Economic inequality and belief in meritocracy in the United States

The crucial relationship between economic inequality and system-supporting beliefs like meritocracy is the subject of two diametrically opposed theories: the conflict theory and the relative power theory. We briefly review these two theories in this section.

As noted above, NJL advocate the conflict theory. This theory maintains that, for lower-income individuals, being confronted with higher levels of local inequality “increases the salience of their disadvantaged position within a conspicuous local economic hierarchy” (NJL, p.327), promotes class consciousness, and, in turn, increases demand for redistribution. Higher-income individuals in high inequality contexts, on the other hand, are expected to avoid guilt while simultaneously and self-interestedly protecting their privilege by becoming even more likely to believe in the importance of individual effort to the distribution of economic rewards. The conflict theory has received sustained attention, particularly in studies of political participation, but empirical support has been at best mixed (on the US case, see Brady, 2004; Oliver, 2001; Solt, 2010).

The relative power theory, on the other hand, starts with the proposition that money is a political resource; that is, that it can be used to influence others. Therefore, the theory contends, where the rich are richer relative to the poor, they will also be more powerful relative to the poor (Goodin and Dryzek, 1980). With regard to attitudes and beliefs like meritocracy, this theory suggests that the greater power imbalance that results from higher levels of economic inequality provides higher-income people with more resources to spread their views in the public sphere while depriving poorer people to a greater degree of the resources needed to resist these efforts. This gives poorer people “a greater susceptibility to the internalization of the values, beliefs, or rules of the game of the powerful as a further adaptive response—i.e., as a means of escaping the subjective sense of powerlessness, if not its objective condition” (Gaventa, 1980: 17). Patterns of religiosity (Solt, 2014; Solt et al., 2011), respect for authority (Solt, 2012), and policy mood (Kelly and Enns, 2010) have provided support for this theory.

As Huber and Stephens (2012: 37) summarize the two theories, the relative power theory can be seen as a straightforward implication of “the usual assumption in sociology, political science, and anthropology . . . that social structures reproduce themselves,” while the conflict theory is grounded in the seemingly implausible premise that social structures are self-negating. Regardless of the theories’ plausibility a priori, which of the two is actually supported by the evidence is crucial to understanding the political consequences of economic inequality.

Using the data files and commands provided (Newman et al., 2015b), we were able to reproduce a close approximation of the article’s results on belief in meritocracy, as presented in the article’s Table 1, Model 1 (“White Rs”).² As these files note that the authors are themselves unable to reproduce the published estimates exactly, and the differences are indeed quite small, we proceed to interpretation.

NJL (p.334) claim that its analysis

This claim is incorrect.

The error lies in the interpretation of the multiplicative interaction term. Though it has been well known for over a decade that models containing multiplicative interaction terms require particular care in interpretation (see, e.g., Brambor et al., 2006; Braumoeller, 2004; Golder, 2003), many political scientists continue to struggle with them; improperly specified or interpreted interaction terms appear at the top of Nyhan’s (2015) list of “recurring statistical errors” for which reviewers should be sure to check. In the multilevel logistic regression model employed in NJL, the logged odds of rejecting meritocracy for individual $i$ in local context $j$ are estimated as follows

NJL (p.334) offers two pieces of evidence as support for the above-quoted claim: first, that the coefficient for local income inequality (i.e. ${\gamma }_{01}$) is estimated to be positive and statistically significant, and second, that the coefficient for the interaction between inequality and respondents’ incomes (${\gamma }_{11}$) is negative and statistically significant (NJL, p.334).

