2.1 Choosing Latin as an immediate strategy of cultural distinction

Building on Bourdieu’s theory of social class (Bourdieu 1984), the extent to which a certain behavior can be considered a “distinctive” cultural practice depends on three factors: its exclusiveness, the concealment of its utility, and its historically established symbolic meaning. We argue that learning Latin from grade five onwards matches all these characteristics and can therefore be considered a strategy of cultural distinction.

Exclusiveness: Learning Latin in the German educational system is only possible if students decide to attend the Gymnasium, the highest secondary track in Germany. At the Gymnasium students can decide to learn Latin from either grade five or grade six onwards, or to learn French or another modern language instead of Latin. In 2017, about 6 % of all students attending the Gymnasium learned Latin from grade five onwards, whereas about 30 % learned Latin from grade six onwards. Learning French was the most popular alternative, with more than 40 % of all students (Federal Statistical Office 2017). Hence, whereas learning French and Latin from grade six onwards is not exclusive at all, learning Latin from grade five onwards is.

Apparent uselessness: The most obvious function of learning foreign languages is the enhancement of one’s communicational potential to communicate with people who speak another mother tongue.^[4] Foreign languages increase the capability to interact with people in other linguistic regions of the world. Whereas English is in this respect the most useful language, followed by other modern languages such as Spanish, French, or Chinese, there is no obvious communicational benefit of learning Latin. Hence, Latin does not fulfill any direct function. Apparent uselessness is precisely one of the features that Bourdieu considers to be constitutive for a strategy of distinction which is typically practiced by academic social classes.^[5]

Historically established symbolic meaning:^[6] In order to understand the reputation attributed to people who have learned Latin, one must understand the role Latin played in the history of the German educational system. For centuries, Latin served as the European lingua franca of the upper classes. During the development of the German school system in the 19th century, learning Latin and ancient Greek became part of Humboldt’s educational ideals (Fuhrmann 2002; v. Friedeburg 1992). Traditionally, 40 % of all schooling hours at the Gymnasium in Germany were spent on learning Latin and ancient Greek. Both languages were associated with the proclamation of an autonomous educational ideal. The humanist educational ideal was not only associated with the ideal of education for its own sake, but also with a highly exclusive and socially selective access to the Abitur (Herrlitz et al. 1993: 38; Sawert 2019). Even though Humboldt’s Gymnasium gradually lost its importance because of educational expansion, the historical idea that learning Latin is something special and elite has remained. Whereas learning modern foreign languages is associated with an orientation towards acquiring educational knowledge to prepare for the labor market, acquiring classical knowledge is associated with an ideal of autonomous “elite” education (Sawert 2021).

Compared to other languages, learning Latin, especially from the fifth grade onwards, fulfils all requirements of a distinctive cultural practice: It is exclusive, its utility is not obvious, and it has a historically established meaning as “elite” education. Following Bourdieu, we expect privileged families and especially traditional academic families, in which not only parents but also grandparents have an academic degree, to follow an educational strategy directed towards cultural distinction.

Several empirical studies have shown that being from an educationally privileged family increases the probability of practicing distinctive educational strategies or distinctive leisure activities (e. g., Reimer & Pollak 2010; Kraaykamp 2003). More recent empirical studies have focused on the effect of the grandparents’ educational background on the grandchildren’s education (e. g., Sheppard & Monden 2018; Hällsten & Pfeffer 2017; Møllegaard & Jæger 2015) and grandchildren’s leisure activities (Sawert & Gerhards 2019). These studies claim that the grandparents’ cultural capital impacts on the acquisition of the cultural capital of the parents, which they again transmit to their children. Hence, whereas parents who originate from a household in which they experienced a distinctive cultural lifestyle develop such a taste for it themselves, parents who did not experience such a lifestyle have a lower probability of being oriented towards distinctive cultural practices even if they experienced educational mobility. Consequently, we assume that historically academic families (those in which parents and grandparents hold academic degrees) are more likely to prefer distinctive educational tracks such as learning Latin from grade five onwards.

