Informed or Overwhelmed? Disentangling the Effects of Cognitive Ability and Information on Public Opinion

Abstract

Received wisdom in political science holds that informed citizens are better able to develop coherent, stable policy preferences. However, past research fails to differentiate between the effects of information and cognitive ability. I show that, for people with low levels of ability, consuming more political information predicts lower levels of ideological constraint and response stability. This effect is driven by relatively technical issues, suggesting that attempts to inform the electorate may backfire by overwhelming some voters. More broadly, these results suggest that an increasingly saturated information environment may exacerbate, rather than ameliorate, differences in political sophistication.

It is undeniably important to improve our knowledge of the dynamics of evaluation, in cases where we may assume that certain raw materials are given. Yet it seems of equal importance to understand the consequences of initial differences in these raw materials, whether they involve cognitive capacity or background of political lore.

Campbell, Converse, Miller, and Stokes, The American Voter (1960, 255)

For over half a century, research has shown that politically informed citizens are more likely to hold stable and ideologically consistent attitudes on public policy (Ansolabehere, Rodden, and Snyder 2008; Converse 1964; Freeder, Lenz, and Turney 2019; Kalmoe 2020). Political scientists often interpret this relationship in causal terms, arguing that knowledge of politics breeds conviction and coherence in one’s beliefs (Althaus 1998; 2003; Alvarez 1997; Arnold 2012; Bartels 1996; Delli Carpini and Keeter 1996; Zaller 1992; but see Goren 2013; Kraft 2024; Lupia 2016). In this view, an uninformed citizenry will struggle to translate its interests and values into votes, giving free rein to demagogues who bypass issues and prey on emotions (Barber and Pope 2019; Dahl 1989; Delli Carpini and Keeter 1996). Some scholars go so far as to argue that, without high levels of political knowledge, democracy is unsustainable (Somin 2013; Brennan 2016).

In making these arguments, political scientists tend to take for granted that information is the active ingredient that causes political knowledge to correlate with opinion quality. However, there is reason to doubt this assumption. Decades of research in cognitive psychology have shown that people differ in their cognitive abilities and that these differences crystallize in early adolescence and persist across the lifespan (Breit et al. 2024; Neisser et al. 1996). Consistent with this principle, cognitive skills in adolescence predict people’s acquisition of political knowledge in later life independently of whether they attend college (Highton 2009; Rasmussen 2016b). Similarly, lab experiments find that cognitive ability predicts people’s ability to encode and organize novel political information above and beyond prior engagement with politics (Lodge and Hamill 1986; Hamill, Lodge, and Blake 1985). Thus, despite decades of research, we have little idea to what extent the active ingredient in political knowledge scales is actually information – or to what extent it might be cognitive ability.

Building on these findings, I propose a model of information effects on public opinion that explicitly accounts for cognitive ability. I argue that qualities like attitude stability and ideological coherence are better understood, not as outcomes of being politically informed, but as products of an interaction between the specific facts, arguments, and ideas that a person encounters and their ability to parse semantic information. All else equal, people with more cognitive skills should find it easier to track whether novel arguments accord with their existing beliefs and commitments; for them, information consumption should encourage the formation of coherent and stable preferences. By contrast, people low in cognitive ability may struggle to make sense of the political rhetoric that they encounter, absorbing talking points from different sources while failing to note contradictions and gravitating to whatever appeals are most salient at a given moment; for them, information consumption may overwhelm and confuse rather than clarify, leading to preferences that are less coherent and less stable.

I test these predictions in three nationally representative US panels. Among Americans with high levels of verbal ability, I replicate the classic finding that the politically informed hold attitudes that are more constrained by left-right ideology and more stable over time. But among Americans with low levels of verbal ability, I find the opposite – for them, information consumption is often negatively related to constraint and stability. Moreover, this pattern holds across a wide range of strategies for measuring information consumption. To probe whether these effects are driven by the cognitive demands of attitude formation, I conduct two additional tests. First, I show that information backfire among low-ability respondents occurs for attitudes in a relatively technical and unintuitive issue domain – size of government – but not in an intuitive and emotionally charged issue domain – social policy. Second, I show that the moderating effect of verbal ability cannot be explained by demographics; education; income; party identification; partisan extremity; psychological motivations; or the specific print, television, radio, and online media that people consume. While these analyses are far from dispositive and cannot tell us whether or not relationships are causal, they help to rule out alternative explanations and provide context for judging the plausibility of my model (Spirling and Stewart 2024).

The implications of these findings are sobering. They suggest that an increasingly saturated information environment will exacerbate, rather than ameliorate, differences in the quality of opinions held by the most and least politically sophisticated members of the American public. These findings challenge the widespread view that information alone explains individual differences in attitude structure and stability, while also uncovering an additional mechanism by which the growth of mass media may have increased inequalities in political sophistication (Prior 2007).

Cognition, Information, and Political Sophistication

Are all people capable of developing firmly held, well-considered opinions about public policy? Observers have debated this question, in one form or another, for thousands of years. Critics of direct democracy, like Socrates of Plato’s Republic and the Irish political theorist Edmund Burke, insist that most people are too fickle to render coherent judgements about politics. Even the architects of the American Constitution take a somewhat pessimistic view of citizen competence. In a famous passage, James Madison argues that elevating mass publics to the intellectual and moral level necessary for direct democracy is a fool’s errand. He instead advocates a representative system that will ‘refine and enlarge the public views, by passing them through the medium of a chosen body of citizens, whose wisdom may best discern the true interest of their country’ (Madison 1987 [1788], 126). While scholars have occasionally espoused Madisonian views on citizen competence (for example, Lippmann 1922; Luskin 1990; Schumpeter 1942), most take for granted that rational deliberation is within reach of anyone who decides to seek out a steady diet of political information (for example, Dahl 1989; Delli Carpini and Keeter 1996; Nie, Verba, and Petrocik 1979). Thus, while political scientists often fret about the public’s lack of civic skills, the field has largely accepted the premise that most people could become politically sophisticated if they put in the work. ¹

