Trumpism, climate and COVID: Social bases of the new science rejection

significant interactions. In each case, the gap between Republicans and Democrats widens when most of the respondent’s friends identify with the same party they do. Almost 90 percent of Republicans with mostly same-party friends approve of Trump, compared with fewer than 70 percent of Republicans with mixed or different-party friends (Fig 3A). Impacts on climate and vaccine rejection are smaller in absolute magnitude but have similar meaning. The Demo- crat–Republican gap in COVID vaccine or ACC rejection widens from about .01 (when they have mixed or different-party friends) to 0.17 or 0.20 (when they have mostly same-party friends). 5.3 Robustness of results How stable are the findings described here? This section reviews diagnostic and goodness-of- fit statistics and evaluates some alternative model specifications. By inflating standard errors (making estimates less precise), multicollinearity can raise problems in multivariate analyses. This is particularly a risk when interaction terms are involved. Testing for multicollinearity, we used ordinary least squares to approximate variance inflation factors (VIF) and tolerances (1/VIF) for independent variables in each equation of our GSEM analysis. Chatterjee and Hadi [92] write that VIF of 10.0 or higher indicate serious multicollinearity; other sources suggest that a VIF of 4.0 could be cause for concern. For the NoACC and NoVax equations in Table 2, VIF ranged from 1.04 to 2.85, with a mean of 1.61, well below thresholds for concern. VIF are even lower for the Trumpism equation: 1.02 to 2.44,

with a mean of 1.51. These results indicate that despite the interaction terms, multicollinearity is not problematic for our analysis. Goodness of fit is an ambiguous concept with logit and ordered-logit analyses, which lack a simple counterpart to linear regression’s explained-variance measure, R2. A variety of pseudo R2 statistics have been proposed, but these can give divergent results and should be interpreted alongside other goodness-of-fit statistics [93]. McFadden’s pseudo R2, the most widely used, equals 0.20 for the NoACC equation in Table 2. Correct classification rate for this NoACC equation is 77% with 87% specificity and 54% sensitivity. A Hosmer-Lemeshow goodness-of- fit χ2 test does not reject the fitted model, nor does a specification link test. For the NoVax equation in Table 2, McFadden’s pseudo R2 equals 0.19; correct classifica- tion rate is 81% with 96% specificity and 23% sensitivity. As with NoACC, a Hosmer-Leme- show goodness-of-fit χ2 test does not reject the fitted model, nor does a specification link test. Unlike the binomial NoACC and NoVax equations, the equation for ordinal Trump support employs ordered logit regression, for which correct classification rates and Hosmer-Lemeshow tests are less appropriate. Pseudo R2 for this equation equals 0.21. Specification link tests sug- gest that we could obtain a better fit by using a dichotomous version of this variable (support Trump or not) instead of the 7-category ordinal measure employed. But a dichotomous indica-

tor for Trumpism is substantively less appealing. It also works less well analytically: using dichotomous instead of ordinal Trump support worsens the fit of both NoACC and NoVax equations, as indicated by comparisons of their respective Bayesian information criteria (BIC). That is, ordinal Trump support predicts NoACC and NoVax better than dichotomous Trump support does. These results support keeping the ordinal version. Two exogenous variables in our analysis, education (4 categories) and party (3 categories) were modeled as ordinal predictors having approximately linear effects. Such treatment fol- lows many previous studies (for graphed examples involving climate change and vaccines, see Figs 2 and 3 in [53]) and yields parsimonious, interpretable results consistent with hypotheses. An alternative approach preferred by some authors would be to enter education and party as unordered categorical variables. Although using categorical education and party requires PLOS ONE Trumpism, climate and COVID PLOS ONE | https://doi.org/10.1371/journal.pone.0293059 January 10, 2024 11 / 20