2.2 Choosing a socially exclusive learning environment

The second mechanism of social closure involves learning Latin as an unintended consequence of the choice of a specific learning environment. The argument proceeds in three steps: First, it explains that learning environments differ between humanist Gymnasiums on the one hand and other Gymnasiums on the other. Second, it states that privileged classes more often prefer the learning environment of the humanist Gymnasium than lower strata. Finally, it requires that the selection of a learning environment dictates the choice of the language.

Socially exclusive learning environment: Baumert and colleagues (2010) show that humanist Gymnasiums in Germany are composed of socially selective student bodies in terms of academic performance of the students, and with respect to the social status and the educational level of the parents. This in turn has a twofold impact on pupils’ development (Baumert et al. 2006; Rindermann 2007; Bellin 2009). First, the social composition of the student body is important for the interactions within the class and impacts on students’ individual learning performance (Hoxby 2000; Schneeweis & Winter-Ebmer 2007). As children from higher educational classes on average perform better at school than students from lower educational classes (Jaksztat 2014), a socially selective school environment should provide a more efficient learning environment, hence, leading to a better skill development. Second, as pupils interact with their peers in school, the school environment might influence the development of normative and cultural orientations. The higher the educational level of a school environment, the more prevalent the culture of the upper-middle classes.

Preferences for a socially exclusive learning environment: There are four reasons why social strata might differ in their preferences for a socially selective learning environment. First, the differences in the learning environment can only affect the school choice if the family knows about these differences. It seems plausible to assume that this knowledge is more widespread among educationally privileged families than among less privileged families. Second, the cultural orientations prevalent in the humanist Gymnasium are more attractive to families with similar cultural orientations. Third, upper classes are more reliant on a good learning environment since they require educational success to maintain their status (Keller & Zavalloni 1964; Boudon 1974). Finally, educational expansion puts educationally privileged families under pressure to look for new ways to transmit their status to the next generation as an upper secondary certificate is no longer distinctive enough (Erikson & Jonsson 1996; Breen & Goldthorpe 1997).

Correlation between a socially exclusive learning environment and Latin as a mandatory language: If the school with the preferred learning environment requires students to learn Latin from grade five onwards, the language choice is determined by the choice of the learning environment. More generally, if the feasible set of language options correlates with the learning environment, the language choice will be affected by the choice of the learning environment. In practice, this is indeed the case, particularly for learning Latin from grade five onwards. Whereas most schools in Germany offer Latin and French from grade six onwards, the option to learn Latin from grade five onwards is only offered at humanist Gymnasiums.

To sum up our argument at this point: Educationally privileged families are more likely to choose Latin from grade five onwards than less privileged families due to their stronger preference for a socially selective school environment.

2.3 Secondary instrumental function of learning Latin

The third possible mechanism builds upon the idea that learning Latin may have a “secondary instrumental function”. Building on transfer theories of learning (e. g. Greeno et al. 1993; Singley & Anderson 1985), several authors assume that learning Latin increases cognitive abilities in other fields, particularly students’ linguistic skills in their mother tongue, their ability to learn Romance languages, and their ability to think logically. Although empirical studies do not support such hypotheses (Thorndike 1923; Haag & Stern 2000, 2003), the belief in positive transfer effects remains quite popular, as we have shown in a recently published paper, analyzing the ascription of positive transfer effects by parents of pupils attending the German Gymnasium (Gerhards, Sawert & Kohler 2019). Of course, families who believe in the positive transfer effects of learning Latin should be more likely to choose Latin at school than those who do not. The same is true for families whose offspring strives for occupations for which Latin skills are useful or necessary. Medical doctors, lawyers, priests, and historians are typical examples of such occupations. Using the University of Heidelberg as an example, we have listed elsewhere for which BA and MA programs knowledge of Latin is necessary or recommended (Sawert 2018: 79). As we have shown elsewhere (Gerhards, Sawert & Kohler 2019), the belief that learning Latin results in positive transfer effects is more prevalent among higher educated families than among lower educated families. Hence, it seems plausible to argue that expectations about a secondary instrumental function of Latin might mediate an effect between educational background and learning Latin from grade five.

