Does College Influence Sociopolitical Attitudes?

The Effects of College on Sociopolitical Attitudes

Sociologists, psychologists, and educators have long argued for the effects of college on political orientation, support for civil liberties, and egalitarian gender-role beliefs. Researchers posit different mechanisms to account for this association, but most scholars claim that college attendance socializes students to adopt more left-leaning attitudes. Generally speaking, they suggest this effect occurs primarily because of interactions with peers and student social norms rather than through classroom experiences.

Scholars have documented a strong association between education and liberal political orientation. Gross and Fosse (2012) argue that this is a major reason why college professors are themselves liberal, as they have spent far more time earning college degrees than the rest of the population. Popular media accounts suggest that college professors then inculcate liberalism in their students, reproducing political beliefs in the classroom (Gross and Fosse 2012; Mariani and Hewitt 2008). Dey (1996, 1997) finds that individuals attending more liberal academic institutions become more politically liberal, but he suggests this is related to a broader institutional context rather than student–faculty interaction. And Mariani and Hewitt (2008) find that faculty political attitudes have no effect on changes in student political attitudes. The evidence that institutional context—rather than faculty interaction—changes political values extends to conservatism as well; Dey (1997) finds that students who enter more conservative academic institutions also become more politically conservative.

Prior research has also found associations between college attendance and increased support for civil liberties, net of other demographic and background characteristics (Kingston et al. 2003). Bobo and Licari (1989) argue that college changes attitudes toward civil liberties because it increases cognitive sophistication (see also Ohlander, Batalova, and Treas 2005). However, there is evidence that learning about civil liberties fails to increase support for them (Green et al. 2011). Phelan and colleagues (1995) conclude instead that college socializes individuals to adopt the “official culture” of the United States: all individuals deserve “equal respect” but not “equal outcomes” (see also Finney 1974).

Considerable evidence shows an association between college education and greater support for gender egalitarianism and less support for traditional gender-role restrictions, and this relationship remains consistent over time (Bolzendahl and Myers 2004; Cunningham 2008; Harris and Firestone 1998; Kingston et al. 2003). However, this relationship depends on the type of peers a person socializes with. Bryant (2003) finds that men who socialize with other men or with people from conservative religious upbringings espouse more traditional gender-role beliefs. In contrast, talking about politics with peers weakens gender-role stereotyping.

Three potential mechanisms could help explain the relationship between college education and sociopolitical attitudes. First, Bowman (2013) argues that college provides an opportunity to interact with individuals from different social backgrounds, and as a result, students express more comfort with out-groups and a greater desire to learn about other cultures. This theoretical explanation is consistent with social psychological studies of intergroup bias (Tadmor et al. 2012), and college campuses do sometimes offer a diverse and multicultural student body to promote out-group interaction and support for new sociopolitical attitudes. Second, college campuses provide “free spaces” to develop alternative ideologies; these free spaces have been instrumental in developing left-wing and libertarian approaches to individual liberties as well as feminist thought (Klatch 1999; Polletta 1999, 2004). Participation in free spaces also provides ample opportunity to engage in political discussion. Finally, Dey (1997) argues that students directly socialize each other to adopt new sociopolitical attitudes. These proposed mechanisms are likely intertwined; for example, a set of individuals who participate in a free space may later directly socialize new students to adopt new sociopolitical beliefs.

In summary, the college-effects model posits that individuals who attend college are constantly exposed to interactions with a diverse set of individuals (Moffatt 1989). These interactions take the form of informal conversations and learning about different viewpoints (Bowman 2013), participation in activist free spaces (Polletta 1999), and direct socialization from peers (Dey 1997). As a result, students are likely to become more liberal (Dey 1996, 1997), express greater support for civil liberties (Bobo and Licari 1989; Finney 1974; Phelan et al. 1995), and support loosening gender-role restrictions (Bolzendahl and Myers 2004; Cunningham 2008; Harris and Firestone 1998; Kingston et al. 2003).

The Effects of Family on Sociopolitical Attitudes

Substantial evidence shows that college affects sociopolitical attitudes, but family influences play a large role in determining who attends and completes college (Blau and Duncan 1967; Conley 2001). Prior scholarship often accounts for the effect of measurable family influences on sociopolitical attitudes by controlling for observed parent characteristics. Kingston and colleagues (2003), for example, account for the role of parental socioeconomic status to estimate the effect of college on social outcomes, net of family background. But family socioeconomic status is not the only way that parents influence children; in particular, existing research emphasizes how parents socialize their children in ways that do not map neatly onto class-based variables.