Does knowing more about politics have large, salutary effects on political behaviour? It is certainly true that citizens who score higher on political knowledge scales demonstrate sounder political reasoning in a variety of contexts. They are more likely to hold attitudes and cast votes that are consistent with their material interests and values (Delli Carpini and Keeter 1996; Zaller 1992); they are better at learning and applying decision-making rules when choosing among political candidates (Lau, Andersen, and Redlawsk 2008; Lau and Redlawsk 2001; Sniderman, Brody, and Tetlock 1991); they are better at spotting and discounting false information even when it flatters their biases (Vegetti and Mancosu 2020); they engage in a broader and more effective information search before arriving at a decision (Bernhard and Freeder 2020; Lau and Redlawsk 2006; Singh and Roy 2014); and they make better use of the information that they encounter (Funk 1997; Gilens 2001; Lau and Redlawsk 2006). Given these findings, it is no wonder that many political scientists view knowledge as ‘an instrumental good that helps to enlighten one’s self-interest and to translate it into effective political action’ (Delli Carpini, and Keeter 1996, 218) and believe that knowledge disparities cause ‘systematically different vote choices by citizens in otherwise similar political circumstances’ (Bartels 1996, 202).

Yet, this causal story is not necessarily the correct one. Just because people who know more facts about politics exhibit better reasoning skills does not mean that learning those facts caused them to become better thinkers. Instead, political knowledge may be in part a reflection of pre-existing differences in cognitive skill. Several lines of evidence support this view. For one, people who do well on political knowledge tests tend to have high levels of cognitive ability even after accounting for differences in education, media consumption, and political interest (Harvey and Harvey 1970; Hamill and Lodge 1986; Hamill, Lodge, and Blake 1985; Highton 2009; Neuman 1986; Neuman, Just, and Crigler 1992; Nie, Junn, and Stehlik-Barry 1996; Rasmussen 2016b). ² Another line of evidence emerges from research on the dimensionality of political knowledge. Burnett and McCubbins (2020) show that people with more political knowledge also know more about shopping, sports, popular culture, geography, economics, and the rules of the road. Moreover, they find that a single latent factor explains over 80 per cent of the variance in knowledge across domains, suggesting that political knowledge is largely a measure of general learning propensity rather than anything specific to politics. In another measurement study, Pietryka and MacIntosh (2013) find that the residualized covariances among political knowledge items are negligible, implying that an underlying latent variable causes people to accumulate political knowledge (see Bollen and Ting 2000). Lastly, several experiments have found that people who score low on political knowledge scales struggle to use relevant facts to inform their decisions, even when the information is provided to them (Gilens 2001; Rahn, Aldrich, and Borgida 1994; Rahn and Cramer 1996). Together, these findings raise the likelihood that political knowledge is confounded with pre-political cognitive skills, calling into question the meaning of these scales’ tendency to predict political sophistication.

The Role of Verbal Ability

One cognitive skill in particular has been found to correlate with and predict the same host of outcomes as political knowledge scales: verbal ability, which captures a person’s aptitude for understanding, retaining, and reasoning about semantic (as opposed to mechanical or spatial) information (Wechsler 1958). High levels of verbal ability appear to reflect both greater working memory capacity and ease of memory retrieval, allowing people to rapidly interpret and encode semantic information while freeing up resources for effortful cognition (Hunt 1978; Perfetti 1985). Experimental studies find that people with high levels of verbal ability are better able to comprehend and recall political information, often outstripping the combined effects of topic-specific interest and knowledge (Eckhardt, Wood, and Jacobvitz 1991; Lodge and Hamill 1986; Neuman, Just, and Crigler 1992). Graber (1984, 195) reports similar results in a qualitative study of news consumption, noting that participants with ‘greater language facility and better ability to articulate ideas’ excelled at processing and retrieving political information encountered in the media, while ‘panelists at lower intelligence levels omitted more stories from processing and had more difficulty in retrieving complex information’. Other studies find that respondents with higher levels of verbal ability are less susceptible to question order effects (Krosnick and Alwin 1987) and are more likely to update their attitudes when the information on which those attitudes were based is shown to be false (Brydges, Gignac, and Ecker 2018; De keersmaecker and Roets 2017; McIlhiney et al. 2023). These results are consistent with the idea that verbal ability facilitates automatic, low-level components of political information processing such as parsing and storing information in long-term memory and updating cognitive representations.

Verbal ability also appears to facilitate more effortful, high-level forms of political cognition – namely, the ability to use ideology as a descriptive and inferential tool (Converse 1964). Hamill and Lodge find that verbal ability is a powerful predictor of people’s ability to map issue positions onto abstractions like liberal versus conservative, surpassing the effects of political interest, education, media consumption, income, and participation (Hamill, Lodge, and Blake 1985; Hamill and Lodge 1986). Similarly, Kinder and Kalmoe (2017, 172, n6) show that verbal ability predicts people’s ability to place themselves on an ideological continuum net of controls for knowledge, participation, and education. As one would expect given their grasp of how abstractions map onto concrete political phenomena, people with higher levels of verbal ability are also more likely to identify with the party that represents their issue positions and ideology (Ganzach 2018; Gooch 2015; Rasmussen 2016a).

