2.1 Trust in science and university research

Previous studies have shown that Americans’ trust in scientists has historically been high, but there are differences in Americans’ trust in science across regional, religious, and partisan lines [20]. More generally, misinformation about science arises from many sources, including individual, group, and societal level factors [10,21]. Past research has identified a variety of factors as being associated with a lack of trust in climate science:

political conservativism [10,20,22–25]; expert cultures and skeptic rationalization [26–28]; and institutional transparency, fairness and public accountability [11,29,30]. Rekker [14] further generalized these factors associated with political polarization regarding science into two interrelated processes: Psychological Science Rejection (PSR) and Ideological Science Rejection (ISR).

This lack of trust in climate science has critical environmental, social and political consequences [11]. It weakens the science-society compact [12] and enables the active resistance of powerful actors with vested interests to change the status quo from which they disproportionately profit [13]. To bridge this trust gap, scholars have studied how to “manage trust” in ways that can strengthen climate action and decision-making. First, trust is a psychological state that is context-specific, involving a trustor and a trustee whose interaction is asymmetric and action-specific [13,29,31]). However, secondly there is no clear understanding of how trust operates at the science-policy interface, which in turn limits the uptake of climate science into policymaking processes [32–35]. Third, it is important to consider maintaining a balance between excessive trust (faith, favoritism, contentment, loyalty) and insufficient trust (skepticism, impartiality, exigency, opportunism) [31].

Cross-national social science research has shown that the public wants scientists to play an active role in the science-policy interface of climate change mitigation that includes advocacy and engagement [4,36]. On a public policy scale, recent evidence shows that public acceptance is a precondition for implementing policies like climate change taxes (carbon taxes) [30]. In this context, factors like distributional fairness, effectiveness and concerns about climate change were found to be strongly associated with public opinion about climate policy [30,37].

Previous research has also shown that knowledge is positively related to climate change opinions, although knowledge alone is not a sufficient precondition to public acceptance of climate policies [38]. Public engagement and accountability by scientists are critical drivers of knowledge translation, further influencing public acceptance of climate research [28,29]. Bergquist et al.’s (2022) meta-analysis found knowledge to be weakly related to public opinion about climate change taxes and laws with no significant differences between subjective and objective knowledge [30]. Also other factors like trust, values and demographic factors (like education, age, income and gender) were not strongly correlated with climate change opinions [11,29,30].

In the case of the United States, opponents and critics of the scientific consensus over anthropogenic climate change have been much more vocal and influential globally. As a result, climate deniers have been successful in confusing public opinion and delaying decisive action [4,39]. However, the origins and motives of the climate change denial movement are highly complex and defined by ideological forces shaped by Republican politics over the last few decades, which bodes ill for climate science and university research [40,41]. For example, a 2016 poll of U.S. adults (n = 1534) found that conservative Republicans are more likely to believe that climate research findings are influenced by scientists’ desire to advance their careers (57%) or their political leanings (54%) most of the time [42]. Moreover, this group is less inclined to anticipate adverse effects from climate change or to judge proposed solutions as making any difference in mitigation.

Additionally, this poll found a stark division in opinions about the causes of global climate change. Nearly 50% of U.S. adults say climate change is due to human activity. A similar percentage says either that the Earth’s warming stems from natural causes or that there is no evidence of warming [42]. This poll also found that most Americans (∼67%) say climate scientists should have a role in policy decisions about climate issues, with ∼23% saying that climate scientists should play only a minor role, and about 9% saying that climate scientists should play no role at all in climate policy [42]. This divide between whether scientists should have a role in policy decisions or not fuels skepticism, with ∼22% of the Americans having no trust or not too much trust in climate scientists [42]. In this paper, we update and extend these earlier studies using recent nationally representative survey data and detailed multivariate analyses.

2.2 Conceptualizing trust in climate science

Climate change is a global challenge and needs an interdisciplinary collaborative approach to understand and address this ‘wicked problem’. Climate science provides a crucial intersectional example for examining public trust in science (and research). The Intergovernmental Panel on Climate Change (IPCC) emphasizes that public confidence in climate science is needed to ensure that the public and governments that receive their mandates from the public implement consensual mitigation and adaptation policies to prevent the predicted devastation from further global warming [29,43].

