Reasons to challenge survey-based findings of the lack of mass public antisemitism in western democracies

Besides the nonattitudes critique, there are other reasons to challenge the view that mass-based antisemitism is virtually nonexistent in western nations, which are worth mentioning, although will not be addressed in depth here. First, the number of antisemitic incidents and hate crimes has burgeoned recently, with antisemitic events comprising the bulk of hate crimes in the United States. Such incidents spike in response to conflict between Israel and Palestinians, the behavior and public rhetoric of major political figures, and traditional scapegoating events, such as economic downturns and the Covid-19 pandemic (Feinberg, Reference Feinberg2020; Teter, Reference Teter2020; Lockwood, Reference Lockwood2021; Vergani et al., Reference Vergani, Goodhardt, Link, Adamczyk, Freilich and Chermak2022; Enstad, Reference Enstad2023). After October 7, 2023, Hamas attack and during the Israel–Hamas attack, antisemitic incidents in the United States surged (Contreras, Reference Contreras2024), indicting the power of events to spur antisemitic incidents.

Second, major public figures, such as Donald Trump, Elon Musk, and Kanye West, among others, have made public and highly visible antisemitic remarks (Steiner, Reference Steiner2020). Their public rhetoric potentially can induce higher levels of outgroup blame and prejudice, including antisemitism (Crandall et al., Reference Crandall, Miller and White2018).Footnote ¹ Some research suggests that leader rhetoric may affect public opinion. Not only may leader rhetoric affect the public, but shifts in public opinion may affect leader rhetoric (Cohen, Reference Cohen1995; Druckman and Holmes, Reference Druckman and Holmes2004; Druckman and Jacobs, Reference Druckman and Jacobs2015). Increased antisemitic rhetoric of leaders may reflect greater acceptance of antisemitic behavior and rhetoric than public opinion polls suggest. Third, antisemitic content appears rife on the internet and social media (Hübscher and Von Mering, Reference Hübscher and Von Mering2022; Riedl et al., Reference Riedl, Joseff, Soorholtz and Woolley2022), rising in the wake of October 7, 2023, Hamas attack and the Israel–Hamas war (Halamish, Reference Halamish2024).

Fourth, antisemitism has a protean quality, altering in form and content in response to changing conditions and contexts (Laqueur, Reference Laqueur2006). One argument is that antisemitism has evolved into “the new antisemitism,” which conflates antagonism toward Israel and sympathy toward Palestinians with a generalized antipathy to Jews worldwide (Klug, Reference Klug2003; Dinnerstein, Reference Dinnerstein2004; Wistrich, Reference Wistrich2004; Kaplan and Small, Reference Kaplan and Small2006; Baum and Nakazawa, Reference Baum and Nakazawa2007; Gerstenfeld, Reference Gerstenfeld2007; Hirsh, Reference Hirsh2007; Judaken, Reference Judaken2008; Cohen et al., Reference Cohen, Jussim, Harber and Bhasin2009, Reference Cohen, Jussim, Bhasin and Salib2011; Kempf, Reference Kempf2012, Reference Kempf2015; Lipstadt, Reference Lipstadt2012; Wistrich, Reference Wistrich2012, Reference Wistrich2015; Rosenfeld, Reference Rosenfeld2015; Baum et al., Reference Baum, Kressel, Cohen and Jacobs2016; Chanes, Reference Chanes, Baum, Kressel, Cohen and Jacobs2016; Jaspal, Reference Jaspal2016; Kressel, Reference Kressel, Baum, Kressel, Cohen and Jacobs2016; Beattie, Reference Beattie2017; Bobako, Reference Bobako2017; Silva, Reference Silva, Casaca and Wolf2017; Brym, Reference Brym2019; Staetsky, Reference Staetsky, Campbell and Klaff2019). The new antisemitism seems especially prevalent among college students in recent years (Alterman, Reference Alterman2016; Saxe et al., Reference Saxe, Sasson, Wright and Hecht2018; Wright et al., Reference Wright, Shain, Hecht and Saxe2018, Reference Wright, Volodarsky, Hecht and Saxe2021; Royden and Hersh, Reference Royden and Hersh2022; Hersh and Royden, Reference Hersh and Royden2023) and young minorities in the United States (Hersh and Royden, Reference Hersh and Royden2023). The melding of anti-Israel with anti-Jewish attitudes became publicly prominent with the Israel–Hamas war, as evidenced with large demonstrations across western nations, where marchers chanted “from the river to the sea,” a slogan espousing the destruction of Israel, as well as the spike of Pro-Gazan protests on college campuses, especially in the United States. Jewish students perceived an increase in campus antisemitism during the Gaza protests, often feeling unsafe (Hartocollis, Reference Hartocollis2024); research suggests that characteristics of Jewish students, including their religiosity, may affect their perceptions of antisemitism (Kosmin and Keysar, Reference Kosmin and Keysar2015). Similar attributes of Jews in the population at-large also appear to affect their perceptions of antisemitism (Cohen, Reference Cohen2010; Rebhun, Reference Rebhun2014).

