A macro-level analysis of language learning and migration

2.1 Dependent variables: The Goethe-Institut dataset

The Goethe-Institut (GI) is a German association that promotes the study of the German language and culture abroad. Since its foundation in 1951, it has been maintaining institutes worldwide. At many of these institutes, locals can study the German language and obtain language certificates which are widely recognised. The GI is mainly funded by the German government and through course fees (Goethe-Institut e.V. 2019). The GI reports key statistics about language learning in its yearbooks. The datasets from which our dependent variables are taken were created by digitizing these statistics.^[5]

The first dataset covers language learning abroad. It reports yearly observations of course and exam participation at Goethe institutes around the world. We use the exam participation variable from this dataset, because it has two advantages over course participation: First, the exams are standardised, cover fixed amounts of learning content (e. g. the A1/1 level in the Common European Framework of Reference for Languages, CEFR) and are therefore not affected by decisions of individual Goethe institutes. Second, exam participation is open to learners who did not participate in a language course at one of the Goethe institutes. As a consequence, it may act as a better proxy for the extent of language learning activities in the respective country.

The second dataset reports the yearly number of language course participants summed up for all Goethe institutes in Germany and disaggregated by the nationality of the participants. Figure A1 shows the distribution of the German institutes and informs about their presence during the time period from 1992 to 2006. This is the period of our analysis as data from both datasets and migration data for a large number of countries are available. At the same time, this period pre-dates regulation about the “A1 requirement” for family reunification, which was introduced in 2007.

While the aim of the analysis in this paper is to address language learning and its relation with migration in general, our data are limited to learners that come into contact with a Goethe institute. Naturally, there are a number of other language learning opportunities, including universities, private language schools, and internet platforms. We want to address two possible issues. First, there could be concerns about the general importance of the GI as a provider of language courses and exams and about a link with migration to Germany. Figures 1 and 2 provide an overview of the numbers of exams and course participants in our dataset and relate them to the number of migrants to Germany. As our sample is not balanced, the numbers are averages per countries per continent and year. For illustration, let us look at the numbers for Europe for 2006: With the ‘Germany’ specification, we find that in 2006, there were on average almost 230 course participants in each of the 36 European countries in our sample (amounting to a total of close to 8,200 course registrations). In the same year, about 10,000 individuals migrated on average to Germany from the respective countries. Analogously with the ‘Abroad’ specification, there were on average almost 490 exams taken in each of the 20 European countries with one or more institutes in our sample (amounting to a total of almost 9,800 exams). In the same year, about 25,600 individuals migrated on average to Germany from the respective countries.^[6]

Figure 1

Germany specification – averages per continent and year.

Figure 2

Abroad specification – averages per countries per continent and year.

Even though the data on exams and courses do not allow any conclusion about the market shares of the GI in the different countries, the data show that the numbers of exams and courses taken is non-negligible. Furthermore, in the context of our study, our language data, which very likely only cover part of German language learning, make it less likely to find any association. To say it differently, any association which we might find can be seen as an ‘a fortiori’ association. Also note that for language learning abroad, we partially alleviate this problem by using exam numbers. They cover a larger group of learners who may have learned the language elsewhere but have their skills certified at a Goethe institute (for language learning in Germany, this data is not available).

Second, the multitude of alternative learning opportunities gives rise to further concerns regarding the self-selection of language learners into courses offered by the Goethe institutes. Three characteristics on which self-selection may be based are willingness or ability to pay, location, and age.

Selection on willingness to pay could occur if the prices of courses at the Goethe institutes differed significantly from the costs of other equally suitable learning options. On the one hand, one might suspect the Goethe institutes to be somewhat of a premium provider of language courses, because they are part of a semi-official German organization with a long tradition and a good reputation. Such a status would allow them to charge higher prices. On the other hand, one might suspect courses to be particularly cheap, because the majority of the Goethe institutes’ funds comes from the German government. There are several reasons why both concerns might not be warranted. First, in conversations with us, employees of the GI have stated that language courses are priced to be self-financing and that government funding is used for non-language-course-related activities only. Second, prices of language courses do not point towards any systematic deviation relative to competitors’ prices. Historical price data on language courses are not available to the best of our knowledge. For a rough idea, Table B1 in appendix B.1 contains current price data on comparable language courses offered by the Goethe institutes and by other institutes in six cities in different countries.^[7] While the data are far from complete or representative, they do not indicate that the Goethe institutes are usually the most expensive provider in the market. Additionally, according to employees of the GI, the price policies of the individual institutes take the prices of local competitors into account and are at least indirectly related to the countries’ income levels. It should also be noted that we have 137 institutes in 76 countries in our sample; courses offered by institutes in Germany are attended by 157 nationalities. At the country-level, there is no indication that the demand for courses comes from higher-income countries only; this does, however, not rule out that within countries, participants might be self-selected. Still, a priori there is no reasons to expect that results are upward biased.