Neither of the two actually provide any support. First, the coefficient ${\gamma }_{01}$ indicates only the effect of $Inequality$ when the other variable in the interaction, $Income$, takes on the value of zero (see, e.g., Brambor et al., 2006: 72). However, inspection of the NJL data reveals that the nine categories of income in the Pew surveys the article employs were recoded to take on nine evenly spaced values ranging from 0.21 to 1 (see Newman et al., 2015b).³ Because this income variable never actually equals zero in the analyzed sample, ${\gamma }_{01}$ is not directly interpretable. Second, Brambor et al. (2006: 74) specifically advise that one “cannot even infer whether $X$ has a meaningful conditional effect on $Y$ from the magnitude and significance of the coefficient on the interaction term.” Instead, the conditional effect of inequality is found by taking the partial derivative of equation (1) with respect to inequality

In short, ${\gamma }_{11}$ is only part of the conditional effect; the magnitude and statistical significance depend also on ${\gamma }_{01}$ and the value of $Income$. To properly interpret the conditional effect of $Inequality$, then, one cannot examine ${\gamma }_{01}$ or ${\gamma }_{11}$ in isolation, as NJL does; instead the conditional effect must be calculated using equation (2) for all observed values of $Income$ (see, e.g., Berry et al., 2012; Brambor et al., 2006; Braumoeller, 2004).

We plot the conditional effect of $Inequality$ at each observed value of $Income$ in Figure 1 using the package interplot (Solt and Hu, 2015b). Contrary to the interpretation offered in NJL (p.334), this plot reveals that the coefficient for county income inequality fails to reach statistical significance at any observed level of respondent income. The coefficient estimate for those with the lowest incomes, under US$10,000 per year, approaches but does not reach statistical significance at conventional levels. In any event, only 4% of the sample employed in NJL had incomes this low; even if this coefficient did reach statistical significance, it would provide little support for the conflict theory (see Berry et al., 2012: 660–661). At the other end of the income scale, there is no hint of support for the NJL claim that these results support the conflict theory’s prediction that higher-income individuals will be less likely to reject meritocracy in contexts of greater income inequality. The article’s own results provide little, if any, evidence for its conclusion that poorer people are more likely to reject and richer people more likely to embrace meritocracy where local income inequality is greater.

Download

Open in new tab

Download in PowerPoint

Figure 1. Logit coefficients of local income inequality on rejection of meritocracy by income (Newman et al., 2015a: Table 1, Model 1, from replication data).

Note: the dots represent coefficients of income inequality within respondents’ county on their rejection of meritocracy for all values of respondent income, estimated from the data and model provided in Newman et al. (2015b); the whiskers represent the 95% confidence intervals of these estimates. Contrary to the interpretation offered in Newman et al. (2015a: 334), the coefficient for county income inequality fails to reach statistical significance at any observed level of respondent income. The histogram presented below the main plot depicts the relative frequency of each observed value of respondent income: only 4% of respondents in this sample, for example, reported incomes below US$10,000.

Beyond the relatively common problem of misinterpreting its interaction term, NJL suffers a more fundamental problem with measurement. Ostensibly to amass observations from a sufficient range of local contexts, NJL (pp.330–331) combines in a single analysis data from four surveys using three different measures of its dependent variable, rejection of meritocracy. Measure 1 was drawn from 2005 and 2006 surveys that asked respondents which of two statements came closest to their own opinion: “Most people who want to get ahead can make it if they’re willing to work hard” or “Hard work and determination are no guarantee of success for most people.” Those who chose the latter were coded as rejecting meritocracy. The 2007 and 2009 surveys employed did not include this item. Instead, they asked respondents to assess on a four-point agree–disagree scale two separate statements: (1) “Hard work offers little guarantee of success” and (2) “Success in life is pretty much determined by forces outside our control.” In Measure 2, used with data from the 2007 survey, those who mostly or completely agreed with both statements were coded as rejecting meritocracy. Although the 2009 survey included these same two statements, and so made Measure 2 easily calculable, respondents to that survey were coded in yet a third manner: in Measure 3, those who mostly or completely agreed with statement (1) were coded as rejecting meritocracy regardless of how they responded to statement (2).⁴

To assess whether these three very different measures are in fact comparable, we collected Pew surveys conducted between 1999 and 2012 that asked any of the items just described and plotted the estimated percentage of the population to reject meritocracy according to each of the three versions of the dependent variable in Figure 2. The figure reveals that Measure 2 results in much lower levels of rejection of meritocracy than either of the others, and that Measure 3 often yields considerably higher levels than Measure 1. Their evident lack of comparability suggests that pooling them in a single analysis is difficult to justify.