A Model of Ability and Information Effects on Public Opinion

What does it mean for the claim that political knowledge is an ‘instrumental good’ that many of its purported effects can be explained, in part, by domain-general reasoning ability? Clearly, people need to have some contextual knowledge about politics to form policy preferences; verbal ability cannot make up for political ignorance. Rather than generating political sophistication from scratch, verbal ability should shape how effectively people process conflicting signals. Zaller (1992) shows that politically sophisticated individuals are better able to identify and reject arguments that conflict with their values, principles, interests, and group attachments, leaving them with more highly structured and stable belief systems. Meanwhile, people who are less sophisticated but who persist in attending to politics tend to ‘fill up their minds with large stores of only partially consistent ideas, arguments, and considerations’, leaving them with attitudes that are more unstable and incoherent than if they had received no information at all (Zaller 1992, 36). Similarly, Lau and Redlawsk (2006, 220) argue that ‘at least in politics, more information does not always result in better decisions … Evidently, additional information beyond cognitive capacity often confuses voters (or tires them out?) and actually lowers the probability of a correct value-maximizing decision.’

Take, for example, the roughly 30 per cent of Americans who hold policy positions that are at odds with their stated ideological convictions (Claassen, Tucker, and Smith 2015; Ellis and Stimson 2012). Given that ‘[p]aying attention to the news is one of the hallmarks of an informed and engaged citizen’, one might expect this group to be less attentive to political media than their peers (Ellis and Stimson 2012, 167). In fact, Ellis and Stimson show that, among Americans with middling levels of political knowledge, those who read or watch the news often are more likely to report inconsistent beliefs. Ellis and Stimson attribute their findings to the fact that American media regularly broadcast two contradictory messages: first, that social safety net programmes are both desirable and compatible with Americans’ ethos of hard work and self-reliance, and, second, that liberals give a free pass to criminals, reward deadbeats, and disrespect traditional ways of life. Given this media environment, Ellis and Stimson argue that ‘exposure to political news may not help – and may even hinder – the ability of citizens to align their own operational and symbolic beliefs’ (2012, 167–8).

The idea that encountering more political information can make people less politically sophisticated may seem counterintuitive. However, research on how the mind allocates cognitive resources suggests that lower verbal ability can force a trade-off between memorization and comprehension, which can make it difficult to learn effectively when information is abundant. Cognitive psychologists argue that the mind draws on a limited pool of attention to parse incoming signals and form long-term memories, placing these two processes in conflict (Popov et al. 2019; Popov and Reder 2020). Because people with higher levels of verbal ability possess greater working memory capacity, they are better equipped to engage in effortful processing while also encoding semantic information in long-term memory (Frischkorn, Wilhelm, and Oberauer 2022; Hunt 1978; Perfetti 1985). By contrast, people with lower levels of verbal ability more often face a trade-off – remember many facts without having thought particularly hard about what they mean, or ignore much of what you’ve heard and focus on understanding the implications of a few things at a time.

Research on voter decision-making provides direct evidence that too much information can impair political cognition for those low in political knowledge, even as it helps the knowledgeable. In a series of experiments, Rahn and colleagues vary whether information about political candidates is presented at higher or lower levels of complexity (Rahn, Aldrich, and Borgida 1994; Rahn and Cramer 1996). They find that conveying information at a greater level of complexity helped people with high levels of political knowledge and hindered people with low levels of political knowledge. Conveying information in a simpler format muted these differences, leading Rahn and Cramer (1996, 198) to conclude that low knowledge subjects ‘suffered overload in the more complex environment’. Moreover, these authors note that alternative moderators such as interest, participation, and newspaper reading all fail to produce the focal interaction, ‘suggesting that the effects of political sophistication in [their] results are based on the cognitive ability component of this construct, as would be expected by a limited capacity framework’ (Rahn and Cramer 1996, 206).

Lau and colleagues apply a self-guided version of this paradigm, allowing subjects to browse information in a simulated campaign environment before casting their votes for fictional candidates (see Andersen, Redlawsk, and Lau 2019). Subjects are then scored on the ‘correctness’ of their vote – that is, whether they chose the candidate whose policy positions align with their own. Using this paradigm, Lau and Redlawsk 2006)(2001; compare the quality of the vote choices made by people with higher and lower levels of political knowledge, conditional on how much and what type of information they viewed. As expected, subjects who scored high on political knowledge scales – and who were, therefore, likely high in verbal ability – benefited from consuming more information. However, subjects who scored low on political knowledge scales were less likely to vote correctly when they consumed more information. In another study, Kleinberg and Lau (2021) examine the effects of telling people that they can look up political facts later rather than having to remember them. They find that subjects who were told they could look up facts later viewed less information but remembered more, and made better decisions, consistent with the idea that ‘extra information (beyond one’s cognitive capacity to handle it) actually hurts decision making – that is, bad (excess) information crowding out good (processible) – by confusing voters and making them less likely to remember crucial information’ (Lau and Redlawsk 2006, 212).

While Rahn, Lau, and colleagues’ results were obtained in simulated campaign environments, quantitative analyses confirm that television and print media present more information than people can process and do so at a level of complexity that makes it difficult to fully interpret (Graber 2004; Neuman, Just, and Crigler 1992). As a result, people with lower levels of verbal ability may come away from each newscast or article with a more contradictory mix of considerations in mind than they had going in. When asked to give their opinions on policy, the mix of considerations that they pull from the top of their head will be more numerous but less coherent than a person with similar verbal ability who simply ignored the news (Zaller 1992; Zaller and Feldman 1992). Meanwhile, high-ability individuals will take advantage of additional information to identify the policy positions that best represent their interests and values. These predictions yield my first hypothesis:

Hypothesis 1 (Opposite Effects): Greater information consumption will correspond to increased constraint and stability among people with high levels of verbal ability and decreased constraint and stability among people with low levels of verbal ability.