There is little interdisciplinary agreement on how to conceptualize trust or distrust. Fage-Butler et al., [29] provide a meta-narrative literature review of the interdisciplinary meaning of trust in climate science. The authors argue that there are six key narratives of trust in the climate science literature, attitudinal, cognitive, affective, contextual, communicated, and contingent (that there are six key narratives underscores the lack of consensus about how to conceptualize trust). In our work we focus on one specific type of attitudinal trust, namely survey-based measurement of trust in university research centers. As we describe in detail below, we use a national survey to ask subjects about whether they trust university research centers. Fage-Butler et al. note that this type of attitudinal trust is relational in nature. In their paper they state that “Relational trust, on the other hand is more specific; it is articulated with respect to particular institutions or certain actors as sources of information or as those taking some form of action on climate change.” [29] Thus in our paper we define trust in this attitudinal-relational manner, which is quite similar to what other researchers have called “institutional trust.”

Many scholars have defined trust in this way, and measures of institutional trust like the one we use in this paper have been included in surveys for decades. This has produced an extensive literature that studies attitudinal trust in many different institutions and organizations. There are important social science studies about trust in governmental institutions [44,45], trust in the news media [46], in how governments responded to the COVID-19 pandemic [47], and of course in the many institutions and organizations that work in the climate change and sustainability area [48].

In the institutional trust literature, emphasis is placed on the relationship between institutional performance and institutional trust. [49] This of course makes sense; individuals interact with institutions, hear about the experiences of others with the institutions, and they develop trust (or a lack of trust) based on this information. If the institutional experience is not positive, for example in situations where the institutions are corrupt, those negative experiences can diminish trust in institutions [50,51].

However, we are not studying trust in governmental institutions, or other major organizations like the news media. Rather we are studying trust in scientists and specifically in university research centers, which are unlikely to be institutions that many Americans have a direct experience with. Thus, the evaluations of many Americans towards scientists and university research centers, and their trust in those institutions, may be based less on direct experience but more on heuristics like partisanship, ideology, and religiosity [52–55]. These factors form an important component of our model for trust in university research science centers, as we discuss in the next section.

Data and survey design

The survey data we use in our analysis was collected May 18–24, 2022. The survey was implemented online by YouGov, using respondents from their online opt-in panel to be representative of all U.S registered voters. For this study, the sample definition was self-reported active registered voters in the U.S., who were stratified on age (18–29, 30–44, 45–64, and 65+), race/ethnicity (white non-Hispanic, Black, Hispanic, and other), gender (male, female), education (high school or less, some college, college graduates, and post-graduate), and geographic region (Northeast, Midwest, South, West). The population targets for these strata come from model estimates by YouGov using the 2019 American Community Survey and November 2020 Current Population Survey, conducted by the U.S. Bureau of the Census, and TargetSmart Voter Files.

This sample was weighted to gender, age, race, education, geographic region, and vote in the 2020 U.S. Presidential election. The weights were developed using a raking methodology [56]. Raking is commonly used in survey methodology for estimating weights when the population information is not completely known, usually situations where the univariate population parameters are known but their multivariate distributions are not known. Raking proceeds iteratively: it starts by post-stratification of the sample using one population feature, then the next feature, and so on until the weight variable no longer changes. The weights range from 0.1 to 4.2, with a mean of 1.0 and a standard deviation of 0.4. The implied design effect for unequal weights is 7.7%. With this adjustment, the margin of error for population proportions would be 2.3%.

There are many potential sources of variance and bias that tend to inflate the root mean squared error above the reported margin of error. The sampling frame does not include any voters without access to the internet (representing approximately 7% of the U.S. population), but the sample is balanced on the variables described above to the population of adult voters. Participation in the panel is voluntary and potentially correlated with variables measured in the survey. The estimates are approximately unbiased if sample selection is conditionally independent of survey variables conditional upon the weighting variables and target population estimates are consistent. Similar assumptions are needed for alternative sampling methodologies and interviewing modes.

The survey included a number of questions about climate change. We use three questions from our survey as dependent variables in the analyses that we report in the next section of this paper. The wording of these questions is presented in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The wording of the survey questions used in this study.

