwit effect.

It is odd that there is no midwit effect for gender roles in either gender, but there is one for conservative identification. Even though the raw sum of the results suggest the theory is not true, it’s hard to ignore the WLS results as a fluke, though that could be attributed to using self-identification as a measurement instead of composite measurements using various questions.

I plan to double check the code, clean these results up into an actual study and publish it somewhere within the next year.

Appendix/notes

Questions used to measure belief in government responsibility:

Questions used in NLSY79 analysis:

WLS variable used to measure conservative identification: z_iz103rer, (R05) Where would you place yourself on a liberal and conservative political scale?

WLS variable used to measure party identification: z_iz103rer, (R05) Generally speaking, do you think of yourself as a Republican, Democrat, Independent, or other? (1-5 scale).

Right wing identification was measured by combining republican/conservative identification.

In the WLS, IQ was measured by taking the average score of the score they got on the Henmon-Nelson IQ test in Freshman and Junior year.

In the NLSY datasets, IQ was measured using the ASVAB.