Another implication of this theory is that the backfire effect proposed in Hypothesis 1 should be small or absent for policies that are highly salient and hence require less cognitive skill to evaluate. Social issues like same-sex marriage, abortion, and transgender rights are likely candidates because they tend to trigger rapid ‘gut-level’ emotional responses regardless of people’s level of political expertise (Hetherington and Weiler 2018, 2009; Johnston, Lavine, and Federico 2017; Johnston and Wronski 2015). These gut-level responses should, in turn, diminish the role played by cognitive processing in turning information into attitudes. Using same-sex marriage as an example, Johnston and Wronski explain that ‘the key referent for the issue of gay marriage (i.e., homosexuality) is inseparable from the policy itself, and the activation of feelings and beliefs related to moral traditionalism is unlikely to require much political knowledge’ (2015, 37). Indeed, Johnston and colleagues have demonstrated that politically disengaged people often possess strong intuitions about issues like homosexuality and abortion but not economic policy (Johnston, Lavine, and Federico 2017; Johnston and Wronski 2015).

By contrast, debates over economic policies like spending, taxes, and regulation tend not to inspire much conviction beyond the most knowledgeable and ideological parts of the electorate (Carmines and Stimson 1980; Pollock, Lilie, and Vittes 1993). Here, verbal ability should play a decisive role. For example, Neuman, Just, and Crigler (1992) show that verbal ability has no bearing on how effectively people learn about emotionally charged topics like drug abuse and AIDS. However, when it comes to issues ‘that involved technical information – Star Wars [a proposed missile defence system] and the stock market crash – subjects with high cognitive skill and low attention learned significantly more than those with low cognitive skill and high attention, even though the two groups had scored about the same before the news exposure’ (Neuman, Just, and Crigler 1992, 105–6). These predictions yield my second hypothesis:

Hypothesis 2 (Domain Difficulty): The negative effect of information consumption at low levels of verbal ability proposed in Hypothesis 1 should appear for economic issues but not social issues.

Before moving on, it is important to note that my theory rests on the assumption that cognitive ability is exogenous to information consumption. As Figure 1 illustrates, this assumption departs from prevailing theories of political knowledge that emphasize the role of information exposure in shaping cognitive ability. This is not to deny the evidence that reading during childhood promotes cognitive development (Ritchie, Bates, and Plomin 2015). Rather, my argument is specifically that the kind of media exposure that people rely on for political information – reading the newspaper, browsing political blogs, watching cable news, and receiving second-hand reports from friends and family – does not affect verbal ability in adulthood. In Appendix B, I test this assumption using three General Social Survey (GSS) panel studies. Results from a bivariate dual change score model show that within-person changes in newspaper reading are unrelated to subsequent within-person changes in verbal ability, providing strong support for exogeneity (Table B1; Grimm et al. 2012).

Figure 1. Models of ability and information effects on public opinion.

Lastly, I note that my model does not assume that the information that people consume is exogenous to their level of ability (see Figure 1). Therefore, if some proxy measure of information consumption (for example, self-reported attention to politics in the media) has different effects at different levels of verbal ability, it could be that the higher-ability respondents are simply paying attention to different, more politically informative, media. While plausible, this concern is not borne out in my analysis. Using detailed data on where people get their political news, I show in a later section that verbal ability continues to moderate the effects of information consumption even when controlling for the moderating effects of media diet.

Data and Methods

Data

My data consists of three American National Election Studies (ANES) panels spanning 2008–2010, 2012-2013, and 2016-2020, respectively. The first sample was recruited as part of the ANES 2008-2009 Panel Study, which consisted of a telephone recruitment interview and demographic data collection in November 2007, followed by twenty-one internet-based surveys from January 2008 through September 2009 (DeBell, Krosnick, and Lupia 2010). Participants were interviewed again in June 2010 as part of the 2010 Panel Recontact Survey (DeBell et al. 2010). The 2008-2010 sample consists of two cohorts recruited via random digit dialing in November 2007 and the summer of 2008, respectively. I focus only on the first cohort, who completed all key independent variables. The recruitment interview for this cohort yielded data for 2,360 respondents; in subsequent waves, the number of respondents fluctuated between 1,108 and 1,623, with 856 completing the 2010 survey. The second sample was drawn from the ANES 2012 Time Series study (ANES 2014). Participants were recruited using a combination of address-based sampling and random digit dialing, yielding 2,054 face-to-face interviews and 3,860 internet-based interviews conducted in the months before and after the 2012 presidential election. Of the internet-based sample, 1,563 were reinterviewed in July 2013 as part of the 2013 Internet Recontact Study (ANES 2013). The third sample was drawn from the ANES 2016 Time Series study (ANES 2019). Participants were recruited via address-based sampling, yielding 1,180 face-to-face interviews and 3,090 internet-based interviews conducted before and after the 2016 presidential election. Of the combined face-to-face and internet-based samples, 2,839 were reinterviewed as part of the ANES 2020 Time Series Study (ANES 2021).

Dependent Variables

To measure ideological constraint, I use all items administered in the 2008-2009 Panel Study, 2012 Time Series, and 2016 Times Series that ask respondents to place themselves on a policy debate on which the Democratic and Republican parties hold stable, principled disagreements. Following this criterion, I exclude questions about defence spending, crime spending, reducing the budget deficit by taxing the wealthy, international trade, and civil liberties from my constraint indices (see Appendix D). I first rescale each item to range from 0 to 1, where low scores indicate more liberal/left-wing positions and high scores indicate more conservative/right-wing positions. Respondents who refused to answer an item or said ‘don’t know’ are coded as missing, and respondents who replied ‘I haven’t thought much about this’ when prompted are assigned to the midpoint of the scale. ³ I then calculate the standard deviation of respondent’s policy attitudes across all items (Barton and Parsons 1977). Lastly, I reverse and rescale the resulting measure to range from 0 (least constrained) to 1 (most constrained).

To measure attitude stability, I used all policy items that were fielded two or more times in a panel (Appendix D). After recoding all items as described above, I take the standard deviation of responses to the same item across time (Elder and O’Brian 2022). I then average, reverse, and rescale these standard deviations to create an individual-level stability index that ranges from 0 (least stable) to 1 (most stable). For both constraint and stability, respondents who answered fewer than 90 per cent of the selected policy items are coded as missing. The remaining respondents are only scored on the items that they answered. I use a more lenient threshold of 50 per cent for calculating stability in the 2008-2010 panel to accommodate high levels of wave non-response (see Appendix E).