Further details about the dependable variables are presented in Table A to Table D in the S1 Text.

https://doi.org/10.1371/journal.pclm.0000147.t001

For our purposes, two of those climate change questions are used in our analysis: how important is climate change as an issue for the U.S. in the next decade, and whether climate change is caused by human activities or is a natural event. The responses to those questions form the dependent variables in two of our models. In the original survey data, the importance of climate change question had four answers: that it is very important, somewhat important, not important, or is not a problem. We recode these four categories into important (very or somewhat important) and not important (not important or not a problem). We recode this dependent variable to be binary for ease of interpretation, in the paper’s S1 Text we present alternative model specifications that analyze the four-category ordinal variable; those models produce results that are more difficult to interpret but which are substantively similar to the results we report in the paper.

The cause of climate change question in our survey had only two responses, whether climate change was caused by human activities or is a natural event. We used this two-response format in our survey to align the survey responses the recent scientific consensus from the IPCC that while climate is always changing due to natural forces, human activities have accelerated this change. For our analysis we use this dichotomy, that climate change is being caused by human activities or by natural events.

The original response distributions for both of these questions from our survey are given in Table A in the S1 Text. As we discuss below, as these two outcome variables are binary responses, we use logistic regression to estimate the model coefficients and standard errors.

The important feature in all our models is trust in university research centers. In our survey, we asked respondents a number of questions about their trust in different institutions, and for this analysis we use the responses to the question asking if they trust in research university centers. The answers to that question provide the dependent variable for another of our statistical models (see Fig 1, further details is presented in the S1 Text).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Distributions of dependent measures (in %, n = 2,096).

https://doi.org/10.1371/journal.pclm.0000147.g001

Empirical models

The three models contain a number of other measures from the survey. First, we include measures of domain specific information (questions that assess the respondent’s knowledge about climate change) and general political information (questions that measure the respondent’s knowledge about American politics). The wording of both our domain specific and general political information questions are in Table 1 and in the Supplementary Information. Finally, our models contain measures controlling for race and ethnicity, gender, educational attainment, region, religious affiliation, frequency of attending religious services, partisanship, ideology, and news sources. Additional information about these variables in Table 1 and the S1 Text.

First, we estimate a logistic regression model for the importance of climate change. The critical hypotheses that we test with this model regard partisan and ideological polarization, as we expect to find that more liberal and Democratic registered voters are more likely to respond that climate change is an important problem for the U.S. The other critical hypothesis is that controlling for partisanship, ideology, and all of the other features in our model, we expect to see that voters with low or moderate trust in university research centers will be less likely to respond that climate change is an important problem, relative to those with high trust in university research centers (we recode the trust responses into these three categories to simplify the analysis).

Second, we study whether our survey respondents respond that climate change is being caused by human action or by natural changes. Again, the key hypotheses regard partisanship and ideology, as we expect to find that those who are ideologically liberal or who are Democratic are more likely to respond that climate change is caused by human activity, than are conservatives or Republicans. The other key hypothesis is that controlling for these political and ideological factors (and the others in the model), those who have low or moderate trust in university research scientists will be less likely to respond that climate change is caused by human actions than are those with high trust.

In our case, the logistic regression model has a functional form (following [57]): (1) where i indexes individuals, j indexes alternatives, X_ijβ represents the linear function of coefficients β and features X_ij in our model, and F represents the logistic transformation. Our models have binary outcomes (so j = 1), and our features (X_ij) represent the set of demographics, identity-related, partisan and ideological, and trust variables discussed above. In particular, we rely on self-reported levels trust in university centers as a measure of relational trust. We ask respondents to report levels of trust towards a generic class of institutions which perform scientific research. One way to interpret this model is that we measure the extent to which self-identified high-trust and low-trust respondents report agreement with the scientific consensus related to climate change. As the logistic regression model is nonlinear, we provide the nonlinear estimates in the paper’s S1 Text and present the estimated average marginal effects below in the paper [57].

The final research question we discuss in the next section regards who trusts university researchers. Establishing this will provide insight into our paper’s closing discussion regarding how climate change scientists might reframe or change their approach to public messaging about the importance and causes of climate change. This model has as a dependent variable response to our survey’s trust in university research centers question: as that question gave respondents the ability to use a 0–10 scale, we will use ordinary least squares regression to estimate the parameters of this model. Here, we estimate the correlates of self-reported levels of trust in university researcher centers directly. This allows us to make inferences related to which demographic variables are correlated with self-reported levels of trust in relation to an important class of scientific institutions.

As we use ordinary least squares regression, the model parameters are easier to depict and interpret (following [58]): (2) where y_i is our outcome variable, i indexes individuals, α is a constant, β_k are coefficients to estimate, the X_ik are the k features in our model, and μ_i is an error term. Ordinary least squares regression estimates can be interpreted directly, and thus we provide in the paper the coefficients and their estimated standard errors in graphical form.