Goethe institutes are usually located in capitals and other major cities. The lack of institutes in rural areas is likely to lead to an under-representation of language learners from these areas among participants at the Goethe institutes. However, the bias need not be as large as one would initially expect: Goethe institutes offer both extensive and intensive language courses. Extensive courses are based on weekly lessons and last for several months, but intensive courses are taught en-block. Participants of intensive courses do not necessarily have to live in the vicinity of the respective institute. They may also stay there for the duration of the course only. This holds even more so for exam participation where a day-trip to the institute might not be so uncommon.

The language courses taught by the Goethe institutes are a traditional “offline” form of language learning (at least during the years of our analysis). At the other end of the spectrum, there are pure online courses like those offered by “myngle” or “babble”. The latter kind of courses may be more attractive to a younger generation of language students, which is more familiar with using the internet in general. While this difference may lead to an over-representation of older participants in language courses at the GI, the advent of online language learning platforms in the late 2000s falls in the very last years of the period of observation (1992–2006) of our analysis. Consequently, we are confident that the age bias only has a small, if any, effect on our results.

Last, in order to isolate the incentive channel, i. e. how migration intention affects language learning incentives, the possible relation between children-age and adult-age language learning needs to be discussed. Participating in a course or exam at a Goethe institute does not rule out that some language skills have already been acquired before and possibly for reasons not related to migration. Rather it provides a complementary perspective to the sorting channel with a focus on deliberate decisions by adults who want to continue learning, brush up their skills and possibly work towards a certificate. Furthermore, there is evidence that a substantial part of course and exam participants start learning the German language as an adult. The number of courses offered at the basic A level exceeds by a factor of at least 4 the number of courses offered at the advanced C level in most institutes.^[8]

More generally, it should be noted that systematic differences across countries, for example, related to the organization of courses or the composition of the group of language learners are captured by our fixed-effect approach. In addition, using exam numbers in our ‘Abroad’ specification alleviates some of the concerns mentioned above.

2.2 Explanatory variables

In the ‘Abroad’ specification, the explanatory variables are characteristics of the countries where the respective institutes are located or characteristics of the relationships of these countries with Germany. In the ‘Germany’ specification, we use the characteristics of the countries of origin of the language course participants and these countries’ relationships with Germany instead.

2.3 Hypotheses

Previous studies that explore the relationship between language and migration fall into one of two groups as outlined in Section 1: The first group connects linguistic characteristics or aggregate measures of language skills to migration flows between locations, usually between countries (see, for example, Adserà and Pytliková 2015; Belot and Ederveen 2012). The seconds group connects individual language skills to migration decisions or to a measure of migration success (examples are Dustmann and Soest 2001, 2002, for the effect of language skills on wages). None of these studies look at language learning decisions explicitly, but a clear set of predictions regarding the motivation of those decisions emerges from their results. While both groups of studies form the basis for our theoretical considerations, our empirical approach is closer to that of the first group because we look at aggregate data. As a consequence our hypotheses aim to establish links between aggregate measures on the basis of individual motivations.

4.1 Main estimation results

Table 3 shows the results of our ‘Abroad’ estimations. For institutes in the ‘Free Move’ countries, we find a positive coefficient for immigration flows. It reduces in size as we add further migration-related variables, but remains significant in all specification. We also find a similarly sized significant coefficient for student immigration. In our preferred specification ‘Free Move 3’, a 1 % change in immigration corresponds to a 0.3 % change in exam participation, while a 1 % change in student immigration corresponds to a 0.4 % change. The correlation is stronger for countries with a larger migrant stock in Germany and with a higher GDP per capita. In contrast to the results for the ‘Free Move’ countries, neither general nor student immigration flows are significantly associated with exam participation in the ‘Restricted Move’ countries. Instead, we find a large and highly significant coefficient for migrant stocks. After adding controls in ‘Restricted Move 3’, a 1 % change in migrant stocks corresponds to a 0.7 % change in exam participation. In summary, our abroad estimations indicate a strong link between language learning and migration flows for ‘Free Move’ countries, which is replaced by a strong link between migrant stocks and language learning for ‘Restricted Move’ countries.

Table 4 shows the results of our ‘Germany’ estimations. We find no significant results for ‘Free Move’ countries. Furthermore the R 2 of even the most basic specification ‘Free Move 1’ is 97 %, indicating that country and year-specific fixed effects already explain most of the variation in language course participation by citizens of ‘Free Move’ countries in Germany.^[20] For citizens from ‘Restricted Move’ countries, we find a positive coefficient for immigration flows. It remains stable in size and significance as we add our additional migration-related variables. In our preferred specification ‘Restricted Move 3’, a 1 % increase in immigration flows corresponds to a 0.3 % increase in language course participation. The coefficient for student immigration is also positive, significant and of similar magnitude as the coefficient for general immigration.

