1. Framework

A “university” will be defined here as any post-secondary institution that offers instruction in the professions and the liberal arts—excluding vocational schools, arts schools, or theological seminaries.

Our main hypothesis concerns the influence of university on the cultural dimension of party politics. This dimension encompasses a great variety of issues that do not directly affect the pocketbooks of most voters, e.g., sexuality and sexual identity, prostitution, drug addiction, abortion, crime, capital punishment, marriage and divorce, parenting, gender, race, immigration, the nation, religion, science, technology, the environment, civil liberties, and democracy.Footnote ⁴

There are reasons for believing that the experience of attending a university might engender a more liberal (i.e. left-leaning) position on each of the foregoing issues, i.e., more tolerant of deviance, more protective of individual rights, more concerned with public goods, more skeptical of traditional norms and practices, and more cosmopolitan in orientation.

Several possible reasons for this liberalizing effect may be gleaned from the literature. Liberal-leaning professors may inflect course material with their political views—an indoctrination effect (Phelan et al., Reference Phelan, Link, Stueve and Moore1995). Among students, liberal activists may influence fellow students—a peer effect (Sacerdote, Reference Sacerdote2001; Strother et al., Reference Strother, Piston, Golberstein, Gollust and Eisenberg2020). The opportunity to study and reflect, along with cognitive developments occurring in early adulthood, may engender an intellectual dynamic with liberal conclusions—an enlightenment effect (Hyman and Wright, Reference Hyman and Wright1979). Finally, the experience of higher education may enhance feelings of competence and security—a psychological effect (McClosky and Brill, Reference McClosky and Brill1983). By contrast, those who do not attend university are likely to have very different life-experiences on average, which may result in different likelihoods of supporting more liberal parties.

In addition to our main hypothesis, about the role of university education in fostering cultural liberalism, we proposed six subsidiary hypotheses in our pre-registered report. Specifically:

These subsidiary hypotheses are explored briefly in Section 4 below, and more fully in Appendix E.

2. Research design

Our research site—the country of Romania—may be described as middling in economic development and higher education. Like other countries in Central Europe it bears the legacy of Soviet rule, while moving toward Western Europe in recent decades as a member of the European Union. Educational attainment among college-age citizens is nearly at the global mean (Barro and Lee, Reference Barro and Lee2013; see also Appendix D). The higher education curriculum is similar to other countries in Europe, a product of Romania's membership in the European Higher Education Area, which ensures comparability across the EU.Footnote ⁵

With respect to its political system, Romania may be considered a fairly typical case within Central and Eastern Europe (see Economist Intelligence Unit, 2021). Within the context of the OECD, its electoral system, party system size, methods of internal party governance, party finance, media, campaign finance, turnout, and level of partisanship are by no means unusual (Dalton et al., Reference Dalton, Farrell and McAllister2011). As in any newly democratized country, parties are fairly new and volatility is moderately high (though not extreme). Party identification is certainly not as entrenched as in the United States or Britain, and less likely to be perceived along a single left-right spectrum than in most other OECD countries (Dalton et al., Reference Dalton, Farrell and McAllister2011: 87).Footnote ⁶ Nonetheless, parties occupy identifiable ideological niches across several dimensions, as suggested by coding provided by the Chapel Hill expert survey (see Table E4). On this account, one might expect the causal effects registered in this study to be stronger than in two party systems with entrenched political parties but perhaps not as strong as in multi-party systems with weaker partisan attachments.

Thus, along various parameters that might be expected to influence the relationship between higher education and party ideology, Romania appears to be a fairly typical case. However, other features of this case are fairly unique, setting the stage for a natural experiment in which higher education is as-if randomly assigned.