As discussed above, the YouGov provided the survey weights which was used to adjust the sample to make inferences for the national population of American registered voters (these weights are discussed in more detail in the S1 Text. To use these weights in our logistic models and regression analyses, we estimate these models using the ‘survey’ package in R [56,59].

4.1 Is climate change a problem?

We begin with our model that provides results for the registered voter’s opinion that climate change is an important problem. For the purposes of our analysis, we have recorded the four responses to our question into a simple dichotomy, important (very or somewhat important) or not important (not very important or not a problem at all). This allows us to estimate the model using logistic regression, and we report the logistic regression coefficient estimates and confidence intervals in the Supplementary Information (see Fig A and Table E in S1 Text). In order to produce more easily understood estimates, we transform the logistic regression results into average marginal effects, which for this first model are shown in Fig 2. This allows easy comparison of the estimated magnitude of each feature’s correlation with the outcome, controlling for the other features in the model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. How much of a problem is climate change?

logistic regression AME results. Note: Average marginal effects from a logistic regression model and 95% confidence intervals. Model coefficients and standard errors are reported in the SI.

https://doi.org/10.1371/journal.pclm.0000147.g002

Starting with the demographic features in the model, we see a few statistically significant demographic features in this model. The first regards age: relative to those under 30, these results indicate that those aged 45 to 64 and those aged 65 and older are less likely to say that climate change is an important problem. We also see that women are more likely than men to say that climate change is an important problem, though the estimate is of slight magnitude. The other demographic features—race/ethnicity, education, region, religious affiliation and attendance—do not produce results that are statistically distinct from zero.

Next, in these results we see some statistically significant information and information source associations. Those who score higher on our domain-specific environmental information scale are more likely to say that climate change is an important problem for the U.S., while those who score higher on our political information scale are less likely to say that climate change is an important problem; both of these estimates are of relatively small magnitude. Similarly for information sources, the results reported in Fig 2 indicate that those who follow the news on television or print are more likely to believe that climate change is an important issue. We do not obtain statistically significant results regarding those who get their information from radio or online sources.

Next, in this model we see statistically significant and relatively sizeable estimates for the political features: partisanship and ideology. Relative to Republicans, we see that Democrats are more likely to say that climate change is an important problem. And relative to conservatives, those who identify as moderates or liberals are more likely to say that climate change is an important problem in the U.S. These estimates are larger in magnitude that any of the demographic and informational features in the model.

Finally, and perhaps most importantly, at the top of the graph we present the average marginal effects for two of the response categories for our trust in science feature (here the baseline category is high trust). These results indicate that registered voters in our sample who say they have low trust in science are much less likely to respond that climate change is an important problem in the U.S, relative to those who have high trust. Also, those who said that they have moderate levels of trust in scientists are less likely to say that climate change is an important problem in the U.S, relative to those who have high trust. These results are statistically significant, and of sizable magnitude (roughly comparable in magnitude to the estimated average marginal effects for partisanship and ideology).

In conclusion, so far, we have established that trust in science is strongly and statistically associated with whether a registered voter in the U.S. believes that climate change is an important problem in the country. As the model controls for partisanship, ideology, information and information sources, and demographic features that are may be correlated with climate change opinions, this helps to make the case that trust in science has an association with climate change concerns, independent of these other features.

4.2 Is climate change caused by humans or nature?

To examine the association between trust in university scientists and the importance of climate change, and to whether climate change is caused by humans or by natural events, we use a similar methodological approach. We present the logistic regression average marginal effects in Fig 3. The logistic regression coefficients are shown in Fig B and Table E in the S1 Text. Recall that the outcome variable is coded so that if the respondent said that climate change is caused by human activities that takes a value of 1 while if they said it is caused by natural events it takes a value of 0. Thus positive estimates indicate that a change in the feature increases the likelihood that the respondent would say that climate change is caused by human activities, while negative estimates indicate that a change in the feature is associated with an increased likelihood that the respondent would say that natural events are causing climate change.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Is climate change caused by humans or nature? logistic regression AME results.