2.4 Spatial proximity between school location and socially privileged neighborhoods

The fourth mechanism linking social origin to language choice stems from the geographic location of families on the one hand, and the location of schools offering Latin from grade five onwards on the other hand. An important motive for the families’ school choice is the sheer proximity of a school to the family’s place of residence. The closer the school, the less time must be invested in transit, and the more time can be spent on other things. Starting from the assumption that humanist Gymnasiums are located closer to the neighborhoods of privileged social strata, choosing Latin becomes an unintended consequence of choosing the nearest school. Unfortunately, there are no representative data available to check to which degree humanist Gymnasiums are located in privileged neighborhoods across Germany. For Berlin, Sawert (2018) shows that the average rent of the neighborhoods in which humanist Gymnasiums are located is above the average rent in Berlin. In order to avoid biased estimates, we designed our study in a way to make sure that we can control for this mechanism to a reasonable extent.

2.5 Hypotheses

The theoretical arguments of the previous subsections are now formalized by means of Directed Acyclic Graphs (DAGs). DAGs can be seen as a generalization of the classical path diagram (Wright 1934). They provide formally proven rules for selecting control variables when estimating causal effects. Although DAGs are vividly discussed in the methodological community, they are not very often used in applied sociological studies. Unfortunately, space limitations only allow a very brief introduction at this point; however, a colloquial introduction streamlined to the content of this paper is given in the online-appendix. Furthermore, we refer to different excellent introductions to DAGs (Morgan & Winship 2007; Elwert 2013; Pearl et al. 2016; Schüssler 2018).

The DAG in figure 1 shows our assumptions about the causal relationships between those variables we regard as relevant for the association between families’ educational background and the decision to learn Latin. The DAG includes the parents’ “educational background”, the decision to “learn Latin”, the three possible mechanisms, namely “cultural distinction”, “secondary instrumental function”, and “exclusive learning environment”, the related variable “school choice” and finally “spatial social proximity structure”. Aside from these variables of immediate theoretical interest, the DAG also includes three further relevant variables, namely “parents’ age”, “child’s sex”, and the “child’s educational performance”. Since these further variables are merely considered nuisances that require attention in the empirical analysis, they are drawn in light gray, allowing us to concentrate on the variables of immediate theoretical interest. The other variables will be discussed in more detail in subsection 3.1 below.

The DAG assumes that the family’s educational background affects language choice directly and indirectly by three mechanisms: cultural distinction, socially exclusive learning environment, and the expected secondary instrumental function of learning Latin. The socially exclusive learning environment is thereby partly a consequence of the spatial-social proximity structure and is a cause of the school choice.

>Figure 1:

Directed Acyclic Graph of the mechanisms linking educational background and language choice

The three indirect effects of educational background on the decision to learn Latin constitute the parameters of interest of our paper, namely:

1. There is a positive indirect effect of educational background on learning Latin through cultural distinction. This indirect effect is, however, only expected for historically academic families, whereas it should not be observed for non- or new academic families.^[7]

2. There is a positive indirect effect of educational background on learning Latin through the choice of a socially exclusive learning environment. This indirect effect should exist for new academic and historically academic families alike as we expect both types of families to feel the same pressure to send their children to a distinctive educational environment.

3. There is a positive indirect effect of educational background on learning Latin through the beliefs in a secondary instrumental function of this choice.

3.1 Identification strategy

The common design to identify the size of an indirect effect of some variable X on some outcome variable Y through the mechanism Z is to estimate the association between X and Y with and without adjusting for Z (commonly termed the “full” and the “reduced model”) and then to compare the two associations. However, this identification strategy has serious risks for so-called “collider biases” (Breen 2018; Elwert & Winship 2014). Therefore, we will discuss whether and how the standard identification strategy can be applied to the three mechanisms in question. To ease the discussion, we provide a simplified version of figure 1 that shows how families’ level of education impacts on the likelihood of learning Latin without taking into account the path through the chosen type of school (figure 2). This is exactly the situation underlying the Gymnasium survey because here the choice of the type of school has already been made (figure 2).