Correspondingly, support for Israel has declined and sympathy for Palestinians has risen over the past 20–30 years, and a gap in support for Israel is especially pronounced among younger people (BenLevi et al., Reference BenLevi, Cavari and Terris2019; Cavari and Freedman, Reference Cavari and Freedman2020). Polls regarding the Israel–Hamas war of 2023–2024 detect a large generational divide, with young people more likely to side with the Palestinians, and even Hamas, while older individuals still display sympathy toward Israel. This age divide exists across partisanship.Footnote ² It is unclear whether this generational divide is due to cohort or life cycle processes, and there are indications that support for Israel has suffered even among young American Jews (Waxman, Reference Waxman1994, Reference Waxman2017; Sasson et al., Reference Sasson, Kadushin and Saxe2010; Sasson et al., Reference Sasson, Phillips, Wright, Kadushin and Saxe2012; Saxe and Boxer, Reference Saxe and Boxer2012), as well as among non-Jewish young people in the United States (Cavari and Freedman, Reference Cavari and Freedman2020).

Attitudes, nonattitudes, and antisemitism in public opinion surveys

A fifth reason to challenge the idea of virtually non-existent antisemitism in western mass publics is that we know little about the properties of public opinion regarding Jews, especially as measured with public opinion surveys. For instance, are the low levels of anti-Jewish opinion a function of social desirability effects, where respondents give pollsters socially acceptable answers (positive responses regarding Jews), although respondents true opinions are less favorable? Several studies have tested for social desirability effects, with mixed results (Kane et al., Reference Kane, Craig and Wald2004; Berinsky and Mendelberg, Reference Berinsky and Mendelberg2005; Winiewski and Bilewicz, Reference Winiewski, Bilewicz, Kovács and Miller2008; Krumpal, Reference Krumpal2013; Cohen, Reference Cohen2021; Cheng et al., Reference Cheng, Greene and Kingsbury2022).

Further, are attitudes toward Jews true attitudes or non-attitudes? Non-attitudes are weakly held beliefs which are likely to change, often randomly, in response to various stimuli, including previous questions or events of the day (Converse, Reference Converse1962, Reference Converse and Apter1964; Achen, Reference Achen1975; Erikson, Reference Erikson1979; Zaller, Reference Zaller1992; Zaller and Feldman, Reference Zaller and Feldman1992; Iyengar, Reference Iyengar2023). To accept the findings that most non-Jews in western nations hold positive attitudes toward Jews requires their responses to survey questions to be true, meaningful attitudes toward Jews.

Political scientists and other disciplines have investigated non-attitudes in the mass public since Converse's (Reference Converse and Apter1964) study. Early nonattitudes research focused primarily on policy positions, such as welfare, education, the economy, etc. (Achen, Reference Achen1975; Erikson, Reference Erikson1979). Over time, research expanded to other opinions and attitudes, like party identification (Schickler and Green, Reference Schickler and Green1997; Green et al., Reference Green, Palmquist and Schickler2004; Green and Platzman, Reference Green and Platzman2022), symbolic political orientations (Krosnick, Reference Krosnick1991), immigration (Kustov et al., Reference Kustov, Laaker and Reller2021), economic redistribution (O'Grady, Reference O'Grady2019), political interest, morality, culture (Kiley and Vaisey Reference Kiley and Vaisey2020) in the United States and other nations. These latter studies often unearthed considerable stability, especially after incorporating measurement error models recommended by early critics of Converse's analysis (e.g., Achen, Reference Achen1975; Erikson, Reference Erikson1979; Zaller, Reference Zaller1992; Zaller and Feldman, Reference Zaller and Feldman1992).

