Racial Resentment and Whites’ Gun Policy Preferences in Contemporary America

Kleck (1996) does not explain how his measure of racial attitudes was constructed, so it is difficult to evaluate his results.

We conducted analyses of the 2004–2012 ANES data using a difference in stereotypes measure and also a difference in affect (thermometers) measure. Neither measure was statistically significant in any of the models. This further suggests that racial attitudes linked to old fashioned racism are not driving whites’ opposition to gun control. Results are available upon request.

We excluded from the analysis the 26 individuals who completed the survey in less than 6 min or more than 90 min. The rationale for this exclusion was that ‘speeders” were not paying enough attention to the questions while those who took an excessive amount of time to complete the survey most likely did it in parts and these interruptions could have an effect on the data. Including these 26 individuals does not change the results.

In addition to gun policy questions, respondents were asked about other policy preferences, attitudes toward government, fear of crime, racial attitude questions, and a long demographics battery.

For example, the mean income of the subject pool was lower than that of the U.S. white population ($45,000 vs. $55,000, respectively).

We opted for using the IAT pictures as a racial prime, rather than administering the full IAT because of time limitations.

We were concerned that the gender of the images’ subjects might affect the treatment and its influence on subjects’ gun policy preferences. Consequently, half the treatment group received images of females only and the other half were exposed to images of males only. Interestingly, our analyses found no systematic differences between the effect of the two treatments. Consequently, for purposes of this paper, we treat exposure to either set of images as a single treatment.

Of the six policy proposals that did not fall on the first dimension, five fell on a second dimension (which had an eigenvalue of 1.15) and the sixth fell on a third dimension (which had an eigenvalue of 0.53). This third dimension item asked about immigrants. The five items that fell on the second dimension asked questions about safety rather than gun regulation per se. Examples of policy proposals that fell on this second dimension included: “More teachers and school officials with guns in schools,” “Teach school children how to use and operate guns safely,” and “Place armed guards in all public and government buildings and events (e.g., schools).”

Since the responses to the 15 policy and belief items were coded on a four point scale, we analyzed the effects of the treatment using ordered logit; but none of the inferences or substantive results based upon the ordered logit analyses differed from the inferences or substantive results based upon the OLS analyses. In order to simplify interpretation, we are reporting the OLS results. Ordered logit results are available upon request.

Our measure of symbolic racism is the standard used in the discipline (e.g., Sears and Henry 2003). See Appendix B for the items.

Based upon the ANES data, support for gun regulation among whites has declined during the 2000-2012 period. This is consistent with data from other surveys.

Though not reported, we performed an identical analysis of the 2000 ANES (sans the gun in household item). Inferences and substantive effects from this analysis were not meaningfully different from the results reported in Table 2. Results of an ordered logit analysis of the 2000 ANES are available upon request.

Consistent with standard practice in the field, included tables indicate statistical significance based upon two-tailed hypothesis tests. However, we think a one-tailed test is more appropriate given that our theory provides an anticipated direction of effect. For the analysis of the 2004 ANES dataset, symbolic racism is statistically significant at the p < 0.05 level based upon a one-tailed test.

Technically, we could provide such a change in probability table for each response category. However, providing such additional information seems to us overly burdensome for the reader. Moreover, since so few cases (4–6 %) fall into the “make it easier” response category, the probabilities of being in the bottom category are very small for the overwhelming percentage of cases, so changes in the probability of being in this category due to a maximum change in any one predictor is quite small. In other words, most of the “action” is in changes in the probability associated with being in the “keep these rules about the same” and the “make it more difficult … to buy a gun” categories. Also, since the changes in probability associated with being in the top category tend to be greater than the changes in probabilities associated with a respondent selecting either of the other two response categories, we think that providing the change in probability associated with being in the top category provides an analysis most akin to the standard understanding of maximum effects.

Based upon a Wald test, we find that the analyses of the 2004 and 2008 ANES datasets are consistent with the parallel regression assumption, but that this assumption is not met by the ordered logit analysis of the 2012 ANES dataset. However, the relationship between symbolic racism and gun regulation preferences as reflected in the 2012 ANES dataset appears robust to model choice. For example, a multinomial probit analysis of the 2012 ANES dataset using the identical variables yields two coefficients associated with symbolic racism, both of which are in the anticipated direction and both of which are statistically significant at the p < 0.05 level (two-tailed). We discuss the results of multinomial probit analyses rather than multinomial logit analyses because the former is not dependent on the independence of irrelevant alternatives (IIA) assumption. Multinomial results are available upon request.

Analysts have noted a few disadvantages to including an LDV on the RHS. For example, in the context of time-series analysis, Achen (2000) notes that including a LDV on the RHS often biases estimates downwards. However, we are primarily interested in checking the robustness of our inference regarding the relationship between racial resentment and gun policy preference. Also, we find that the includsion of the LDV on the RHS does not reduce the estimated coefficient associated with symbolic racism by more than 17 % (see Table 3).

Issues regarding the consistency of maximum likelihood estimates of logistic regression using panel data have long been recognized (Hsiao 1986). Even when possible, consistent estimates of fixed effects logistic models involve the loss of significant information since only those cases in which the dependent variable changes state contribute to the log-likelihood function (and note that a fixed effects model is not possible given that T = 1 for the predictors). Because of these concerns, development economists often utilize the LPM as the least worst option (for a discussion, see: Beck 2011). Given these issues, we have opted to do the same.

Though the difference in coefficients is not statistically significant, it is possible that racial attitudes had a stronger relationship with gun policy preferences in July 2013 than in November 2012. In the middle of January 2013, less than a month after the Newtown, CT massacre, Obama announced 23 executive actions his administration would take in regards to gun control and he asked Congress to pass several gun control measures. Congress debated a variety of measures and the Senate voted on a series of gun control measures in late April 2013. It is conceivable that the push by Obama’s administration and Congressional Democrats on gun control during this period might have made racial attitudes more salient on this issue among the public.

Indeed, a simple bivariate AR(1) model accounts for approximately 38 % of the variance of the dependent variable.

Technically speaking, the coefficient associated with the lagged dependent variable can be interpreted as would any other coefficient using an OLS specification. A different approach would be to think of the model as a panel analogue to a “dead start” time series model. In this model, the coefficient associated with a lagged dependent variable can be thought of as an estimate of a decay rate that can be used to help parse the short-term and long-term effects of a change in the predictor on the dependent variable. In this approach, the −0.260 can be thought of as the short term change in support for “making it more difficult to purchase a gun” due to a one unit increase in symbolic racism. But the long term effect of such a one unit increase in symbolic racism would be −0.576 (which is simply \((\beta_{0} + \beta_{ 1} )/\left( { 1{-}\delta } \right) = (0 \pm 0. 2 60)/\left( { 1{-}0. 5 4 9} \right)\)).

Maximum effect of \({\text{X}}_{k} = {\text{b}}_{k} \times ({ \hbox{max} }.{\text{ value of X}}_{k } - { \hbox{min} }.{\text{ value of X}}_{k} )\).

Factor analysis are available upon request.