Download

Open in new tab

Download in PowerPoint

Figure 2. Comparing the three measures of rejection of meritocracy pooled by Newman et al. (2015a).

Note: the analyses presented in Table 1 of Newman et al. (2015a: 333) were conducted on pooled observations with the dependent variable, rejection of meritocracy, measured in one of three different ways (see Newman et al., 2015a: 331). Here, solid circles represent the data used by Newman et al. (2015a); hollow circles represent data in other available Pew surveys. The whiskers are 95% confidence intervals for each estimate. Plotting the weighted percentage of respondents to reject meritocracy by each of these measures reveals that the second measure results in much lower levels of rejection of meritocracy than either of the others, and the third often yields considerably higher levels than the first. In light of the evident lack of comparability of these three measures, pooling them into a single analysis cannot be justified.

In a footnote and the article’s online appendix, NJL nevertheless argue that mixing these different measures is not problematic on the basis that the 2005 and 2006 surveys alone, which use Measure 1, yield results for the variables of interest similar to those produced by the four surveys together (NJL, p.331). In analyses we report in online Appendix A, we demonstrate that the results obtained for the NJL model differ considerably across the three measures of the dependent variable. Regardless of how the dependent variable is measured, however, the conditional effect of inequality fails to reach—or even approach—statistical significance at any observed level of income. The results presented in NJL are an artifact generated by the decision to pool these three very different measures together.

What then can be discerned regarding the relationship between economic inequality and belief in meritocracy? We employ the US Religious Landscape Survey (RLS) conducted by the Pew Forum on Religion and Public Life in 2007 to better investigate the question. With more than 10 times the number of respondents of the much smaller Pew surveys examined in NJL, the RLS was designed to provide a particularly fine-grained picture of geographic variation in attitudes and beliefs across the continental US. The RLS is, therefore, perfectly suited to providing observations across a broad range of local contexts, and it includes Measure 1 of the dependent variable. To make use of all of the available data, we analyze the entire sample of survey respondents, rather than only non-Latino white respondents; as NJL (p.330) notes, this should be expected to bias the results toward the expectations of the conflict theory. We use the package mi to multiply impute missing data (Su et al., 2011).

Otherwise, we adopt the approach employed in NJL. As in that article, we use the Gini coefficient of household income inequality for each county calculated by the US Census Bureau from the 2005–2009 American Community Survey (five-year estimates) to measure income inequality at the local level.⁵ Like in the Pew surveys analyzed in NJL, respondents’ incomes are measured in the RLS on a nine-point scale ranging from less than US$10,000 to over US$150,000, which we straightforwardly coded with values 1 to 9.

At the contextual level, we follow NJL in controlling for average income, the black share of the population, the percentage of votes won by George W Bush in 2004, and the total population size.⁶ At the individual level, the analyses include controls for age, education, sex, race, citizenship, partisan identification, political ideology, and church attendance. As in NJL, the model is estimated using multilevel logistic regression of individuals nested in counties, with both the intercept and the coefficient for income allowed to vary across the counties.

Figure 3 displays a dot-and-whisker plot of the results: the dots represent the estimated change in the logged odds of rejecting meritocracy for a change of two standard deviations in each variable in the model, and the whiskers represent the 95% confidence intervals of these estimates (see Kastellec and Leoni, 2007; Solt and Hu, 2015a).

Download

Open in new tab

Download in PowerPoint

Figure 3. Predicting rejection of meritocracy.

Note: the dots represent the estimated change in the logged odds of rejecting meritocracy for a change of two standard deviations in the independent variable; the whiskers represent the 95% confidence intervals of these estimates. Multilevel logistic regression analyses of 35,556 individual respondents living in 2740 counties.