Given the situation depicted in figure 2, our first parameter of interest, namely the indirect effect of the educational background on learning Latin through cultural distinction, can be identified in a standard way: Just subtract the effect of the educational background net of cultural distinction from the total effect of the educational background. Thereby, the total effect of educational background on learning Latin can be estimated with a regression of learning Latin on educational background adjusting for parent’s age. For the estimation of the controlled effect, the variables for immediate cultural distinction and child’s sex must be added to the regression model. The additional adjustment of child’s sex is non-standard but necessary to prevent the collider bias described by Breen (2018). However, since child’s sex does not contribute to the total effect parents’ educational background has on learning Latin, we further simplify the estimation by adjusting for child’s sex in both models. The difference in the effects of educational background on learning Latin between both models is our first strategy to estimate the first parameter of interest.

Figure 2:

Simplified DAG for indirect effect through cultural distinction

A direct application of the standard research ritual would identify the indirect effect of educational background on learning Latin through “cultural distinction” by subtracting the association between educational background and learning Latin adjusted for cultural distinction from the corresponding association without controlling for cultural distinction. It must be noted though that under the proposed data-generating process, the adjustment of the immediate cultural distinction would create the so called “collider bias” originating from the assumed role of child’s sex (see Elwert and Winship 2014, or Breen 2018 for elucidations). The same problem would arise if the secondary instrumental function would be at least partly a cause of immediate cultural distinction (i. e., if there really should be an arrow from the secondary instrumental function to cultural distinction in figures 1 and 2). To prevent the collider bias through child’s sex and to make the analysis more robust against possible flaws in the assumed data-generating process we estimate the indirect effect in three different ways:

(1) Under the conditions shown in figure 2, the difference between the regression coefficients with and without cultural distinction should be the same regardless of whether we additionally adjust for the secondary instrumental function or not. Thus, we replicate the first strategy using the secondary instrumental function as an additional control variable. Since this second strategy estimates the same parameter of interest, we regard this second strategy as a robustness check for the first strategy. (2) Another way to estimate the indirect effect in question is to again use the same principle but to adjust for educational performance as well (only shown in figure 1). While it is true that this additional adjustment variable introduces an endogenous selection bias into the estimation of the effects of educational background on learning Latin (Elwert 2013), this bias should be the same for the full and the reduced model. Thus, calculating the difference of the regression coefficients of the educational background with and without this additional control variable is a further estimate of the indirect effect in question. In the following, we refer to the three estimation strategies as the minimal, medium, and full adjustment set. Since we will use the Gymnasium Survey for all three strategies, we also adjust for school choice throughout (see section 3.2 below).

The identification strategy for parents’ belief in the secondary instrumental function of learning Latin corresponds to the strategy for cultural distinction, except that this time the medium set contains cultural distinction instead of the secondary instrumental value. As before, all replications adjust for school type by using the design of the Gymnasium Survey.

The identification strategy for the exclusive learning environment is a bit more complicated. To clarify the situation, figure 3 shows those paths of figure 1 that create an association between educational background and learning Latin through an exclusive learning environment. The DAG makes clear that the standard strategy applied above would unavoidably fail here: If the DAG were correct, one could estimate the total effect of the educational background on learning Latin without further adjustment. The estimation of the effect of educational background on learning Latin, adjusting for the exclusive learning environment, will however also “unblock” the non-causal path from educational background to learning Latin through the exclusive learning environment, the spatial-social proximity, and school choice (see Elwert & Winship 2014 for an explanation).