3. Analysis

We use a fuzzy RD design in our main analysis. Since our sample consists solely of students whose bac scores are in the narrow range of 5.8–6.2, variation in this score variable is relatively small. This means it is unnecessary to include multiple polynomial terms in our estimation of the relationship between score and the dependent variable.Footnote ⁹ Furthermore, we do not employ automatic bandwidth selection procedures such as those proposed by Calonico et al. (Reference Calonico, Cattaneo and Farrell2020) since these are unlikely to be appropriate when applied to an already narrowly selected bandwidth of data. Instead, we use all observations within our bac score sampling bandwidth of 5.8–6.2. We employ the continuity-based framework in which it is assumed that the relationship between bac score and both treatment probability and potential outcomes for the dependent variable are continuous away from the cutoff. This allows for the estimation of each of these relationships separately for those scoring below and above 6.0. Assuming that all respondents scoring below 6.0 who attended university would also have done so if they had scored above 6.0 and that all who scored above 6.0 and did not attend university also would not have done so if they scored below this value (the so-called monotonicity assumption), we can estimate the average treatment effect among compliers who score at the bac passage threshold (see Imbens and Lemieux, Reference Imbens and Lemieux2008; Cattaneo et al., Reference Cattaneo, Keele, Titiunik and Vazquez-Bare2016a, Reference Cattaneo, Titiunik and Vazquez-Bare2016b for further discussions of this). An alternative assumption that would justify the same approach is based on assuming independence of potential outcomes and treatment decisions (see Hahn et al., Reference Hahn, Todd and van der Klaauw2001). In our application, we regard this setup as less plausible since people might have a sense for how much university would benefit them, which could influence their decision to attend university.

Another approach would be to rely on the local randomization framework (Cattaneo et al., Reference Cattaneo, Frandsen and Titiunik2015). However, before collecting our data, it was not obvious how wide a bandwidth would be reasonable to use in this framework. Given that there were a limited number of students with bac scores close to the threshold, and given the challenges of locating and contacting these people for our survey, our PAP proposed an approach that would use a continuity-based RD framework allowing for the use of a somewhat wider range of data around the score variable cutoff, but also with some adjustment for potential relationships between the score variable and the dependent variable on either side of the cutoff. In any event, Appendix C presents local randomization analyses as well as multiple robustness checks. The results of these analyses are similar overall to the ones shown in this article.

Our analysis uses the methods developed by Calonico et al. (Reference Calonico, Cattaneo and Titiunik2014), in particular the rdrobust function from the rdrobust R package (Calonico et al., Reference Calonico, Cattaneo and Titiunik2015). We use the default options except as noted below. This approach estimates a linear regression separately on each side of the threshold with weights based on a triangular kernel. Since bac scores are slightly lumpy rather than continuous, we use the approach proposed by follow Lee and Card (Reference Lee and Card2008) and cluster our standard errors by exam score.Footnote ¹⁰

University attendance is our treatment variable and the score variable is respondents' bac scores, with the (fuzzy) treatment threshold occurring at 6.0. In order to obtain cleaner comparisons between treated and untreated respondents, we drop those who graduated high school in 2019 from our analyses since these people would have little or no university experience before our survey was fielded in the fall of that same year. Treatment was then coded as 1 for those with at least some university education and 0 for those with no university education.

A key requirement of a fuzzy RD is that there is a discontinuous increase in the probability of treatment that occurs as the score variable crosses the threshold and that the relationship between the score variable and the probability of treatment is smooth away from the threshold (at least within the bandwidth used). The left panel of Figure 1 shows the proportion of respondents attending university among those having each unique value of the bac score. The jump in treatment probability at the cutoff is clearly dramatic. Respondents with bac scores of 6.0 or higher are much more likely to attend university than are those with scores below this threshold. In fact, 86 percent of those who pass the bac attended university, while only 21 percent of those who did not received this treatment.

Figure 1. Relationship of bac score with treatment and with the cultural dimension of party conflict. Left panel shows proportion attending university among respondents having each unique value of bac score. Right panel shows average score along the cultural dimension of party conflict among respondents having each unique value of bac score. Vertical line denotes (fuzzy) treatment threshold of 6. Horizontal lines on each panel show averages for all respondents above/below threshold. The size of each point is proportional to the number of observations at that bac score.