To my knowledge, no study of non-attitudes in public opinion toward Jews exists. One way of distinguishing attitudes from non-attitudes is to look at the stability of responses across survey panels. Non-attitudes will exhibit random change; responses at panel t ₀ will not be highly correlated with responses at panel t ₁, and sometimes the correlation of attitudes at t ₀ and t ₂ will be higher than that between t ₀ and t ₁. Although survey questions regarding opinion toward Jews have become more common in recent years (Enstad, Reference Enstad2023), rarely have panel studies asked questions about Jews, one reason no study of non-attitudes in opinion toward Jews exists. This study employs numerous waves of the large VSG panels from 2011 through 2020. VSG asks respondents to rate Jews on a 0–100 degree feeling thermometer. To put findings regarding the stability of ratings toward Jews into perspective and to set as benchmarks for comparison, I run parallel analyses of ratings of Muslims, Democrats, and Republicans, also using the feeling thermometer.

Sentiment toward Jews in western nations: evidence from public opinion surveys

There are several ways of estimating the amount of antisemitism toward Jews. One method looks at antisemitic incidents and hate crimes (Enstad, Reference Enstad2023). Depending on the country, both private and government agencies have collected incident and hate crime data. In the United States both the Anti-Defamation League (ADL) and the FBI collect incident and hate crime data. Incidents and hate crimes often receive considerable news coverage, sometimes because of the publicity efforts of the reporting organizations and sometimes because of characteristics of the incidents itself, such as the number of people killed and/or injured, the facility damaged, etc. Yet there are issues with incidents and hate crimes as measures of societal antisemitism. The counts of incidents and hate crimes may be inaccurate because some not reported and affected individuals/institutions may not think that antisemitism was a cause or related to the incident. Further, across nations, there are different reporting requirements and definitions, which affects comparisons. Finally, even in nations that have reported high numbers of incidents and hate crimes, these counts are quite small relative to population size.

A second method for measuring the amount of antisemitism is to ask Jews about their experiences. This method too has its limitations. First, in most western nations, the Jewish population is small, which requires complex survey techniques. Second, some Jews, especially those who are very religious, may refuse to participate in surveys (Weisberg, Reference Weisberg2019), which may skew results. Third, some Jews may perceive an incident as antisemitically motived when some other factor, like race or gender, was the stronger motivation.

However, useful incident/hate crime and Jewish individuals' reports and perceptions of antisemitism are to understanding and tracking the level of antisemitism, public opinion surveys of non-Jews also play a prominent role in estimating the amount of antisemitism in a nation. When similar survey methodologies are employed, including sampling and question wording, public opinion surveys can be used to track antisemitic sentiment over time and across nations.

Public opinion surveys across western nations, such as the United States, Canada, and western Europe, for the past several decades indicate broad-based acceptance and positive attitudes by non-Jews toward Jews. The ADL Global 100 has been surveying attitudes toward Jews using a 10-item scale since 2014. In their 2023 administration, the percentage of respondents which the ADL classifies as antisemitic are: the Netherlands 6%, UK 10%, Germany 12%, France 15%, Belgium 22%, and Spain 26%. The scores for the United States (2015) and Canada (2019) are 9% and 8%, with Australia and New Zealand, surveyed in 2014, both scoring 14%.Footnote ³ The Pew Global Attitudes Project surveys voters, asking their favorability toward Jews in 16 countries in 2019. Among western nations, the percentage favorable, based on adding very and mostly, was 97% for Sweden, 95% for the Netherlands, 94% for France and the United Kingdom, 93% for Germany, 83% for Italy, and 81% for Spain.Footnote ⁴ Finally, the European Values Study (EVS) surveyed respondents in 36 nations since 2017, asking whether they would object to a Jew as a neighbor.Footnote ⁵ Across these nations, 86.6% raised no objection. In Great Britain, 98% voiced no objection, as did 98% in Denmark, 97% Netherlands and Norway, 96% Germany, 93% Switzerland, and 92% Italy.Footnote ⁶ But that the Pew and EVS scores of 81%, and 87% for Russia cast some doubt on these high scores.