The coefficient of income inequality is negative and the coefficient of the interaction term is positive. Both are statistically significant, but as income never takes on a value of zero and the coefficient of the interaction term is only part of the conditional effect, these results do not reveal much. Figure 4 plots conditional effects for inequality at each observed value of income. It shows that inequality’s estimated marginal effects on rejecting meritocracy are negative and statistically significant for those with incomes of up to US$50,000; that is, for those in the bottom half of the sample by income. They are not distinguishable from zero for those with higher incomes.

Download

Open in new tab

Download in PowerPoint

Figure 4. Estimated coefficients of income inequality by income.

Note: results presented in Figure 3. The dots represent estimated coefficients of income inequality within respondents’ commuting zones on their rejection of meritocracy for all values of respondent family income; the whiskers represent the 95% confidence intervals of these estimates. The estimates are negative and statistically significant for those with lower incomes, while the coefficients for those with higher incomes are not distinguishable from zero. The histogram presented below the main plot depicts the relative frequency of each observed value of respondent income: 49% of respondents reported incomes below US$50,000.

Of course, given the dichotomous dependent variable, these estimates are in logits and so their magnitudes are not easily interpretable directly. Figure 5 presents the predicted probability of rejecting meritocracy according to Model 1 across the observed range of local income inequality at various incomes when all other variables are assumed to take on their median values. Given that assumption, those with the lowest incomes living where the context of income inequality is at the highest observed level are 19 percentage points (plus or minus 7 points) less likely to reject meritocracy than similarly low-income people living where inequality is at its lowest observed level. For people with incomes between US$40,000 and US$50,000 and otherwise median characteristics, the predicted probability of rejecting meritocracy declines from $37\pm 2$% to $30\pm 3$% over the observed range of inequality—a drop of $7\pm 4$ percentage points. For those with the highest incomes, given the confidence intervals, the predicted probabilities of rejecting meritocracy are consistent with no change across all levels of local income inequality. These results are contrary to the predictions of the conflict theory, but consistent with those of the relative power theory.

Download

Open in new tab

Download in PowerPoint

Figure 5. Predicted probability of rejecting meritocracy by income and level of inequality.

Note: results presented in Figure 3. Solid lines represent predicted probabilities and shaded regions represent the 95% confidence intervals of these predictions. The predicted probabilities were generated by fixing all other variables at their median values.

One of the most important questions underlying recent research on economic inequality and democracy is whether inequality is self-correcting or instead self-reinforcing. NJL argue in favor of the former, more optimistic view. It contends that mere exposure to high levels of local income inequality prompt those with lower incomes to become more likely to reject an important ideological prop of persistent inequality, the idea that those who work hard can get ahead, supporting the conflict theory. Unfortunately, the article’s own reported results do not support this conclusion, and its analysis suffers from serious measurement issues that render even those results untrustworthy. Our analysis of data from a larger and more representative survey that employs a uniform measure of the dependent variable, in fact, yields the opposite result: lower-income people living where local levels of income inequality are higher are less likely to reject meritocracy than those living where the income distribution is more egalitarian, in line with the predictions of the relative power theory.

These results have important political implications. For those who would prefer higher levels of redistribution, whether the conflict theory or the relative power theory better describes political reality is crucial to understanding the effort that will be required to reverse the decades-long trend of rising income inequality in the US. If the NJL results were sound, and the conflict theory supported, advocates of greater redistribution could remain relatively complacent, confident that by activating conflict, higher levels of inequality would more or less automatically deliver the votes needed for the policies they prefer.

The results presented here, however, suggest that the current social structure will not simply undermine itself in this fashion. That lower-income individuals living in localities with higher levels of income inequality tend to be less likely to reject the meritocratic ideology that sustains the status quo means that change, if it is to occur, will only result from concerted effort, from the difficult and much-constrained work of organization and mobilization. Absent such undertakings, the greater relative power of the wealthy to shape the views of their poorer fellow citizens in contexts of greater inequality will go uncontested.