The second panel of Figure 1 shows the average value of the dependent variable (cultural ideology score) among respondents with each bac score in the sample, revealing evidence of a jump at the threshold. This visual presentation, although intuitive, is informal. Note, however, that because this is a fuzzy RD, the jump in the dependent variable at the cutoff understates the effect of treatment.

Table 1 displays the estimated effect of university attendance on party affiliation based on our formal RD analysis. The estimated effect is 0.54, which is highly significant with a p-value of 0.01. The GALTAN index, in theory, can range from 0 to 10, but among the political parties in our sample it ranges from 1.75 to 7.38 (see Table 2) with a sample standard deviation of 1.88. This estimate represents an increase in support for culturally liberal parties by more than one-quarter of a sample standard deviation when one attends university. The 95 percent confidence interval for this estimate ranges from 0.19 to 1.09 which translates to 11 and 62 percent of a sample standard deviation. This estimate is of the effect of university attendance for compliers—people who would not have attended university if they scored below 6.0 on the bac but who would have attended university if they scored at least a 6.0.Footnote ¹¹

Table 1. RD estimate of university attendance on the cultural dimension of party ideology

Analysis using rdrobust function from rdrobust R package (version 0.99.9) clustering standard errors by bac score, using all of the data in our sampling bandwidth of 5.8–6.2, and otherwise using the function's default arguments. Confidence interval and p-value are based on robust bias-corrected results. N = 1121 (503 obs. below threshold, 618 obs. above threshold).

Table 2. Political parties and their scores on the cultural dimension of party ideology

Party names, number of respondents (N), and score on the GALTAN index (reversed scale: higher numbers refer to a more liberal position) from the Chapel Hill Expert Survey. Note that the UDMR is under-represented in our sample (relative to the country at-large) because our contacting procedure eliminates students who pursue a non-Romanian curriculum; this describes many of the supporters of this ethnically based party, which appeals primarily to Romanians of Hungarian descent.

This result demonstrates that attending university has a powerful effect on ideology in Romania, encouraging those who attend university to support more culturally liberal political parties than non-attenders. Importantly, because the effect estimated from RD designs applies to observations at the threshold (here, people scoring 6 on the bac), we cannot make direct statements about the effect of university on those with different scores. For example, our design does not allow us to estimate the effect of university for those scoring very low or very high on the bac. But the local effect at this passage threshold is likely to be of particular interest in our case. If university education were made either more or less widely available, this would be unlikely to affect those who did very well on the bac or those who did very poorly, because these two groups would likely remain very likely and very unlikely, respectively, to attend university after the change, just as they were before it. Instead, the narrow passers, as well as those with scores close to the passage threshold, are precisely those who are most likely to be affected by an expansion or contraction of the opportunity to attend university, making our estimated effect both theoretically and practically important. (For further discussion, see Section 4.)

A related question that may affect interpretation of our results is how the compliers in our sample, based on whom we estimate our treatment effect, compare to other groups on relevant covariates. Because our fuzzy RD estimates the average effect of treatment for compliers, understanding the characteristics of these compliers helps put our findings into perspective. In Figure B3 (Appendix B), we present the results of analyses following Marbach and Hangartner (Reference Marbach and Hangartner2020), comparing compliers (for which our effects are estimated) to always-takers (those who would have attended university whether they passed or failed the bac) and never-takers (those who would not have attended university whether they passed or failed the bac). We estimate that roughly two-thirds of our sample consists of compliers, meaning that our estimated effect applies to the majority of narrow bac passers. Furthermore, compliers are estimated to have similar average values of several pre-treatment variables including father's education, childhood socio-economic status (SES), and gender. The only covariate showing statistically significant differences between compliers and other types is attending an urban (versus small town or rural) high school. In this case, the estimated proportion urban among always-takers is 72 percent urban, among never-takers is 36 percent, and among compliers is 44 percent. Overall, then, compliers appear to resemble the general sample in most respects with the exception of urban residence.