Overall, across numerous nations in the West, and even Russia, non-Jews display positive attitudes toward Jews, but there are reasons to question this extremely positive sentiment. This paper investigates the critique that survey responses to non-Jews about Jews may be non-attitudes for many survey participants. Cohen (Reference Cohen2024) argues that Jews are not highly salient to many non-Jews, hence they rely on leadership from trusted sources in thinking about Jews; they do not have well-developed or deeply held opinions about Jews. Cohen hypothesizes that the consensual positivity toward Jews among leaders across western nations may in part account for the highly positive regard of average people toward Jews. To test one assumption of Cohen's thesis, that non-attitudes characterize regard toward Jews for many non-Jews, requires panel data which is done in this paper.

Data: the voter study group panel, 2011–202

Data for this study come from the VSG panel. The Democracy Fund/VSG was formally launched in 2016 by reinterviewing respondents from the YouGov 2011–2012 survey, which was the foundation for the 2012 Cooperative Campaign Analysis Project (CCAP). For purposes here, all the panels are designated as VSG. VSG asked respondents about Jews on five panels, 2011, 2016, 2017, November 2019, and September 2020, employing a feeling thermometer question. Feeling thermometers have been used in prior research to investigate attitudes toward Jews, particularly the American National Election study, which has employed this instrument since the mid-1960s (Cohen, Reference Cohen2018, Reference Cohen2024).

Some argue that feeling thermometers and other explicit questions (e.g., favorability) does not do a good job of measuring attitudes toward groups due to factors like social desirability bias. Research on attitudes toward groups suggests the superiority of implicit questions, like the implicit association test. VSG does not provide an implicit measure of attitudes toward Jews, obviating that measurement strategy here. A recent paper (Lee et al., Reference Lee, Lindquist and Keith Payne2023) finds a very strong relationship between the thermometer rating and implicit measures on attitudes toward Blacks, which may partially alleviate this concern. Further, it is not clear that the explicit/implicit critique is relevant here with its focus on attitude stability and uses attitudes toward other groups to benchmark stability in thermometer ratings for Jews.

Table 1 lists the feeling thermometer ratings for Jews, Muslims, Democrats, and Republicans for the VSG panels by year, and separately for just those who participated in all the panels that asked about the specific group. (The ratings reported toward Jews exclude Jews and toward Muslims exclude Muslims, but the Democrat and Republican ratings include all partisans.) There is considerable panel attrition, perhaps due to the long time from the first to last panel, nearly a decade. Green and Platzman (Reference Green and Platzman2022) show that panel attrition produces little sampling bias in their study of stability of party identification using the VSG data, and the other panels they analyzed.

Table 1.

Mean feeling thermometer ratings of Jews, Moslems, Democrats, and Republicans

Source: Voter Study Group Panels, https://www.voterstudygroup.org/

Measuring stability in ratings of Jews by non-Jews

Following existing research on response stability, this study employs numerous methods to assess stability in ratings toward Jews. These include standard deviations of change from panel-to-panel, correlations of ratings across panels, regression of past panel values on current panel values, regression of past panel values on current panel values using instrument variables, Wiley–Wiley corrections for correlations across panels, and indexing which averages past measures of ratings to predict current ratings. To preview the results, ratings of the two parties are quite stable, those for Muslims are moderately stable, but the ratings for Jews suggest a high degree of instability, a possible indication of non-attitudes in how Jews are rated. For all analyses, unless otherwise indicated, only respondents who participated in all relevant panels are included.

Response variability over time: standard deviations

Table 2 presents standard deviations of the difference in ratings across panels. Like Green and Platzman (Reference Green and Platzman2022), unweighted data are presented because of panel attrition (see Table 1A). The standard deviations, which measure the difference in ratings, appear quite large compared to other studies of opinion stability, such as Green and Platzman's study of party identification. The standard deviations range from 15 to nearly 25. Green and Platzman report much smaller standard deviations, but they analyzed a seven point scale, where the feeling thermometer scale has 101 points, from 0 to 100. Thus, we should expect higher standard deviations. Kustov et al. (Reference Kustov, Laaker and Reller2021) find that the larger the number of response categories, the greater the response instability in their study of attitudes toward immigrants. But the largest number of categories in their study is seven.

Table 2.