A potential objection to treating response stability as a sign of firmly held attitudes is that most policy items were asked only twice, meaning that we cannot be certain whether change represents vacillation or genuine opinion change. While this represents a limitation of the current study, existing evidence suggests that the vast majority of change observed in public opinion surveys is random rather than durable. Hill and Kriesi (2001) analyze attitudes toward six environmental regulations in a four-wave, nationally representative Swiss panel. They place the proportion of citizens moving from one stably held opinion to another at between 2 per cent and 8 per cent over a 2-year period, depending on the issue. Similarly, Feldman’s (1989) analysis of a five-wave panel reveals virtually no durable change in Americans’ issue positions in the months leading up to the 1976 US presidential election.

Independent Variables

To measure verbal ability, I use two versions of Wordsum, a short vocabulary test developed for use in public opinion surveys (Thorndike 1942; Thorndike and Gallup 1944; see Malhotra, Krosnick, and Haertel 2007). A 10-item version of the test was included in the ANES 2012 and 2016 Time Series studies (ANES 2014, 2019). A 14-item version developed by Cor, Haertel, Krosnick, and Malhotra (2012) was fielded in the eighth wave of the ANES 2008-2009 Panel Study. Both versions consist of items that ask respondents to indicate which of five words is closest in meaning to a target word (for example, ‘beast: 1. afraid, 2. words, 3. large, 4. animal, 5. separate, 6. don’t know’). To achieve a correct answer, respondents must either know or infer the meanings of the words and weigh their relative similarities. I assign 0’s to incorrect and ‘don’t know’ responses and 1’s to correct responses. I then average the scores to create an additive scale ranging from 0 (lowest ability) to 1 (highest ability).

Despite its brevity and simplicity, studies report consistently high correlations between Wordsum and tests of verbal ability, abstract reasoning, math ability, and general intelligence (r ≈ .45-.85; Hagen and Thorndike 1955; J. L. Huang and DeSimone 2021; M.-H. Huang and Hauser 1998; Miner 1957; 1961). However, much of the evidence for Wordsum’s breadth and convergent validity hinges on data that is now at least seventy years old, and some have speculated that the test’s validity has decayed as its vocabulary words have gone out of use (Wilson and Gove 1999). In Appendix A, I address this concern by testing Wordsum’s convergent validity in the 1987 and 1994 GSS samples, which were administered Wordsum and abstract reasoning tests taken from the Sixteen Personality Factor Questionnaire (16PF; Cattell, Eber, and Tatsuoka 1970) and the Wechsler Adult Intelligence Scale-revised (WAIS-R; Wechsler 1981), respectively. Using structural equation modeling to correct for measurement error, I find that Wordsum and abstract reasoning ability correlate at 0.631 in the 1987 GSS and 0.626 in the 1994 GSS (see Tables A1-A2). The strength of these disattenuated correlations suggests that, as recently as the 1990s, Wordsum retained its ability to measure a broad ability domain that includes abstract reasoning.

The task of measuring information consumption is less straightforward. Political scientists have long sought to identify the best method for measuring the amount of political information that people encounter, with mixed results (Bartels 1993; Dilliplane, Goldman, and Mutz 2013; Price and Zaller 1993; Prior 2009a; 2009b; 2013). Given this lack of consensus, I opt to test my hypotheses across six commonly used information consumption proxies. These are shown with example items in Table 1. I measure each proxy by creating an equally weighted composite of its items and rescaling this composite to range from 0 to 1 (for its items, see Table A2). ⁴ While most of these measures are self-explanatory, two require clarification. Candidate-issue placement knowledge is measured by assigning one point to respondents who place the Democratic candidate to the left of the Republican candidate on a given issue and zero points to all others, including those who do not know one or both candidates’ stances (Delli Carpini and Keeter 1993; Freeder, Lenz, and Turney 2019; Zaller 1992). General political knowledge is measured using a variety of multiple-choice and open-ended factual questions about political figures and institutions. Correct responses receive one point and both incorrect and ‘don’t know’ responses receive zero points (Brown and Pope 2021; Luskin and Bullock 2011). For items where the option is available, partially correct responses receive half a point (DeBell 2013).

Table 1. Information consumption proxy measures

The information proxies in Table 1 can be divided into two broad types – self-reports and objective tests. Among the self-reports, a further distinction can be made between subjective evaluations and frequency estimates. Items tapping political interest, while not measures of information consumption per se, are highly correlated with exposure to political information and boast excellent reliability (Prior 2007; 2019). The same is true of items tapping attention to politics in the media, which have the added benefit of asking directly about political information consumption (Prior 2019). The primary drawback of these items is their use of ambiguous response options such as ‘very’, ‘somewhat’, and ‘not much’. By contrast, media consumption frequency and political discussion frequency items ask about a concrete quantity – the number of days in a week that something occurred – but exhibit high levels of measurement error (Bartels 1993; Konitzer et al. 2021; Morey and Eveland Jr. 2016; Price and Zaller 1993; Prior 2009b; 2009a). Faced with a choice between vaguely categorized evaluations and error-prone frequency estimates, some have recommended eschewing self-reports altogether in favour of objective tests (Price and Zaller 1993; Zaller 1990). However, the objective tests have their own mixture of pros and cons. On the positive side, they display high levels of reliability and criterion validity (Delli Carpini and Keeter 1993; Pietryka and MacIntosh 2013; Price and Zaller 1993; Zaller 1990). On the negative side, both issue placement and general political knowledge are confounded with verbal ability (Hamill, Lodge, and Blake 1985; Highton 2009; Neuman, Just, and Crigler 1992; Rasmussen 2016b).