Table B1 in Appendix B compares the characteristics of our sample to the general population of Romania and the population of those under 24 years old (93 percent of our sample is in this age group). Although our sample is obviously not a representative sample of young Romanian adults (instead it is closer to a random sample of those scoring between 5.8 and 6.2 on the bac in the years we consider), it is still informative to consider how they compare on important characteristics. Our sample is similar to these young adults in terms of the being unmarried, childless, and ethnically Romanian. But our sample is less male, more likely to be employed, and more likely to identify as being in one of the lower social classes. It should be noted, though, that we cannot make these comparisons to the population of those scoring between 5.8 and 6.2 on the bac, but instead look at those in a similar age range to our respondents.

3.1 Robustness checks and RD design validation

We conduct several analyses in Appendix C to assess the robustness of our main finding that university attendance causes an increase in support for culturally liberal parties. First, we repeat the same type of analysis in Table 1, but change the bandwidth around the treatment threshold. This type of analysis can be useful insofar as it provides information about whether RD estimates are sensitive to this arbitrary choice. In our case, we can only reduce the bandwidth used since our data collection strategy focused on contacting people whose scores fall between 5.8 and 6.2. Analyzing several narrower bandwidth choices produces similar estimates to our main results, albeit with the estimates becoming less precise as narrower bandwidths (containing fewer observations) are analyzed.

We also performed “donut hole” robustness checks (Bajari et al., Reference Bajari, Hong, Park and Town2011), changing the bandwidth by removing observations lying within a certain distance of the cutoff. If students could somehow sort themselves very narrowly around the passage threshold these analyses may reduce the bias of RD estimates. In general, these results produced similar estimates to those in Table 1, the exception being when the radius of the “donut hole” is widened to 0.08 or 0.1, discarding a large portion of our data. In these cases the degree of extrapolation is very high the confidence intervals for the treatment effects are so wide as to make point estimates uninformative.

Another common method for evaluating the plausibility of RD assumptions is to choose one or more relevant pre-treatment covariates and substitute them as the dependent variable in the same sort of RD analysis used for the main RD analysis. In Table 3 we estimate the effect of university attendance on several variables: father's education, reported childhood SES, the disciplinary track that students selected for the bac exam, and attending an urban high school. It should be noted that while father's education level and reported childhood SES were measured as part of our survey, both disciplinary track and high school location are drawn from the administrative data about bac results and thus were measured pre-treatment. We conduct RD analyses separately, treating each of these covariates as if it were the dependent variable in a simple local randomization framework (essentially a simple difference in means between those below and above the threshold) and also the continuity-based RD analysis used for our main results in Table 1. All analyses are “intent to treat” (ITT). Using the continuity-based framework paralleling our main analysis above, we find no significant differences in three of the pre-treatment characteristics, but we do find some evidence that the humanities/social science track is more common above rather than below the threshold. Using local randomization analyses, both disciplinary track and father's education show significant differences, although it should be noted that narrow passers report having less educated fathers than narrow failers, which might be regarded as a bias against our findings. Appendix Figure B4 shows RD plots for these pre-treatment covariates, which does indicate that very narrow passing scores include some of the highest proportions of humanities or social science students, but the sample sizes are relatively small in these specific score bins.

Table 3. RD estimates of effect of bac passage on pre-treatment covariates

Difference in means: Local randomization estimates (with 95 percent confidence intervals and p-values underneath) from linear regressions predicting specified covariate with exam passage (i.e. having average bac score of at least 6).

Continuity-based RD: Estimates (with 95 percent confidence intervals and p-values underneath) from continuity-based RD analysis. Results are based on intent-to-treat analyses using linear regressions predicting specified covariate with exam passage (i.e. having average bac score of at least 6), estimating first order (linear) polynomial for score separately on either side of cutoff, estimated and standard errors clustered by exam score, using rdrobust function in rdrobust R package.

Sample sizes listed are the number of valid bac scores (including those below and above the cutoff). Number of responses to father's education and childhood SES questions are slightly different due to a small number of missing values (less than 5 percent for each variable).