Standard deviations of differences in Jewish, Muslim, Democrat, and Republican feeling thermometers across panels, respondents who participated in all panel waves, VSG

Source: Voter Study Group panels, https://www.voterstudygroup.org/

The standard deviations for the parties can be used to benchmark the standard deviations for ratings of Jews and Muslims, since attitudes to the parties should be among the most stable of political orientations. Unfortunately, there is only one panel paring that includes ratings for all four groups, 2019 November–2020 September. We require identical panel ratings to use the Democrat and Republican ratings as a benchmark for the Jewish and Muslim ratings. For that panel pair, the standard deviations for Democrats and Republicans are 15.6 and 17.4, respectively. The standard deviations for Jews and Muslims are only slightly higher, 18.4 and 20.1. From this perspective attitudes toward Jews and Muslims exhibit some instability, but not much more than for the parties.

Correlations across panels

Correlations across panels provide another perspective on attitude stability. In his seminal study, Converse (Reference Converse and Apter1964) used correlations to demonstrate instability in issue positions. Although the size of the correlation matters, the pattern of correlations over time is also consequential. Converse showed that the correlations of issue attitudes at t ₁–t ₃ were higher than t ₁–t ₂ and t ₂–t ₃, indicating random change – correlations from more distant panels were stronger than correlations between bordering waves. Converse's analyses using correlations have been challenged, primarily because of measurement error in items tapping issue positions (Achen, Reference Achen1975; Erikson, Reference Erikson1979; Zaller, Reference Zaller1992). Rather than the respondent being the source of the instability, critics argued that the survey instrument accounted for the correlation pattern Converse noted. This analysis also employs corrections for possible survey instrument measurement error.

Two types of comparisons are possible with these data. First, we can compare the magnitude of the correlations of the Jewish ratings with those for Muslims and the parties. Second, with the numerous panel waves, we compare the correlation magnitude of proximate panels with distant panels. Following Converse, stability with the possibility of directional change is likely when proximate panels have higher correlations than distant panels. But instability is likely when distant panels exhibit higher correlations or if there is no pattern across panels in the magnitude of the correlations.

Tables 3 and 4 present Pearson Product Moments correlations for the feeling thermometer ratings for Jews and Muslims (Table 3), and Democrats and Republicans (Table 4). The correlations between panels for Democrats and Republicans are quite high, as found in research on the stability of party identification. The Democratic correlations range from 0.85 to 0.92, while the Republican ranges from 0.79 to 0.89. The correlations for Muslims fall below that of the parties, ranging from 0.58 to 0.78. But the correlations for ratings of Jews are even lower, ranging from 0.44 to 0.61. Using the parties as a benchmark, rating for Jews exhibit much more instability.

Table 3.

Pearson product moment correlations of ratings of Jews and Muslims between panels, for respondents who participated in all panels

Note: All correlations are significant at the 0.000 level or better. Observation totals are Jews = 2,738 and Muslims = 2,589.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

Table 4.

Pearson product moment correlations of feeling thermometer of Democrats and Republicans between panels, for respondents who participated in all panels

Note: All correlations are significant at the 0.000 level or better. Observation totals are 2,376 for Democrats and 2,361 for Republicans.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

Across panels, the correlations for Jews weaken as the distance between panels widens, but the decline in the size of the correlations sometimes is meager. For instance, the correlations for the first wave, 2011, declines linearly but slightly from 2016 to September 2020, from 0.47 to 0.44. The 2016 correlation also declines from 0.56 to 0.52 in September 2020, again not much of a change. The largest correlation is for the November 2019–September 2020 pair, 0.61, which is the shortest time between panels, about 10 months. The temporal pattern of the correlations for the Jewish thermometer ratings is consistent with patterned change, but the slight changes in the correlations over time do not provide strong evidence to support that hypothesis. Combined with the parties as benchmarks, the correlational analyses suggest instability rather than stability in ratings of Jews.

Regression analysis on lagged values

Regression of the current feeling thermometer value on lagged value from the previous wave provides information in addition to that from the correlational analysis, specifically the regression coefficient and the R ². The regression coefficient tells us how much of a change in the dependent variable is caused statistically by a change in the independent variable, while the R ² measures the amount of variance in the dependent variable that is due to the independent variable(s). The larger the regression coefficient and the R ², the stronger the case for attitude stability. Since all variables employed here use a feeling thermometer, we can compare these quantities across estimations. Like the correlational analysis, the regression analysis assumes no measurement error, a point we return to below. This regression analysis serves as a foundation for assessing the impact of measurement error.

