This paper makes a very important contribution to our knowledge on intelligence and political opinions. Although it has been speculated for a while that the relationship may be non-linear or u-shaped, this hypothesis has not been tested thoroughly using approaches that are reliable in the extremes of the IQ distribution. The manuscript uses a highly reliable IQ test, with a large sample and a highly apt methodology (splines as opposed to higher order terms) to test for and identify non-linearities. The paper now allows us to say quite conclusively that the relationship is not perfectly linear, albeit that the effect appears to level off at higher IQs, rather than change direction. I intend to support publication after the author has considered several small issues.
1) Why analyse only one latent factor of political opinions? As the author has previously declared on vieweing political opinions as one dimensional: "However, this model is actually and obviously false in a nontrivial way... Political ideology cannot generally be reduced to 1 dimension without massive loss of information and probably substantial chance of misinterpretation." Yes, it is a useful simplification and a general politics factor should be the main result, but I see little reason not take a two factor solution and analyse the non-linearity on the second factor. My prior guess is that non-linearity will be stronger on the second factor (at least if it was reliably measured). In Figure 2 it appears that it might be the economic questions which are most u-shaped e.g. large companies should be controlled by the state and not private actors. It would be a shame not to run this as a supplementary test. It would be worth writing another paper on the topic if the author were to skip it.
The only argument I can think of against testing the relationship with two latent factors is that the author thinks the reliability will be too low to say anything. I'm generally skeptical of using hard and fast rules regarding reliabilities being too low to publish (because publication bias is annoying and as long as the sample szie is large enough you might still have power to detect an effect). But if this is the author's reasoning it might be worth reporting the reliability of a second latent politics factor.
2) Related to point 1, it might be worth adding the reliability of the politics factor. Part of the author's justification of the paper is that his more reliably IQ testing allows non-linearities to be tested with good power. The reliability of the latent conservatism factor is also related to that claim.
3) The author seems to want to say that there is a positive effect among the very high IQ, but the results as they are presented seem to only clearly show that the effect is somewhat nonlinear.
I view the results as vindicating the scholars who thought the relationship was approximately linear, moreso than those who have trumpeted nonlinearities and certainly moreso than fans of the midwit meme. The change in adjusted R squared from including splines in table 1 is 0.012, or change in "adjusted r" of 0.11. Important, to be sure, but at times over-sold in the paper. For example, in the abstract the author says "becomes positive in the right tail". Looking at Figure 2, I'm not sure you can say the relationship isn't just flat after a certain IQ. Sure, the MARS algorithm gets nice large positive effects at high IQs, but they don't have confidence intervals with which I can judge how strong the reversal is. As another reviewer has pointed out, the author seems to overplay how congruent his results are with the midwit meme.
I see a few possible solutions: a) change the language of the conclusions to be slightly less strong ("the effect levels off", "possible that the effect of IQ changes direction" etc.) b) Literally subset to the high IQ in the sample, and recalculate the linear effect in that portion so you can then see if it is significantly different from 0. (or any other equivalently simple test that can tell us whether there really is a significant positive effect in the right tail) c) Putting confidence intervals on the MARS estimates. "Confidence intervals" not possible. With a bootstrap and computer left on overnight, anything is possible.
The rest of my comments are less substantive and are primarily concerned with writing.
- First sentence - "with political opinions with". Two "withs" so close is ugly.
- First sentence of introduction "long standing interest". The studies cited only go back to 2012. Research on the topics goes back to at least the 40s or 50s. I'd cite Onraet et al.'s meta-analysis here, or cite one of the classic studies to justify the long standing interest claim.
- First paragraph of intro. A lot of claims are made e.g. effect of IQ on party support in multi-party systems, changes over time in the relationship between IQ and politics. I think all the claims are correct, but citing them would make it easier to follow this. Some of this information is surely in the citations given in the first sentence. But it is a pain for me to trawl through to see which citation corresponds to which claim made later in the paragraph. I think the author might want to cite Ludeke et al. (2018) Different political systems suppress or facilitate the impact of intelligence on how you vote: A comparison of the US and Denmark for this claim regarding multi-party systems.
- "measures of free market support (economic liberalism) show positive relationships". I think Jedginer et al meta-analysis says r = 0.05. I'd say "tend to show positive relationships", "tend to show null or positive relationships".
-"This test has quite limited reliability". Personal taste, but I dislike adverbs or qualifying adjectives. Writing is clearer and stronger without them. ie. "This test has limited reliability."
-The first study was for something else. Maybe explain what this something else was or cite it
- Maybe cite your new intelligence test. And explain/cite how you used "an optimally chosen 50-item abbreviation" for the second sample.
- Footnote 2 says "we removed both the effect of age on the central tendency and on the dispersion." I'm dumb but how did you remove the effect of age on dispersion? Calculated the standard deviation within each birth year or within bins of five years? Could use simple words for simple readers. Many readers, I expect, won't know what central tendency means.
- Maybe give more detail on the ethnicity questions. ie. which groups got messed up in the second survey and which became "other".
-related to the above. It might be useful to post the survey on an OSF page
- Table S1 is a little unclear. I'm guess demo_rsq is Rsquare with just demographic variables and linear_rsq is after IQ is controlled linearly? Maybe the names could be clarified or a footnote could do this to.
- endspan.". However... There are two fullstops here. Best to remove the second one.
- "Two prior studies." Three, however are cited. One of which does not actually use polygenic scores, but only the within-family design.
-"which were most shorted and less reliabiliy than the current study". Could change to "which were mostly short and had less reliability..."