In this context, self-reports possess an important quality that makes them valuable, if noisy and subjective, information consumption proxies. On average, it takes no more cognitive skill to pick ‘very interested’ or ‘7 days’ than it does to pick ‘not much interested’ or ‘0 days’, whereas it does take more cognitive skill to answer a knowledge question correctly than it does to answer it incorrectly. Therefore, only self-reports allow us to assess how likely a person is to consume political information without selecting on their ability to store and retrieve memories and hence selecting on verbal ability. In Appendix C, I demonstrate this by using exploratory Item Response Theory models to scale the information proxy and Wordsum items in the 2012 and 2016 ANES. ⁵ As shown in Figure 2, the self-report items load primarily on latent factors defined by the amount of attention that people pay to politics in the media. By contrast, the objective test items load heavily on latent factors defined by verbal ability and load only modestly on the media attention factors. Also noteworthy is the fact that the frequency estimate items do not load as highly on the media attention factors as the subjective assessment items. This is consistent with evidence that these items are prone to measurement error (Bartels 1993; Price and Zaller 1993).

Figure 2. Scaling information consumption and verbal ability items.

Note: Results are discrimination parameters from exploratory multidimensional IRT models. Difficulty parameters are estimated but not shown. Results are varimax rotated to produce two orthogonal latent factors. The model output is in Table C1.

Given that objective tests appear to be confounded with verbal ability, models that use them will effectively be interacting verbal ability (as measured by Wordsum) with itself (as measured by the objective tests). This should bias my results in a predictable way: Because respondents with low Wordsum scores and high knowledge scores will tend to be higher in underlying ability than those with low Wordsum and low knowledge scores, the marginal effect of information at low levels of verbal ability will be biased upward. Similarly, if frequency estimates are especially affected by random measurement error, then models that use them should tend to yield attenuated estimates of information effects. Therefore, I expect self-reports to deliver the strongest confirmation of Hypothesis 1.

Do Information Effects Depend on Verbal Ability?

According to Hypothesis 1, the relationship between political information consumption and attitude quality should be positive for respondents with high levels of verbal ability and negative for respondents with low levels of verbal ability. To test this hypothesis, I begin by estimating thirty-six linear regressions – one for each combination of a dependent variable and an information proxy in each of my three samples. To summarize these results, I also estimate twelve models pooling across samples using hierarchical linear regression. The focal independent variables in each regression are verbal ability, one of the six information proxies, and an ability-information interaction term. Each regression also includes a set of demographic control variables – age, gender, race, education, and income – measured in the first wave of each panel. All variables except for age are scaled to range from 0 to 1. I present the focal interactions from each model in Table 2 and report the full results in Appendix F.

Table 2. Verbally ability moderates the relationship between information consumption and attitudes

Note: Entries are unstandardized regression coefficients with standard errors in parentheses. Bolded coefficients are statistically significant at the p < 0.05 level. For full results and alternative specifications see Tables F1-F16.

All twenty-four of the constraint models yield interaction terms that are positive and statistically significant, providing strong initial support for Hypothesis 1. The models predicting stability, meanwhile, offer only partial support. Among the eighteen stability models estimated on individual panels, four interactions are positive and statistically significant – political interest in 2016-2020, attention to politics in the media in 2012-2013 and 2016-2020, and news consumption frequency in 2008-2010. Of the six stability models estimated on the pooled samples, two interactions are positive and statistically significant – political interest and attention to politics in the media. Contrary to Hypothesis 1, there are also several statistically significant interactions whose coefficients are negative – specifically, the interactions for candidate-issue placement knowledge and general political knowledge in 2012-2013 and in the pooled samples.

To further interpret these results, Table 3 reports marginal effects from the pooled models at the fifth and ninety-fifth percentiles of verbal ability, and Figure 3 visualizes these marginal effects across the entire range of verbal ability. Looking first at the results from the constraint models, I find strong confirmation of Hypothesis 1. At the ninety-fifth percentile of verbal ability, all marginal effects are positive and significant. Here, moving from the lowest to the highest level of an information proxy is associated, on average, with a thirteen-percentage point increase in constraint. But at the fifth percentile of verbal ability, all of the marginal effects are significantly negative with the exception of general political knowledge, which is small and non-significant. Here, moving from the lowest to the highest level of an information proxy is associated, on average, with a roughly six-percentage point decrease in ideological constraint.

Table 3. Effects of information proxies by verbal ability percentile (pooled models)

Note: Entries are unstandardized regression coefficients with standard errors in parentheses. Bolded coefficients are statistically significant at the p < 0.05 level.

Figure 3. The relationship between information consumption and attitudes depends on verbal ability (pooled models).

Note: Plot lines are marginal effects with 95 per cent confidence intervals from Pooled models reported in Table 2.

Turning next to the stability results, I find partial confirmation of Hypothesis 1, but only in models using self-report proxies. The marginal effect of attention to politics in the media is significantly positive at the ninety-fifth percentile of verbal ability and significantly negative at the fifth percentile of verbal ability, mirroring the results for constraint. The marginal effects of two other self-report proxies – political interest and political discussion frequency – are significantly positive at high levels of verbal ability and non-significant at low levels of verbal ability. News consumption frequency is not a significant predictor of stability at either level of verbal ability. Lastly, the marginal effects of candidate-issue placement knowledge and general political knowledge are positively and significantly related to stability at high levels of verbal ability, as expected. Contrary to expectations, however, these effects are even larger at lower levels of verbal ability.

While the stability results are only partly consistent with Hypothesis 1, it is worth noting that this may reflect differences in bias and reliability among the information proxies outlined in Table 1. The proxies that yielded positive interactions in the pooled stability models were both subjective assessments, which are highly reliable and not confounded with verbal ability (Prior 2019). By contrast, both objective tests yielded negative interactions in the pooled stability models. As demonstrated in Figure 2, the objective test items pick up heavily on latent verbal ability, potentially exaggerating the marginal effect of information among those with low Wordsum scores. Meanwhile, the two frequency estimates yielded smaller and non-significant interactions in the pooled stability models, potentially because they are measured with more error than the other proxies (Bartels 1993).