Figure 3:

Simplified DAG for indirect effect through exclusive learning environment

In other words: Any adjustments of the learning environment in an analysis would in fact remove parts of the association between educational background and learning Latin, but at the same time would add an association that was not present in the unadjusted model. This is another example of a collider bias (Breen 2018). Unfortunately, running the analysis by comparing the effects of educational background with and without adjusting for school choice would also not suffice. This is because the adjustment of school choice not only removes the path from the educational background to learning Latin through the exclusive learning environment and school choice, but also the path through social-spatial proximity and school choice.

Given the assumptions of the data-generating process encoded in the DAGs of figure 3 (and 1), a feasible way to estimate the learning environment mechanism is as follows: first, to calculate the association between educational background and learning Latin with a regression that solely adjusts for spatial proximity; second, to calculate the corresponding association that solely adjusts for school choice; finally, to subtract both associations from one another.^[9] To the extent that the Gymnasium survey controls for school choice and the primary school survey controls for the social-spatial proximity structure but not for the learning environment, this estimation strategy boils down to compare the relevant association between these two datasets, which is what we will do. Again, the analysis is replicated using a minimal, medium, and full adjustment set. Thereby, the medium adjustment set adds both cultural distinction and the belief in the secondary instrumental function to the minimal set.

We are well aware that the described identification strategy is unusual and perhaps controversial. However, we would like to defend our strategy by making the most crucial underlying assumptions as clear as possible: The first assumption is that the two surveys either hold constant the social-spatial proximity structure or the school type. This will be explained in the next sub-section. The second assumption is that educational background affects school choice only through the social-spatial proximity structure and/or the exclusive learning environment – an assumption that is similar to the standard assumption of instrumental variable regression. Last, but not least, we assume that the estimated indirect effect is actually homogeneous across individuals; this, of course, mimics a standard assumption applied when inferring results of experimental research on special populations to larger populations (see Kohler et al. 2019; Pearl & Bareinboim 2014). More specifically, any argument suggesting differences in the effects of figure 1 between locations (i. e., states, cities) in the so-called “old states of Western Germany” would place a question mark behind the results of this article. We invite critics to present such an argument alongside a proposal for a better identification strategy for the learning environment hypothesis.

3.2 Data

The estimation of the indirect effects proposed by the hypotheses requires a dataset that allows to adjust for cultural distinction, beliefs in the secondary instrumental function of learning Latin, school choice, and spatial proximity. To the best of our knowledge, no available dataset enables such an analysis. We thus conducted a survey to interview both parents of children in the fourth grade (Primary School Survey) and the eighth grade (Gymnasium Survey). This section describes the survey in some detail (see Arneth et al. 2021a, 2021b for a more thorough description); the dataset is available at the GESIS data archive.

In the Primary School Survey, parents of students who attended the fourth grade of a primary school were surveyed. In the German educational system, students usually attend primary school until the end of the fourth grade and then decide which secondary school to attend at the end of the fourth grade. The survey was conducted from 27 April to 9 May 2017. At this time, the families had already registered for the secondary school the child would be attending from the next school year onwards. The survey was conducted at primary schools in the city center of the German city Düsseldorf. Düsseldorf was selected because we found there several Gymnasiums that differ in their schooling profile but are located in close proximity. The primary schools were then selected from that area. Consequently, students living in this neighborhood can choose between different secondary school types within a very similar distance. Hence, the Primary School Survey holds constant the spatial-social proximity. We distributed 1132 questionnaires to parents living in 566 families, assuming each household consists of two parents and at least the child whose language decision was of interest to us. We received 420 questionnaires back from persons living in 229 different families. Hence, the response rate of the Primary School Survey was 37 %. While this might be regarded as small, it should be noted that it is higher than the response rates achieved for the German sample of the European Social Survey in 2018 (28 %; see ESS 2019: 107) and the ALLBUS, i. e., the German General Survey (32 %; see GESIS 2018: viii). It should also be mentioned that 37 % is a conservative estimate of the response rate as we did not consider single-parent families. Due to item nonresponse, we lost around 11 % (n=26) of the total sample size, resulting in a total number of 203 families for our analysis.