Overall, there is substantial evidence showing that parents are able to instill their political beliefs and gender-role attitudes in their children, net of socioeconomic status (Acock and Bengtson 1978; Dalhouse and Frideres 1996; Glass, Bengtson, and Dunham 1986). Jennings, Stoker, and Bowers (2009) find that parents directly socialize their children to adopt certain political beliefs, and this effect is more pronounced in families with strong political viewpoints. Braungart (1971) finds that parents directly socialize their children’s political beliefs and steer their children toward specific political alignments by providing political discussions and modeling authoritarian or democratic values. In addition, Glass and colleagues (1986) find that while family socioeconomic status exerts a strong effect on children’s gender-role ideology, parental socialization has a statistically significant effect into adulthood.

In other words, childhood family influences have an effect on sociopolitical attitudes beyond parental socioeconomic status; this is problematic because these influences can also affect whether a person attends and completes college. Davis and Pearce (2007) find that children with more gender-egalitarian attitudes also have higher educational expectations. Furthermore, Elchardus and Spruyt (2009) find that college students tend to select into academic majors where they feel most politically comfortable; this implies that the relationship between college education and political attitudes is a reflection of childhood influences. Additionally, some evidence suggests that the effect of education on civic participation is spurious owing to family background (Highton 2009; Persson 2014; Schnittker and Behrman 2012). Specifically, Schnittker and Behrman (2012) provide a more stringent test of educational effects by using a fixed-effects design, which can remove the effect of shared family background on a dependent variable (see also Halaby 2004). They find that the effect of education on multiple social outcomes is overestimated or nonexistent once variation associated with family background is removed from the model. They also find evidence that some prosocial behaviors are reduced by college attendance, suggesting that college may socialize students in ways that make them more attuned to their own interests rather than others.

We know of one study that investigates the effect of education on various social and political attitudes after accounting for a wide range of family influences. Sieben and de Graaf (2004) use a sample of Dutch siblings to investigate the impact of shared family background and educational attainment on religious beliefs, several different measures of political orientation, postmaterialistic values, and gender-role traditionalism. After removing the effects of shared family background, they find that education has no effect on religious beliefs or political orientation and a much smaller effect on postmaterialistic values and gender-role traditionalism. The applicability of these findings to the United States is unclear because of dissimilarities between the U.S. and Dutch education systems, the uniqueness of U.S. political ideologies, and differences in gender stratification in the two countries. In the U.S. context, it seems reasonable to expect some effects of college on sociopolitical attitudes to remain statistically significant, but effect sizes should shrink and others may lose statistical significance after better accounting for family and social background characteristics.

In summary, researchers have traditionally attempted to separate college effects from social background effects through the use of controls in regression models. Because of data limitations, researchers can usually control only for some combination of parent education, family income, and parent religious beliefs (e.g., Kingston et al. 2003). Scholars who draw on the spurious model (e.g., Schnittker and Behrman 2012) suggest that this method of identifying college effects is problematic because it is unlikely a limited set of controls can fully account for social background characteristics. Parents influence the educational attainment and sociopolitical attitudes of their children in complex ways that are difficult to observe, operationalize, and measure. Due to these complex processes, some scholars have expressed concerns that observed college effects are biased by unobserved social background effects. In fact, more recent studies that use matching or within-family designs find more modest or even null college effects on an array of outcomes.

In this study, we disentangle background effects from college effects by basing model estimation on variation within sibling pairs. In doing so, we are able to offer a unique test of the effect of college on sociopolitical attitudes. Specifically, if college influences sociopolitical attitudes net of social background characteristics (as predicted by the college effects model; see Hout 2012; Kingston et al. 2003), then we should expect to observe college effects among siblings with different levels of education. Conversely, if college effects are confounded by social background characteristics (as predicted by the spurious model; see Schnittker and Behrman 2012), then we should expect to observe no statistically significant differences in sociopolitical attitudes among sibling pairs with disparate education levels.

Data

We draw data from the 1994 General Social Survey (GSS; Smith et al. 2011) and the Study of American Families (SAF; Hauser and Mare 1994) to test the relationship between college and sociopolitical attitudes, net of family background. The GSS, conducted by the National Opinion Research Center at the University of Chicago, was administered to a nationally representative sample of non-institutionalized adults and included a variety of items on sociopolitical attitudes and respondent demographics. The SAF is a companion data set to the 1994 GSS, enabling researchers to test for similarities within families. The GSS first collected identifying information for one randomly selected sibling for each of the 2,992 respondents who had at least one sibling above the age of 25. SAF researchers then conducted telephone interviews with 1,155 siblings, asking a subset of the questions in the GSS. By combining SAF and GSS data, we were able to construct a data set that contains data on 2,310 individuals and 1,155 sibling pairs.