Is Ability More Important in Less Intuitive Policy Domains?

The above results show that, among people with low levels of verbal ability, consuming more political information often corresponds to lower levels of constraint and stability. According to Hypothesis 2, these negative relationships should disappear when we look exclusively at attitudes toward social issues like same-sex marriage and abortion. This is because these issues tap directly into gut-level intuitions about social change and autonomy, making information pertaining to them easier for people to process (Johnston, Lavine, and Federico 2017; Johnston and Wronski 2015). Meanwhile, these negative relationships should hold when we look only at issues like government spending and taxation that do not resonate as strongly with people’s intuitions and, therefore, require more thought (Neuman, Just, and Crigler 1992; Pollock, Lilie, and Vittes 1993).

To test this hypothesis, I first calculate domain-specific measures of constraint and stability. I follow work that identifies two core dimensions of political ideology in the American public: social policy – having to do with the tension between traditional religious morality and bodily autonomy – and size of government – having to do with the amount of spending and taxation undertaken in service of the welfare state (Ellis and Stimson 2012; Feldman and Johnston 2014; McClosky and Zaller 1984). The 2008-2010 panel does not have enough items to estimate social policy constraint and the 2012-2013 panel does not have enough repeated items to estimate stability for either domain. Therefore, I focus on the 2016-2020 panel. Table 4 shows the items assigned to each domain and their average stabilities. Reassuringly, the social policy items are generally more stable than the size of government items, as would be expected if social issues were generally easier for citizens to evaluate.

Table 4. Domain-specific constraint and stability items

Note: Entries are mean item stabilities with standard deviations in parentheses.

Using these domain-specific measures of constraint and stability, I estimate a series of linear regressions with the same sets of independent variables as those reported in Table 2. I show the focal interactions in Table 5 and report the full results in Appendix G. As predicted, verbal ability is a more consistent moderator in the size-of-government models than in the social policy models. In the models predicting constraint, all six size-of-government models yield significant positive interactions, compared to only three social policy models. And in the models predicting stability, three size-of-government models yield significant positive interactions, compared to zero social policy models. As before, the subjective evaluation proxies deliver the strongest results, followed by the discussion frequency proxy.

Table 5. Domain-specific results in the 2016-2020 panel

Note: Entries are unstandardized regression coefficients with standard errors in parentheses. Bolded coefficients are statistically significant at the p < 0.05 level. For full results and alternative specifications see Tables G1-G8.

Next, I report the marginal effect of each information proxy at the fifth and ninety-fifth percentiles of verbal ability in Table 6 and plot these marginal effects across the entire range of verbal ability in Figure 4. As expected, nearly all of the marginal effects estimated at the ninety-fifth ability percentile are positive and statistically significant. Notably, for people with high levels of verbal ability, information consumption predicts constraint and stability to roughly the same extent for social attitudes and size-of-government attitudes. This can be seen clearly in Figure 4, where the marginal effects mostly overlap at high levels of ability. By contrast, clear differences between the policy domains emerge at low levels of ability. At the fifth ability percentile, all of the social policy marginal effects are either non-significant or positive, while the size-of-government marginal effects are often significantly negative. In the size-of-government models, the average effect of moving from the lowest to the highest level of an information proxy for low-ability respondents is a six-percentage point reduction in constraint and a four-percentage point reduction in stability. This pattern is most visible for the subjective evaluation proxies.

Table 6. Marginal effects of information proxies by verbal ability percentile and policy domain in the 2016-2020 panel

Note: Entries are unstandardized regression coefficients with standard errors in parentheses. Bolded coefficients are statistically significant at the p < .05 level.

Figure 4. The moderating effect of verbal ability differs by policy domain (2016-2020).

Note: Plot lines are marginal effects with 95 per cent confidence intervals from models reported in Table 5

In short, when people with low levels of verbal ability consume more political information, their economic attitudes tend to be more scattered and unstable, whereas their social attitudes are no less structured or stable than those of their inattentive peers. Given that size-of-government is the more technical and unintuitive issue domain, these results are consistent with the idea that people with low levels of verbal ability may be overwhelmed or confused by political information. Of course, there are other plausible explanations for these results, and my interpretation should not be taken as a claim that I have identified a causal effect of issue difficulty. Rather, the purpose of this analysis is to increase the descriptive information available for judging the plausibility of my theory (Spirling and Stewart 2024).

Addressing Alternative Explanations

As with any cross-sectional analysis, it is possible that my results are biased by the omission of confounding variables. Verbal ability is correlated with age, gender, racial identity, educational attainment, and income (Cor et al. 2012; M.-H. Huang and Hauser 1998; Strenze 2007); therefore, it could be capturing demographic or socioeconomic group differences in the incentives and opportunities that lead people to adopt ideologically consistent postures (Coppock and Green 2022; Groenendyk, Kimbrough, and Pickup 2023; White, Laird, and Allen 2014). Democrats and more opinionated people also tend to score higher on verbal ability (Mazur 2023; Shoots-Reinhard et al. 2021), raising the possibility that verbal ability is capturing partisan differences in elite messaging and coalition structure (Grossmann and Hopkins 2015; Lelkes and Sniderman 2016) or partisan identity strength (Huddy, Mason, and Aarøe 2015). Similarly, three psychological traits that are plausibly correlated with verbal ability – motivations to engage in effortful cognition, experience strong emotions, and render firm judgements – have been shown to moderate the relationship between information consumption and attitudes, and therefore represent potential confounds (Arceneaux and Vander Wielen 2013; Bakker and Lelkes 2018; Federico and Schneider 2007; Holbrook 2006; Lee 2021).