The Gymnasium Survey targeted parents of students in the eighth grade of the Gymnasium. The schools were located in the West German states starting with secondary school at grade 5 (Rhineland-Palatinate, Hesse, Baden-Württemberg, North Rhine-Westphalia, Lower-Saxony and Schleswig Holstein)^[10]. The schools included in the survey were the only ones with a humanist profile in the respective cities and offered the students the possibility to choose between Latin from grade five onward, from grade six onwards or no Latin at all. To the extent that the learning environment is approximately the same for all classes in a school, the Gymnasium survey holds this variable constant. Hence, the decision for the school does not intervene with the decision for a language profile. We distributed 2994 questionnaires to 1497 families and received back 1102 questionnaires from 554 families, resulting in a response rate of the Secondary School Survey of again, 37 %. In this survey we lost around 12 % (n=68) of the families due to item-nonresponse, resulting in a total number of 483 families for the analyses.^[11]

3.3 Variables

The outcome variable of all analyses is whether the children decided to learn Latin from grade five onwards or not. Note that in the Primary School Survey the answer to this question measures an intended choice. Students had already registered to learn Latin but had not actually started. For the Gymnasium Survey, the answers reflect that students started to learn Latin around three years ago. We do not differentiate between learning Latin from grade five onwards and learning Latin from grade six onwards, as Sawert (2018) shows that there is no effect of educational background on learning Latin from grade six onwards, but a substantial effect of educational background on learning Latin from grade five onwards.^[12] The central explanatory variable is the educational background of the households. We distinguished three categories:

• Non-academic: neither parents, nor grandparents have an academic degree.

• New academic: at least one parent, but none of the grandparents have an academic degree.

• Historically academic: at least one parent and at least one grandparent have an academic degree.

The variable is constructed based on the information provided by each parent about their highest educational degree. Furthermore, the parents were asked about the highest educational degree of their parents.

Cultural distinction was operationalized by orientation towards highbrow activities^[13] (see also Goßmann 2018).^[14] We built this variable based on the frequency parents visit the theater, museums, arts galleries, the opera, the ballet, and classical concerts.^[15] If two parents were available in a household, we would calculate the average of both parents’ cultural activities.^[16] In either case we distinguished three categories:

• Rarely, meaning never to a maximum of three times a year,

• Regularly, meaning four to five times a year and,

• Often, meaning more than five times a year.

The belief in a secondary instrumental function of learning Latin was operationalized by asking parents which foreign language they expect to be the best in terms of (1) deepening the understanding of the German mother tongue, (2) improving the skills in a foreign language, (3) improving logical thinking, and (4) realizing future studies. Both parents were asked these questions independently and could choose between English, French, Latin, and Spanish. Based on this information, a variable was generated, describing for each family whether modern languages or Latin are considered more useful. The variable consists of three categories:

• Advantage Latin: Both parents answered in three of the four dimensions that Latin is more useful.

• Advantage modern languages: Both parents answered in three of the categories that a modern language is more useful.

• No difference: The parents disagreed, or the parents’ answers were inconsistent.

We coded “no difference” as the reference category. We expect that an ascribed advantage for modern languages reduces the probability of learning Latin, whereas an ascribed advantage for Latin increases the probability.^[17]

3.4 Estimation

We used logistic regression models for the estimation of the effect of educational background on learning Latin. For the comparison between the regression coefficients of nested logistic regression models, we used the KHB correction suggested by Karlson and Holm (2011). The actual estimation was done with version 2.13 of the user-written Stata command KHB (Kohler et al. 2011). The significance level for significant indirect effects was set to α=0.05. Since our hypotheses suggest the direction of the indirect effect, we used one-sided significance tests and one-sided confidence intervals.