Of the original GSS sample, 2,663 respondents had an eligible sibling, but only 43 percent of the selected siblings were interviewed. The major source of nonresponse was the inability of GSS interviewers to collect sufficient information for SAF researchers to locate siblings for interviews. Although selective nonresponse is a potential source of bias and could limit the generalizability of the findings, we believe these data are representative of sibling pairs in the general population. We compared our analytic sample to eligible GSS respondents to test for differences. We present these findings and additional analyses in an online supplement. Overall, as shown in Table S1 of the online supplement, our analytic sample and the eligible GSS sample are similar on most characteristics. However, the analytic sample does have more white respondents, more respondents with a college degree, and more educated mothers. Respondents in the analytic sample were also less likely to have lived outside of the United States during adolescence. These differences could limit the generalizability of our findings.

Goldstein and Warren (2000), however, find consistent results across the GSS-only sample and the GSS-SAF combined sample, which suggests that nonresponse bias is not a serious issue. Following others who have used these data to examine within-family differences (Freese, Powell, and Steelman 1999; Warren 2001), we note the low response rate as a potential limitation and add that the present analyses should be interpreted cautiously for more disadvantaged populations. That said, we do not believe this biases our findings.

We impose two restrictions on our sample to increase the likelihood that siblings experienced similar social backgrounds when growing up. First, we exclude siblings who do not share the same mother and father (6 percent of sample). Second, we exclude siblings who did not grow up together, measured as living in the same household at age 16 (11 percent of sample). We also exclude sample members who are under the age of 25 and report that they are still in school (less than 1 percent of the sample). Additionally, for reasons further explained in the discussion of our measure of college attendance, we also exclude sibling pairs where one sibling completed an associate’s degree or some college (14 percent of sample). Before accounting for missing data, the analytic sample contains 1,652 individuals and 826 sibling pairs. We lose an additional seven sibling pairs to missing data on independent variables. The GSS uses a split-ballot survey design, which means not all respondents were asked all questions. As a result, the analytic sample size varies by outcome.

The combined GSS-SAF sample has one additional limitation: it was conducted in 1994. Close to half of the participants were college graduation age in the 1960s and 1970s, about one fifth of respondents were college graduation age prior to the 1960s, and about one third were college graduation age in the 1980s and early 1990s. As a result, the findings roughly reflect the effect of college education from the end of World War II until the Republican revolution that elected Newt Gingrich as speaker of the House. Since then, college costs have risen, the rate of college attendance has increased, and universities have become more oriented toward research and grant funding (Brennan 2008; Brint 2005; Horowitz 2015; Slaughter 1985; Slaughter and Rhoades 1996). Thus, the effects of college we report here are limited to a particular time frame and may have changed as the college experience itself has changed.

Nonetheless, the GSS-SAF data are unusual because they are nationally representative, include a wide variety of questions on sociopolitical attitudes, and include matching sibling data. Data that allow us to run fixed-effects sibling models, a powerful way to address family-invariant spuriousness (Halaby 2004), are rare. Although the age of the data is a limitation, these data permit a strong analytic design that more recent data sets cannot match.

Measures

Demographic and background variables

In our analysis, we estimate random-effects and sibling fixed-effects models. We first estimate random-effects models, using control variables for covariate adjustment. This allows us to compare the random-effects estimates to the sibling fixed-effects estimates. Specifically, we include controls for sex (1 = female, 0 = male), race/ethnicity (1 = nonwhite, 0 = white), a continuous measure of age, region of residence during adolescence (0 = South, 1 = Northeast, 2 = Midwest, 3 = West, 4 = foreign), religion during adolescence (0 = Protestant, 1 = Catholic, 2 = Jewish, 3 = none, 4 = other), mother’s education (0 = high school graduate, 1 = did not complete high school, 2 = some college, 3 = college graduate), and whether the respondent is the firstborn sibling (1 = firstborn, 0 = other). Race/ethnicity is limited to a reference variable because the GSS did not collect a more nuanced measure or include a large enough sample of nonwhites. Table 1 presents summary statistics for all variables.

OPEN IN VIEWER

Table 1.

Descriptive Statistics for All Variables.