Tables 5–7 present the impact of lagged feeling thermometers on current feeling thermometers for Jews, Muslims, Democrats, and Republicans. The temporal distance from the lagged to current feeling thermometers varies depending upon the panels. Turning first to the regression coefficient, again using the parties as benchmarks, the lagged feeling thermometer coefficients ranges from 0.87 to 0.94 for Democrats and 0.87 to 0.90 for Republicans. These are strong, indicating nearly a one-to-one correspondence between past and current feeling thermometers for the parties. The coefficients for Muslims are slightly lower, ranging from 0.70 to 0.79. But the coefficients are even smaller for the Jewish feeling thermometers, from 0.45 to 0.61. Compared to the parties and Muslims, again attitudes toward Jews appear unstable.

Table 5.

Regression of Jewish feeling thermometers on lagged feeling thermometers

Note: Respondents participated in all panels. Standard errors are in parentheses.

***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

Table 6.

Regression of Muslim feeling thermometers on lagged feeling thermometers

Note: Respondents participated in all panels. Standard errors are in parentheses.

***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

Table 7.

Impact of lagged partisan feeling thermometer

Note: Respondents participated in all panels. Standard errors are in parentheses.

***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

The regression results also provide information on the R ², the amount of variance in the current feeling thermometer that the past feeling thermometer explains. Studies of stability in party identification find R ²s upward of 90%. Here, the R ²s for the Democrat and Republican feeling thermometers range from 0.75 to 0.84 and 0.67 to 0.79, respectively. Although substantial, they are small compared to studies of party identification, probably because party identification uses a seven-point scale where a feeling thermometer with a 0–100 scale is used here. The Muslims R ²s fall below that for the parties, ranging from 0.40 to 0.61. But the feeling R ²s for Jews are much smaller, ranging from 0.22 to 0.37. All the regression results suggest comparative instability in attitudes toward Jews, especially when using the parties as benchmark, and even when compared to attitudes toward Muslims.

Correcting for measurement error: instrumental variable regression

The correlation and regression analyses above assume no measurement error. Critiques of Converse (Reference Converse and Apter1964) argue that he failed to take into account measurement error in the items used to tap respondent attitudes (Achen, Reference Achen1975; Erikson, Reference Erikson1979; Zaller, Reference Zaller1992). Several approaches have been offered to correct for measurement error, such as instrumental variable regression and measurement error methods across panels (Green and Platzman, Reference Green and Platzman2022). This section employs instrumental variables.

The instrumental variable technique derives from Wiley and Wiley (Reference Wiley and Wiley1970). In this model, the measurement and disturbances across panels are assumed to be independent, that is, equal to 0 (Green and Platzman, Reference Green and Platzman2022). The instrumental variable approach helps purge error in measurement for each panel. To implement the instrumental variable approach requires three waves, where the feeling thermometer at t ₋₂ is used to instrument the feeling thermometer at t ₋₁, with the subscripts denoting the relevant panel. The instrumented feeling thermometer is then used as a predictor variable for the feeling thermometer at t ₀. Like the regression analyses from above, the instrumental variable approach produces two quantities of interest, the regression coefficients for the instrumental variable and the R ².

The instrumental variable results are presented in Tables 8–10. In all cases, the regression coefficients for the instrumental variable are much higher than the simple lagged analysis from above. The regression coefficients for Democrats and Republicans range from 0.97 to 1.02 and 1.001 to 1.004, respectively. There is nearly a perfect one-to-one correspondence of the lagged instrument and current feeling thermometer for both parties. The instrumental regression coefficients for Muslims are also quite high, ranging from 0.89–0.99. Although only once approaching the one-to-one correspondence, these coefficients for Muslims are not very far from that mark. For Jews, the instrumental regression coefficients are comparable to those of the parties and Muslims, ranging from 0.90 to 1.01. Some of the instability in the Jewish feeling thermometers in the correlation and regression analyses may be due to measurement error.

Table 8.

