The contours of constitutional approval.

• Author: Stephanopoulos, Nicholas O.
• Position: III. Descriptive Exploration F. Non-Substantive Features through Conclusion, with footnotes and appendices, p. 152-190

6. Non-Substantive Features

   Our other constitutional variables relate not to the charters' actual content but rather to their key non-substantive features: their age (in years), length (in words), and amendment frequency (in number). As these are all continuous state-level variables, we cannot display them using the same sorts of charts we have used until now. Instead, Figure 9 plots respondents' state constitutional approval scores against these non-substantive features, with best fit lines included as well. These scatter plots reveal no discernible pattern in the data, and thus suggest that there is essentially no relationship between approval and constitutional age, length, or amendment frequency. In fact, the correlations between approval and these variables are just 0.004, -0.002, and 0.002, respectively. Again, this analysis does not hold constant any other drivers of approval, but it certainly lends no support to the notion that people's constitutional attitudes are shaped by their charters' non-substantive dimensions.

   4. EXPLANATORY ANALYSIS

   To this point, our exploration of the data has been entirely descriptive. We have shown how federal and state constitutional approval scores vary along several notable dimensions, including geography, demography, and self-reported knowledge and attitudes. While this sort of analysis helps to detect patterns in the data, it does not take into account various confounding factors. For example, people who are more constitutionally knowledgeable might also follow the news more closely, making it unclear which factor is more associated with constitutional approval once we control for the other. Similarly, race, income, and education might be interrelated in ways that make it impossible to draw reliable conclusions from statistics for a single attribute in isolation.

   To assess the causes of constitutional approval more rigorously, we therefore turn to regression analysis. Unlike descriptive exploration, regression analysis enables us to determine the impact of different variables while holding other variables constant. To illustrate, we can evaluate how constitutional knowledge affects constitutional approval while controlling for the correlation between knowledge and attentiveness to the news. Likewise, we can discern the link between race and approval notwithstanding the many ties between race and other demographic characteristics.

   However, it is important to acknowledge the limitations of this method. The basic advantage of regression analysis is that it allows us to hold other variables constant--but not all other variables can be controlled for. In particular, it is possible that there exist factors, either personal or constitutional, that are correlated with both our independent variables and constitutional approval, but that are omitted from our models. These factors could be the actual drivers of constitutional backing, but we would not be able to observe their impact since they are excluded from our calculations. The possible existence of these omitted variables limits our ability to make causal claims. Nevertheless, regression analysis does at least shed light on the plausibility of different hypotheses, and so launches the systematic study of support for constitutions.

   We begin below with base models of federal and state constitutional approval that include only demographic attributes. We then build up these models in stages, adding in turn civic knowledge, institutional attitudes, and in the state model, constitutional congruence and non-substantive constitutional features. In all regressions, the dependent variables are respondents' stated support for the federal Constitution or their state constitution. Even though these variables are on a ten-point scale, and so ordinal in nature, we use simple linear (OLS) regression models because their results are easier to interpret. (114) We also confirm the robustness of our findings by using ordered probit models that are suited to ordinal data. Furthermore, since there likely exist significant differences between states (after all, each state has its own constitution), we include separate dummy variables for all states. This technique, known as fixed effects estimation, controls for all interstate variations due to politics, economics, demography, or culture. (115) Lastly, since answers from respondents in the same state might be correlated with one another, we use robust standard errors clustered at the state level, thus allowing for serial correlation between same-state respondents.

1. Demographic Attributes

   As just noted, our base models of federal and state constitutional approval include only the demographic attributes asked about by our survey. These are: (1) a binary variable for gender (0 if female, 1 if male); (2) a continuous variable for age (in years); (3) two binary variables for race, one indicating whether the respondent is African American, the other whether she belongs to another minority group (Asian American, Pacific Islander, or Native American); (4) an ordinal variable for education, ranging from less than high school to a doctorate or its equivalent; and (5) an ordinal variable for income, ranging from below $30,000 to above $500,000. (116)

   The two panels in Figure 10 graphically depict the results of these models (federal on the left, state on the right). For each variable, the point represents the best estimate of its coefficient's value--that is, the impact of a one-unit shift in the variable on constitutional approval, holding the other variables constant. The lines to either side of each point denote the 95% confidence interval for the coefficient's value. We can say with 95% certainty that the coefficient's true value falls within this range. And the stars above each point illustrate how confident we are that the coefficient's true value is different from zero. Three stars (***) indicate confidence at the 99% level, two stars (**) confidence at the 95% level, and one star (*) confidence at the 90% level.

   In the federal model, being male, older, better educated, and wealthier all are associated with increased constitutional approval. The findings for gender, age, and education are especially clear, rising to statistical significance at the 99% level. Men support the federal Constitution by roughly 0.35 points more than women. A nineteen-year increase in age (the span famously identified by Thomas Jefferson as the duration of a constitutional generation (117)) results in about a 0.5-point rise in backing. And a one-level increase in education produces an approval boost of 0.1 points or so. On the other hand, being African American is linked to a substantial decrease in constitutional support. Blacks back the federal Constitution by roughly 0.25 points less than whites. And membership in other racial minority groups is statistically unrelated to approval. (118)

   The results at the state level are extremely similar. Men, older people, and wealthier people again support their constitutions more strongly (though the coefficients for gender and age are not quite as large as at the federal level). And African Americans again are less constitutionally satisfied (by an even larger margin than before). The only notable difference between the analyses is that education does not have a significant impact on backing in the state model. Demographics, then, play an almost identical role in explaining federal and state constitutional approval. People's attitudes toward both of their charters are shaped in equivalent fashion by their key personal characteristics. (119)

2. Civic Knowledge

   Of course, the base models include only respondents' demographic attributes, but we also want to test several additional hypotheses about constitutional approval. So we now begin to add in stages more variables to the models, beginning here with the ones related to civic knowledge. Specifically, we now add (6) respondents' self-reported familiarity with the federal Constitution and their state constitution, on a five-point scale; and (7) how closely respondents follow the national and local news, ranging from "not closely at all" to "very closely." We also note that, to save space and avoid confusion, we only discuss the results for the newly added variables in each of the intermediate models we construct. While the insertion of these variables affects the coefficients for the variables already present in the models, we save our discussion of all of the potential causes of constitutional approval until we arrive at the final federal and state models. Again, the final federal model includes only respondent-specific factors, while the final state model includes constitutional features too.

   At the federal level, as Figure 11 indicates, self-reported familiarity with the federal Constitution and attentiveness to the national news are strongly associated with constitutional approval. Both variables are statistically significant at the 99% level, and their coefficients are substantively large as well. A one-point increase in constitutional familiarity results in almost a one-point rise in constitutional support. Similarly, a one-level increase in attentiveness to the national news (e.g., from "not so closely" to "somewhat closely") produces about a 0.25-point bump in backing. These results, which control for all of the demographic attributes already included in the models, are quite substantial. Civic knowledge seems clearly related to federal constitutional approval. (120)

   As Figure 11 also reveals, the same is true at the state level. Self-reported familiarity with the state constitution and attentiveness to the local news both are significantly linked, at the 99% level, to state constitutional support. A one-point increase in constitutional familiarity again results in almost a one-point rise in constitutional backing. Likewise, a one-level increase in local news attentiveness again produces about a 0.3-point...