			GSS		SAF
Variable	Min.	Max.	M	SE	M	SE
Liberal political ideology	0	1	0.266		0.261
Seven-item political ideology scale	1	7	3.77	1.4	3.71	1.41
Support for civil liberties	0	15	10.39	4.68	11.41	4.28
Agree men are better emotionally suited for politics	0	1	0.186		0.299
Agree men should achieve outside of home and women should take care of home	0	1	0.346		0.376
Agree wife should put husband’s career first	0	1	0.207		0.184
College completion	0	1	0.318		0.321
Female	0	1	0.534		0.514
Firstborn	0	1	0.276		0.473
Age	18	93	45.32	15.5	45.07	15.51
Nonwhite	0	1	0.088
Region of residence during adolescence
Raised in South	0	1	0.303
Raised in Northeast	0	1	0.202
Raised in Midwest	0	1	0.329
Raised in West	0	1	0.151
Raised outside of United States	0	1	0.171
Religion during adolescence
Raised Protestant	0	1	0.633
Raised Catholic	0	1	0.294
Raised Jewish	0	1	0.02
Raised without religion	0	1	0.04
Raised with other religion	0	1	0.02
Mother’s educational attainment
Mother did not complete high school	0	1	0.315
Mother completed high school	0	1	0.504
Mother completed some college	0	1	0.03
Mother completed bachelor’s degree	0	1	0.149
N			819		819

Note: GSS = General Social Survey; SAF = Study of American Families. Because the sample is restricted to siblings who reported growing up together and share the same mother and father, there is no variation across the GSS and SAF samples for region of residence during adolescence, religion during adolescence, mother’s education, or race/ethnicity.

Analytic Strategy

Our analysis contains two stages. In the first stage, we use random-effects models to estimate the effect of college completion on sociopolitical attitudes, net of observed social background characteristics and clustering by family unit. Thus, the first part of the analysis largely replicates past research. The basic model can be expressed as follows:

Y i f = α X i f + >β Z i f + ∈ f + ∈ i f

where Y_if is the probability of holding a given attitude (e.g., identifying with a liberal political ideology) for individual i in family f. X_if represents whether respondent i in family f completed college, Z _if is a vector of observed control variables, ϵ _f is a family-specific error term, and ϵ _if is a random error term. The coefficient for X_if (α) is the effect of college completion on attitudes net of observed control variables. However, the college effect will be biased if there is a correlation between college completion (X_if) and unobserved social background characteristics (ϵ _f ). To address this shortcoming, we estimate sibling fixed-effects models—the second stage of our analysis. The basic model can be expressed as follows:

( Y 1 f – Y 2 f ) = α ( X 1 f – X 2 f ) + β ( Z 1 f – Z 2 f ) + ∈ 1 f – ∈ 2 f

where the subscript 1 stands for the first sibling from family f and the subscript 2 stands for the second sibling from family f. The primary difference between Equation 1 and Equation 2 is the removal of ϵ _f . Because ϵ _f , the family-specific error term, is the same for Sibling 1 and Sibling 2, all background characteristics shared by the siblings are removed from the model. Sibling differences in the outcome variable are thus related to sibling differences in college completion and other sibling-variant covariates, but all unobserved family background characteristics that are shared by siblings are removed from the model (see also Halaby 2004).

A variety of studies have used sibling fixed effects to isolate the effect of interest from family background effects (see Campbell 2015; Currie and Thomas 1995; Geronimus and Korenman 1992; Guo and VanWey 1999); however, sibling fixed effects have several limitations. First, siblings are seldom the same age and thus may experience different social contexts. For example, family income or parenting styles may change over time. If these differences are correlated with college attendance, this could bias the college effect. As a robustness check, we conducted a reanalysis where we limit the sample to siblings who are close in age, increasing the likelihood that siblings were exposed to similar environments and experiences while growing up. Findings from these analyses are consistent with the analyses that do not impose an age similarity restriction on the sample (see Tables S3, S4, and S5 in the online supplement).

Second, there may be differences among siblings that are correlated with college completion. For example, a preadult difference in level of intelligence or having different sociopolitical interests could make one sibling more likely to attend college and also to hold different sociopolitical attitudes. This has the potential to bias the college effect. We cannot completely address this issue because our data are cross-sectional, and so postcollege differences may be due to either preadult differences or college effects. That said, because both intelligence and social interests tend to cluster within families, it is likely that much of this variation is removed by our design.

Third, sibling fixed effects cannot be estimated for individuals who do not have a sibling. If the effect of college is different for people without a sibling, the sibling fixed-effects models may over- or underestimate population-level differences in sociopolitical attitudes by college completion. This limits the overall generalizability of our findings.

Additionally, because many of the outcomes of interest are nonlinear and discrete variables, some fixed-effects strategies are not viable. Consequently, we estimate conditional fixed-effects models and linear probability models. The conditional fixed-effects models estimate logistic fixed effects using conditional likelihood functions. The linear probability model eases interpretation and has more statistical power, but it produces predicted probabilities that are theoretically not constrained between 0 and 1 and violates homoscedasticity assumptions. We estimated both models and found that substantively it makes little difference which models are used, as both models produce similar results. We present results from the linear probability models in the main text here, and results from the conditional fixed effects are available in Tables S10 and S11 in the online supplement.