Regression of Jewish feeling thermometers on lagged feeling thermometers, instrumental variable regression

Note: Respondents participated in all panels. Standard errors are in parentheses.

*** p < 0.01, ** p < 0.05, * p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/.

Table 9.

Regression of Muslim feeling thermometers on lagged feeling thermometers, instrumental variable regression

Note: Respondents participated in all panels. Standard errors are in parentheses.

*** p < 0.01, ** p < 0.05, * p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/.

Table 10.

Regression of lagged feeling thermometers on Democrat and Republican feeling thermometers, instrumental variable regression

Note: Respondents participated in all panels. Standard errors are in parentheses.

***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

But instability may also arise because of confounding external factors, such as events and leader rhetoric. For example, near the timing of some of the VSG surveys, there were important events related to Jews, in particular the demonstration by Neo-Nazis in Charlotte, Virginia and Donald Trump's subsequent remarks about there being “fine people on both sides,” from which a media frenzy erupted, including public criticism of Trump by some cabinet officials, most notably Gary Cohn, his director of the National Economic Council.

The R ²s of the instrumental variable regressions show similar magnitudes to that of the simple lagged regressions from above. For Democrats and Republicans, the R ²s range from 0.79 to 0.83 and 0.73 to 0.78, respectively, while the range for Muslims is 0.51 to 0.57. In contrast, the R ²s for the Jewish feeling thermometers drops considerably, to 0.14–0.23 from the 0.22–0.37 of the lagged regression analysis. The instrumental variable analysis suggests strong stability in the feeling thermometers for the parties and moderately strong stability for attitudes toward Muslims, but ambiguous results for Jews, with the instrumental coefficients being quite strong but the R ²s relatively small, and weaker than for the simple lagged regression.

Instrumental variables with correlated errors

The instrumental variable analysis in the previous section assumed the errors in measuring ratings for the four groups across panels are independent. This may not be a reasonable assumption. The analysis in this section builds on the instrumental variable approach, adding corrections for correlated errors across panels. Results are presented in Tables 11–14.

Table 11.

Regression of Jewish feeling thermometers on lagged feeling thermometers, regression correction for correlated errors, VSG

Note: Respondents participated in all panels.

Variables: means (mean), variances (var), covariances (cov) for feeling thermometers (ft) for Jews and panel dates for the independent variables.

Standard errors in parentheses, ***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

Table 12.

Regression of Muslim feeling thermometers on lagged feeling thermometers, regression correction for correlated errors, VSG

Note: Respondents participated in all panels.

Variables: means (mean), variances (var), covariances (cov) for feeling thermometers (ft) for Muslims and panel dates for the independent variables.

Standard errors in parentheses, ***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

Table 13.

Regression of democratic feeling thermometers on lagged feeling thermometers, regression correction for correlated errors, VSG

Note: Respondents participated in all panels.

Variables: means (mean), variances (var), covariances (cov) for feeling thermometers (ft) for Democrats (Dem.) and panel dates for the independent variables.

Standard errors in parentheses, ***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

Table 14.

Regression of Republican feeling thermometers on lagged feeling thermometers, regression correction for correlated errors, VSG

Note: Respondents participated in all panels.

Variables: means (mean), variances (var), covariances (cov) for feeling thermometers (ft) for Republicans (Rep.) and panel dates for the independent variables.

Standard errors in parentheses, ***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

These tables present numerous statistics. The ones of primary interest are the feeling thermometers for the most recent past panel, the lagged variable as above (e.g., t ₋₁), the R ², and the covariance between the lag t ₋₁ and the lag t ₋₂. As in the instrumental variable estimation, the lagged variable at t ₋₂ is used to instrument the lagged variable at t ₋₁.

Turning first to the R ², for the two parties they range from 0.84 to 0.86 (Democrats) and 0.78 to 0.82 (Republicans), and 0.61 to 0.62 for Muslims. Again, the R ²s for the Jewish feeling thermometers are smaller, at 0.36, 0.37, and 0.62. Other than the lone 0.62 value, these R ²s are considerably smaller than for the parties and Muslims.