Overall, these imbalances could be the result of sorting around the threshold or of differential nonresponse to our survey that is related to the covariates. In order to assess whether our results are influenced by covariate imbalance, we re-estimate the main treatment effect from Table 1 above, this time with covariates for each of the four pre-treatment variables in Table 3 (Calonico et al., Reference Calonico, Cattaneo, Farrell and Titiunik2019). Although this analysis relies more strongly on specific modeling assumptions, it adjusts estimates for possible imbalances in these predictors. These estimates are virtually identical to our main estimates, providing further reassurance that covariate imbalances are not driving the results (Table 4).

Table 4. RD estimate of university attendance on the cultural dimension of party ideology including covariates

Analysis using rdrobust function from rdrobust R package (version 0.99.9) clustering standard errors by bac score, using all of the data in our sampling bandwidth of 5.8–6.2, and otherwise using the function's default arguments. Confidence interval and p-value are based on robust bias-corrected results. N = 1051 (474 obs. below threshold, 577 obs. above threshold).

Finally, we also conduct RD analyses in the local randomization framework as a compliment to the continuity-based analyses used here. The assumptions in the local randomization framework are somewhat different than in the continuity-based framework, meaning that if both sets of results provide similar estimates, we might be more confident in our findings. Formally, in the local randomization approach it is assumed that there exists some bandwidth on either side of the treatment threshold within which we can assume that treatment is as if randomly assigned. This allows the data to be analyzed as in a randomized experiment.Footnote ¹² The local randomization analyses generally produce positive and highly significant estimates for university attendance's impact on political liberalism. These estimates are quite similar to our main result above, although when using very small bandwidths these estimates become insignificant, albeit with similar size to the other estimates (see Appendix Table C3). Estimates ranged from 0.45 to 0.61 depending on the bandwidth used, with an estimate of 0.51 when using all of our data (bandwidth of 5.8 to 6.2 as in the main analyses above). This is quite close to the effect of 0.54 estimated in our pre-registered main continuity-based analysis in Table 1 above. The fact that the overall substantive findings are similar under these two different approaches bolsters our confidence in the results. Appendix C provides more information about these local randomization estimates.

Overall, the main finding—that university attendance causes people to be more politically liberal—appears quite robust.

4. Subsidiary hypotheses

In addition to our main hypothesis, we set forth a number of subsidiary hypotheses, listed briefly at the end of Section 1. In this section, we review our findings. Details about the designs, as well as tables and figures showing the statistical tests, are contained in Appendix E. Importantly, most of these hypotheses require subsetting our main sample, which entails a significant loss of power. These results are therefore regarded as highly speculative.

First, there is no evidence that longer attendance (measured in years since matriculation) at university leads to stronger causal effects (Figure E1). At the same time, we note that estimating the effect separately among each of five graduation year cohorts produces imprecise results, so this should not be viewed as a strong null finding.

Second, we do not find that university students from rural backgrounds are more susceptible to the liberalizing effect of college than students from urban backgrounds. In fact, it may be the reverse—though the estimate is not terribly precise and may be stochastic (Table E1).

Third, the liberalizing effect of university education does not appear to differ by intended university concentration. No differences are found when we compare the impact of university on students majoring in (a) humanities and social sciences or (b) natural sciences (Table E2).

Fourth, attending university sharply increases the probability that a student will support a different political party that than of his or her father. The estimated effect of 0.22 (for a binary dependent variable with a sample mean of 0.70) is large and highly significant (Table E3).

Finally, there is no evidence that attendance at university exerts a liberalizing effect on party ideology along economic or left-right dimensions. Indeed, there may be a modest conservative effect along these (highly correlated) dimensions of party ideology (Tables E5–E6), which would run counter to our expectations, although these results are not statistically significant in our main specifications.

We surmise that on economic issues (e.g., taxes, spending, redistribution, regulation) students are likely to encounter a mixture of left- and right-wing views in the scholarly literature, the predispositions of college faculty, and the content of course curricula. Consequently, students presumably receive mixed signals during their tenure at university. Additionally, attendance at a university is likely to enhance students' long-term earning power and class status. Consequently, they may view government intervention—and especially policies with a redistributive goal—with greater skepticism.