Lastly, people with higher verbal ability tend to read the newspaper more often, watch television less (Glenn 1994; Neuman, Just, and Crigler 1992), and seek out news sources that reflect their political biases (Shoots-Reinhard et al. 2021). Thus, verbal ability may be a mere proxy for whether a person consumes ‘high-brow’ media like National Public Radio (NPR) – media that carries ‘the rich diet of national and international news necessary to create political awareness’ – or ‘low-brow’ media like conservative talk radio (Zaller 1992, 34). Along these lines, Claassen, Tucker, and Smith (2015) find that regular Fox News viewers are more likely to incorrectly label liberal policy positions as conservative, even after controlling for political knowledge and education.

Table 7 shows the availability of items used to operationalize these potential confounding variables. Using these measures, I re-estimate the sample-specific models reported in Table 2 under three different specifications – (1) only the focal independent variables and their interaction; (2) with the controls listed in Table 7; and (3) with the controls and their pairwise interactions with the information proxy (Blackwell and Olson 2022). In Table 8, I report the interactions between verbal ability and information consumption from each of these models. The full results are in Appendix H. With the interaction controls added, seven of the twenty-four ability-information interactions predicting constraint become non-significant. However, most of these coefficients remain roughly the same size. Among the stability models, only the political interest interaction in 2016-2020 becomes non-significant, and even here the coefficient only shrinks by about thirty per cent. Because my data is cross-sectional, no combination of controls can demonstrate that the focal regression coefficients represent causal effects. However, the fact that verbal ability continues to moderate the effects of the information proxies when controlling for potential confounders makes it less likely that these effects can be explained by alternative pathways.

Table 7. Operationalizing alternative explanations

Table 8. Controlling for potential confounders fails to eliminate focal interactions

Note: Entries are unstandardized regression coefficients with standard errors in parentheses. Bolded coefficients are statistically significant at the p < .05 level. For full results see Tables H1-H9.

Conclusion

[E]ven under a more facilitative regime, the combination of limited cognitive resources and competing attentional demands may keep politics a minority pursuit, as it seems to have been even in ancient Athens

Luskin (1990, 353)

Observers have long noted that, in America, the mass public and political elite share a secular faith: the belief that the greatest and most essential function of government is to ensure an equal voice in the political process (Arendt 1963; de Tocqueville 2002 [1835-1840]). With the advent of public opinion polling, we can now see just how far America falls short of this ideal in practice. Most citizens – particularly those with less money and education – do not have the influence on political outcomes that they theoretically could (Gilens 2012; Schlozman, Verba, and Brady 2012). Pessimistic accounts like Luskin’s, quoted above, view this as a regrettable but probably inevitable feature of mass politics.

At first glance, my results appear to vindicate Luskin’s conclusion. Across three nationally representative panels, I find that Americans with low levels of verbal ability report attitudes that are less structured and less stable when they consume more political information. Thus, contrary to a widely accepted view in political science, these results suggest exposure to information tends to benefit citizens who are already skilled at parsing it while hindering those who already struggle to keep up. This dynamic may explain why the cheap and abundant learning opportunities afforded by mass media have failed to level the playing field between political experts and political novices (Kinder and Kalmoe 2017; Luskin 1990). Far from closing gaps in political fluency, an increasingly saturated political media environment appears to have left the less politically sophisticated ‘blown about by whatever current of information manages to develop the greatest intensity’ (Zaller 1992, 311).

So, was Luskin right? On the one hand, my results concur with studies by Graber (1984) and others, which show that the media environment places a prohibitive cognitive load on citizens, making information processing skills important for determining who develops firmly held, ideologically structured attitudes. To the extent that politics will always be somewhat cognitively demanding, a limited version of Luskin’s claim may be true – not everybody will have the ability or the inclination to participate meaningfully in mass politics. On the other hand, this does not mean that greater equality of voice is unattainable. As scholars like Prior (2014), Rahn, Aldrich, and Borgida (1994), and Neuman, Just, and Crigler (1992) have shown, presenting information in a more digestible format can greatly diminish differences in learning and competency between political experts and political novices. Therefore, my results speak less to the feasibility of universal political participation than to the consequences of increasing the quantity of available political information without attending to its quality. Not only will this not bring us closer to achieving equality of voice – it may distance us from it (cf. Prior 2007).

I also find that verbal ability conditions the effects of information consumption to a greater extent for economic attitudes than social attitudes, consistent with the idea that verbal ability helps people form attitudes on issues that are less intrinsically polarizing. However, it is important to stress that this evidence is only suggestive, not dispositive. More fine-grained research designs are necessary to infer what is happening inside people’s heads, and experiments are necessary to demonstrate a causal effect of information on attitudes. In this vein, a handful of studies have used treatments such as priming ideology and encouraging subjects to stop and think before answering to induce constraint and stability (Fiske, Kinder, and Larter 1983; Judd and Downing 1990; Keating and Bergan 2017; Lavine, Thomsen, and Gonzales 1997; Milburn 1987; Zaller and Feldman 1992). These treatment effects are consistently stronger among the politically knowledgeable, but researchers have yet to investigate how much of this is due to prior knowledge and how much is due to cognitive skill.

Overall, my results point to a fundamental weakness in the way that political scientists study information effects in cross-sectional data. For decades, the norm has been to use political knowledge scales as measures of information-holding. Following this approach, many have predicted that major changes in political behaviour and policy outcomes would result if the electorate were fully informed (Althaus 1998; 2003; Alvarez 1997; Arnold 2012; Bartels 1996; Delli Carpini and Keeter 1996). Yet, the existence of ‘initial differences [in] raw materials’ makes it difficult to say whether information is the active ingredient that gives political knowledge scales their predictive power (Campbell et al. 1960, 255). Without research designs that can discern the unique effects of information and cognitive ability, we will be left without a clear understanding of how either influences political behaviour.