On the tables, the instrumental lagged variable coefficients are in boldface type for easy identification. The coefficients for ratings of Democrats range from 0.57 to 0.59, with ranges from 0.56 to 0.60 for the Republicans. These are much smaller coefficient values than found in the instrumental variable presentation in the previous section. The coefficients for the Muslim ratings range from 0.43 to 0.65, quite a wide range and like the coefficients for the parties, much smaller than the instrumental variable estimation of the previous section. Finally, the coefficients for the Jewish ratings are 0.36, 0.45, and 0.46. Like for the parties and Muslims, these coefficients are smaller than in the last section's instrumental variable regression. And notably, they are smaller than found for the parties and for all but one of the coefficients found for Muslims. The reason for the massive drop in the coefficient values here is the correlation in the errors of the measurement of variables at t ₋₁ and t ₋₂, which are denoted on the tables in boldface italics. All the covariances are large and statistically significant, supporting the idea of correlated errors.

Index creation

Ansolabehere et al. (Reference Ansolabehere, Rodden and Snyder2008) propose a simple and intuitive approach for studying stability in attitudes. Their approach averages lagged values of the same variable across panels.Footnote ⁷ The index is then used to predict current opinion. For this study, we average past measures of attitudes toward Jews (e.g., t ₋₁ to t _−n) and regress attitudes toward Jews at t ₀ on the index. The logic of the indexing approach is that if there is random measurement error, averaging will cancel out these random errors, providing a clearer signal or measurement of the attitude in question. The indexing approach requires at least two previous panels. From the VSG, we have four administrations of the feeling thermometer toward Jews, allowing use of the indexing method. Also, the index approach assumes no measurement error variance. In as much as there is measurement error variance, the indexing approach may understate the degree of attitude stability. But above we have found instability in attitudes toward Jews, especially compared to attitudes toward Muslims and the two parties. Again, we can use the index approach on these three groups to serve as a benchmark for assessing the degree of attitude stability regarding Jews.

To create the indices, I averaged all available lagged feeling thermometer ratings for the four groups. The aim is to use as much prior data as possible in creating the indexes to correct for measurement error. Table 15 presents the results. The regression coefficients for the index for Jews are substantially larger than for the simple lagged regression, 0.82 for the index versus 0.45–0.61 for the simple lagged regressions. But these regression coefficients for the index are much smaller than the coefficients for the instrumental variable regressions, which range from 0.90 to 1.01. The index value for Muslims is 0.88, which is larger than the coefficients for the simple lagged regressions of 0.70–0.79, but on the lower end of the range for the coefficients for the instrumental variable regressions (0.89–0.99). Finally, the index values for Democrats and Republicans show nearly a perfect one-to-one correspondence with values of 1.02 for each party, which are higher than for the simple lagged regressions, which range from about 0.87 to 0.94, but about the same magnitude as the instrumental variable regressions with ranges of 0.97–1.03. The index improves more for Jews than the other groups compared to the simple lagged regressions, while the index performs less well for Jews and Muslims for the instrumental variable regression, at about the same magnitude for the parties.

Table 15.

Impact of regressing final feeling thermometer ratings for Jews, Muslims, Democrats, and Republicans on index based on averaging of prior panel readings, VSG

Note: Respondents participated in all panels. Standard errors are in parentheses, ***p < 0.01, **p < 0.05, *p < 0.1.

Source: Voter Study Group panels, https://www.voterstudygroup.org/

The index R ²s are 0.44, 0.62, 0.87, and 0.83 for Jews, Muslims, Democrats, and Republicans. This is an improvement over the simple lagged regressions for all groups: Jews (0.22–0.37), Muslims (0.40–0.61), Democrats (0.75–0.84), and Republicans (0.67–0.79). The R ²s for the instrumental variable regressions are for Jews 0.14–0.23, Muslims 0.51–0.57, Democrats 0.79–0.83, and Republicans 0.73–0.78. And the R ²s for the correlated error regressions are 0.36, 0.37, and 0.62 for Jews, 0.61–0.62 for Muslims, 0.84–0.86 for Democrats, and 0.78–0.82 for Republicans. In all cases, the R ²s for Jews show the greatest increases (except for the one case for the correlated error regression value of 0.62). Otherwise, the index R ²s show either small increases or comparable values to the other analyses. Most telling is that the R ² index value for Jews is still a modest 0.44, reinforcing the story of weak stability in attitudes toward Jews, with moderate stability for attitudes toward Muslims, and strong stability for attitudes to the parties.