The effect of the socially exclusive learning environment mechanism is estimated by d=b_Prim – b_Gym, i. e.the difference between the coefficient of the educational background from a logistic regression based on data of the Primary School Survey b_Prim and the equivalent coefficient b_Gym based on data of the Gymnasium survey data. The standard error for the difference d is S.E.(d) = , with Var(b) denoting the squared standard errors of the regression coefficients (Clogg,Petkova & Haritou 1995: 1279).^[18] The one-sided 95 % confidence interval around d then is [d – 1.65⋅S.E.(d), ∞], and z = is a standard normally distributed test statistic that should be larger than 1.65 in order to reject the hypothesis that the coefficient from the Primary School Survey is larger than the coefficient of the Gymnasium data.^[19]

4.1 Cultural distinction mechanism

Table 3 shows estimates for the degree to which the effect of new and historically academic backgrounds on learning Latin can be attributed to “cultural distinction”. The estimates are presented for new academic and historically academic families for the three different control sets. As the estimated indirect effects are very robust regarding the different adjustment sets, we will discuss the results of the medium set only.

Relying on the reduction of the average marginal effect (e. g., Auspurg & Hinz 2011) between the models with and without adjusting for cultural distinction, we observe that new academic families choose Latin more often (~2 percentage points) due to cultural distinction than families without an academic background. While this indirect effect has the expected sign, we consider this to be a rather small mediation effect. It is also not statistically significant. However, it should be mentioned that the effect of a new academic background is already small in the reduced model.

The expectation of hypothesis 1 was that cultural distinction particularly mediates the effect of historical academics. Relying again on the reduction of the average marginal effect, we see that historically academic families choose Latin more often (~4 percentage points) due to cultural distinction than families without an academic background. This is, in fact, a bit larger than the corresponding indirect effect of the new academics, but not much. Nevertheless, the mediation effect of the historically academic families has the expected sign and is statistically significant.

Overall, we find moderate support for the hypothesis that learning Latin from grade five onwards is associated with a general tendency to acquire a distinctive cultural status; this predisposition partly mediates the effect of educational background. However, the effect is rather small, and it remains debatable whether it is stronger for historical academics than for new academics.

4.2 Socially exclusive learning environment mechanism

Table 4 shows estimates for the degree to which the effect of new and historically academic backgrounds on learning Latin can be attributed to a socially exclusive learning environment. The indirect effect that is shown in table 4 is the difference between the effect of educational background on learning Latin in the Primary School Survey and the Gymnasium Survey and is estimated as outlined in section 3.4.

The results show that a substantial part of the effect of educational background can be attributed to the selection of the socially exclusive learning environment. For the new academic families, the effect is between 12 and 19 percentage points, depending on the adjustment set. For the historically academic families, the effect is between 11 and 15 percentage points. Hence, the indirect effect via the path of a socially exclusive learning environment is substantial for new and historically academic families in all adjustment sets. Additionally, it is statistically significant at α=0.05 for all estimated effects except for the indirect effect of new academics in the full set (p=0.08, one-sided test). Overall, we take the results as strong support for hypothesis 2: The selection of a socially exclusive learning environment is an important reason why families with an academic educational background start to learn Latin.

4.3 Belief that Latin has a “secondary instrumental function”

Table 5 shows estimates for the degree to which the effect of new and historically academic backgrounds on learning Latin can be attributed to the belief in its “secondary instrumental function”.

The results show small changes in the average marginal effect between the models with and without the measure for secondary instrumental functions. Depending on the adjustment set, the difference lies between 1 and 2 percentage points for new academic families and between 3 and 4 percentage points for historically academic families. Thus, the mechanism “secondary instrumental function” is overall almost as small as the mechanism of cultural distinction. The indirect effect of secondary instrumental function is only significant for the historically academic families.

With respect to hypothesis 3, the results are somewhat inconclusive. The indirect effect is small and not significant for the new academics. It is also small for the historical academics, but it is significant and about as strong as the cultural distinction mechanism. To be consistent, we thus conclude that there is moderate evidence that the belief in the secondary instrumental function operates as a mechanism for transferring social origin into language choice, but that it is not very important.