A reader of our Nature Climate Change study asks:
I was wondering if the anti-correlation of scientific literacy with climate change understanding is muted or reversed as one moves into the middle of the Hierarchy-Egalitarian/Individualism-Communitarianism Axes? Did you consider dividing the group into quartiles for example rather than in halves?
To start, as you know, the negative correlation (itself very small) between science literacy (or science comprehension, as one might refer to the composite science literacy & numeracy scale) & climate change risk perception doesn't take account of the interaction of science comprehension with cultural worldviews. Once the interaction is measured, it becomes clear that the effect of increased science comprehension isn't uniformly negative; it's *positive* as individuals become more egalitarian & communitarian, & negative only as individuals become more hierarchical & individualist
For this reason, I'd say that it is misleading to talk of any sort of "main effect" of science literacy one way or the other. By analogy, imagine a drug was found to decrease the lifespan of men by 9 yrs & increase that of women by 3 yrs. If someone came along & said, "the main effect of this drug is to *decrease* the average person's lifespan by 3 yrs; what an awful terrible drug, it should be banned!" I think we would be inclined to say, "no, the drug is good for women, bad for men; it's silly to talk about its effect on the 'average' person because people are either men or women." Similarly here: people vary in their worldivews, & the effect of science comprehension on their climate change views depends on the direction in which their worldviews tend.
But that's not really important.
I understand your question to be motivated by the idea that the interaction between science comprehension & culture might itself be concentrated among people who have particularly strong worldviews. Perhaps the effect is uniformly positive for everyone except some small set of extremists (extreme hierarchical individualists, it would have to be). In other words, maybe only hard core partisans are using -- abusing, really -- their science comprehension to fit the evidence to their predispositions. That seems plausible to me, and definitely worth considering.
You are right that there is nothing in the analyses we reported that gets at this "partisan abuse" hypothesis. As you likely saw, the cultural worldview variables are continuous, and in our Figures we plotted regression estimates that reflected the influence of the culture/science comprehension interaction across the entire data set. That way of proceeding imposes on the data a model that *assumes* the interaction of science comprehensionis uniform across both worldview variables -- "hierarchy-egalitarianism" & "individualism-communitarianism." We'd necessarily miss an evidence of the "partisan abuse" hypothesis w/ that model.
But we also did try to fit a polynomial regression model to the data. The idea behind that was to see if in fact the interaction between science comprehension & cultural worldviews seemed to vary w/ intensity of the cultural worldviews-- as the partisan abuse hypothesis implies. The polynomial regression didn't fit the data any better than the linear model, so we had no evidence, in that sense, that the interaction we observed was not uniform across the cultural dimensions.
One could also try to probe the "partisan abuse" hypothesis by slicing the sample up into segments, as you suggest, and seeing if the effect of science comprehension on groups of people who are more or less extreme. But because such effects will always be lumpy in real data, there is a risk that any differences one observes among different segments along the continuum when one splits a continuous measure up into bits will be spurious. See Maxwell, S. E., & Delaney, H. D. (1993). Bivariate Median Splits and Spurious Statistical Significance. Psychological Bulletin, 113, 181-190 (this was one of statistical errors in the scandalously idiotic "beautiful people have more daughters" paper).
Accordingly, it is better to treat continuous measures as continuous in the statistical tests -- and to include in the tests the right sorts of variables for genuine nonlinear effects, if one suspects the effects might vary across the relevant continuum. That's what we did when we tried a polynomial regression model out.
Still, let's slice things up anyway. Really, let's just *look* at the raw data -- something one always should do before trying to fit a model to them! -- to see if we can see anything that looks as interesting as the "partisan abuse" dynamic is going on.
I've attached a Figure that enables that. It fits a smoothed "lowess" regression lines to the risk perception/worldview relationship after splitting the sample at the median into "high" & "low" science comprehension groups. The lines, in effect, show what happens when one regresses risk perception on the worldview "locally" -- to little segments of the sample along the cultural worldview continuum -- for both types (high & low science comprehension) of subjects.
What we're looking for is a pattern that suggests the interaction of science comprehension w/ culture isn't really linear; that in fact, science literacy predicts more concern for everyone until you get to some partisan tipping point for subjects who are culturally predisposed to be skeptical by their intense hierarchy or individualism. I plotted a dashed line that reflects that for comparison.
I don't see it; do you? Both lines slope downward (cultural effect), the green one at a steeper grade (interaction), in roughly a linear way. The difference from perfectly linear is just the lumpy or noisy distribution of data you might expect if the "best" model was linear.
Am open to alternative interpretations or tests!
Oh, since we are on the subject of looking at raw data to be sure one isn't testing a model that one can see really isn't right, here's another picture of the raw data from our study. It's a scatterplot of "hierarchical individualists" and "egalitarian communitarians" (those subjects who score either in the top 50% of both worldview scales or the bottom 50% on both, respectively) that relates their (unstandardized) science-comprehension score to their perception of climate change risk (on the 0-10 industrial strength measure).
I've superimposed a linear regression line for each. Just eyballing it, seems like the interaction of science comprehension & climate change risk perception is indeed more-or-less linear & is the about the same in its slope for both.