Note: Average marginal effects from a logistic regression model and 95% confidence intervals. Model coefficients and standard errors are reported in the appendix.

https://doi.org/10.1371/journal.pclm.0000147.g003

As shown in Table A in the S1 Text, 59% of the survey respondents said climate change is caused by human activities. At the same time, 41% attributed it to natural phenomenon. Our logistic regression results show few statistically significant demographic features attributable to race and ethnicity; the one exception is that Black registered voters in our sample are more likely to believe that climate change is being caused by natural events than white registered voters. Those aged 45 to 64 and older than 65 are less likely to say climate change is caused by human activities. We also see Hispanics/Latinos are more likely (with higher estimates) than any other race to say climate change is caused by human activities (see Fig 2). In the same demographic category, we find that those who have at least a postgraduate degree are more likely to support that climate change is caused by human activities. We check for the robustness of this result to choice of functional form in Fig C and Fig D in the S1 Text. We estimate model coefficients for both a Linear Probability Model and Ordered Logit specification, and we find that the results are consistent with those reported in the main text.

In the model, we see that women are more likely than men to say that climate change is caused by human activities. Our results show that religious affiliation is associated with this belief. For instance, Jewish and non-Catholic respondents are more likely to believe climate change has anthropogenic sources. However, respondents who said they have no religious affiliations were significantly more likely to support this belief (see Fig 2, “Rel:None”). In contrast, those who frequently attend religious services are more likely to believe climate change is due to natural causes. Previous research has argued that those who attend religious services tend to believe that climate change is due to natural causes [60–62].

Finally, our findings support existing literature connecting political conservatism and climate change beliefs, as discussed in detail in Section 2. In the model, Democrats, Moderates and Liberals are statistically more likely to say that climate change is caused by human activities. Similiar to the results from section 4.1 regarding trust in science (baseline category is high trust), the average marginal effects show that those who say they have low trust in science are much less likely to respond that climate change is caused by human activities, relative to those who have high trust (see top of graph in Fig 2). Also, those who said that they have moderate levels of trust in scientists are less likely to say that climate change is caused by human activities, relative to those who have high trust.

4.3 Who trusts university research centers?

We use ordinary least squares regression for our third model—examining who trusts university research centers. We show the results from this regression graphically in Fig 4 and in tabular form in Table E in the S1 Text. The outcome variable is coded so that higher values indicate that the subject had higher levels of trust. Regarding the demographics features in the model, we see some evidence that non-whites in our sample (those in the “Other” category) are more trusting of university research centers than registered white voters. We also see that that those aged 45 to 64 and older than 65 are less likely to trust university research centers than those aged less than 30.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Who trusts university research centers? regression results.

Note: Regression coefficient estimates and 90% confidence intervals. Model dependent variable is trust in university research centers, higher scores indicate higher trust.

https://doi.org/10.1371/journal.pclm.0000147.g004

In this model, religiosity and religious denominations show intriguing correlations with trust in university research centers. Beginning with the religious denominations, the baseline denomination in our model specification is Protestant belief. We see that relative to Protestant belief, all of the other denominations (including “none”) are associated with higher levels of trust in university research centers. Perhaps the easier way to interpret this result is that Protestant registered voters are less trusting of university research centers than those of all other denominations, and those who do not have any religious affiliation. While those who said that they frequently attended religious services were not statistically more likely to be trusting of university research centers than those who never attended religious services, the results in Fig 4 indicate that those who often or sometimes attend religious services reported higher levels of trust in university research centers than those who never attend religious services. As for our previous model, we check for the robustness of this result to functional form in Fig E in the S1 Text. We estimate model coefficients for a Linear Probability Model, and we find that the results are consistent with those reported above.

Regarding race and ethnicity, the model results show that both Black and Hispanic/Latinos were less likely to trust university research centers than other race groups in the survey sample. Moreover, in contrast to the results in sections 4.1 and 4.2, we find that women are statistically less likely to trust university research centers. Finally, in terms of education, respondents with a postgraduate education are significantly more likely to have greater trust than those with at most a high school education (see Fig 4).

Regarding partisanship and ideology, Democrats and moderates are more likely to trust university research centers than independents. The regression model results also show that information sources play a critical role in determining trust in university research. For instance, we find that those who are have better domain-specific information about the environment are more likely to trust university research centers, while those with higher levels of general political information are less likely to trust university research centers. Furthermore, those who get their information from print-based and television news sources have been significantly more likely to trust university research centers. Interestingly, we find that respondents who get their information from online news sources were less likely to trust university research centers, though this result is not statistically significant.