When comparing our preferred specifications from Tables 3 and 4, we observe a clear mirror pattern in the results for migration flows. For ‘Free Move’ countries, both general and student migration are significantly associated with language learning outside of Germany, while for ‘Restricted Move’ countries^[21] there is a significant correlation between both general and student migration and language learning in Germany. This pattern is in line with our expectation (see Section 2.3.2) that individuals from countries with unrestricted and with restricted access to Germany face different trade-offs when it comes to the decision whether to learn German before or after migrating to Germany. The results suggest that there is a strong certainty-of-investment mechanism at play, where membership in the EU or in the Schengen area encourages pre-migration language learning by guaranteeing access to the German labour market. In contrast, we find no evidence of a bridge-the-gap mechanism that encourages citizens of countries outside the EU or the Schengen area to invest into language skills early to facilitate job seeking before their arrival in Germany.

The positive association of migrant stocks with language learning abroad, which we find for ‘Restricted Move’ countries (see Table 3), can also be related to uncertainty considerations: A large stock indicates that – at least in the past – many immigrants of a given country were successful in entering Germany. At the same time, these settled immigrants can help potential new immigrants to reduce uncertainty. Both can be expected to be positively related to preparatory language learning.

4.2 Capitals and large cities

A potential concern with our ‘Abroad’ specification is that institutes in cities with different characteristics attract language learners from different sub-groups of the population of the respective country and that our estimations throw all of these sub-groups together. In particular, institutes in large cities may attract a large number of language learners that work for the government or for large businesses. While we cannot control for the individual characteristics of different language learners in our macro-level study, we can use city-level data to shed light on these differences. Table 5 presents the results of four estimations for language learning abroad, where we split the sample based on city population. ‘Largest’ includes the largest city in each country; ‘Other’ contains the remaining cities.

The sample split does indeed reveal a difference between cities that are and cities that are not the largest in their country. For the relationship between migration flows and language learning in ‘Free Move’ countries, the largest cities seem to be responsible for the results regarding general flows when looking at those countries with a larger migrant stock and those with higher GDP per capita. While for the smaller cities, the positive correlation between migration flows and language learning is also there for countries with a larger migrant stock, we observe a negative correlation for countries with higher GDP per capita. Language learning may be particularly important for migrants connected to large businesses or the government which are more common in the largest cities of each country, especially in high-GDP countries. The smaller cities are, however, responsible for the results regarding student migration. This may be the case because smaller cities have higher relative student populations.

This last point may also apply to language learning outside of the EU or the Schengen area, where the positive association between migrant stocks and language learning is driven entirely by the largest cities. Growing migrant stocks in Germany may create additional demand for language skills among those individuals in big business or the government who establish ties with Germany through the diaspora.

5.1 OLS estimations with country characteristics

For the OLS specification with time-invariant country characteristics, we consider the ‘Free Move’ countries and the ‘Restricted Move’ countries separately as different characteristics might be relevant for the two groups of countries. Also data availability is a larger issue for the countries with restricted access to Germany; this is also the reason why some of the variables are included as time-invariant variables even though they possibly vary across time (Table A1 informs about the chosen years).

For the ‘Free Move’ countries, we include different geographic, linguistic and cultural distance measures as well as information about the educational level and the sectoral composition. Following Ginsburgh et al. (2017), we also add information about the importance of the German language in the origin country and of the immigrants’ native language in the world. For the ‘Restricted Move’ countries, we also include distance measures (due to data availability we cannot control for cultural distance, however). The importance of the German and the native language as well as the sectoral composition are also captured. The educational level is extended by the literacy rate. In addition, we control for a political rights index and a civil liberties index and the presence of a visa requirement for Germany. Tables A2 and A3 provide the descriptive statistics.

We would expect that linguistic and cultural distance is negatively correlated with language learning in general, while geographic distance might differently affect where to learn the language and this might also depend on whether access to Germany is free or restricted. More German speakers in the home country, which indicates some closer economic and social links with Germany, should lead to larger incentives to learn German; more speakers of the native language in the world should, on the contrary, decrease the benefits of learning a foreign language. A higher educational level in the origin country might be positively correlated with learning German. If a visa is required, this makes migration and the benefits from language learning more uncertain and should thus be negatively related to language learning, especially abroad.

Tables A4 and A5 present the results. As the samples become smaller because of the data issues mentioned above, the fixed-effect estimations for the smaller sample are shown next to the respective OLS specifications for easier comparison. We find that the results for the immigration variables do not differ in an important way for the fixed-effect and the OLS specifications. It shows that the coefficients of the population and population-city variables are now highly significant in the OLS estimation; this was apparently absorbed by the institute and country fixed-effects. Looking at the country characteristics, a few patterns can be observed. The linguistic and cultural distance measures mostly show the expected negative link, while the picture for geographic distance is more mixed. The share of German speakers in the home country is not significantly correlated with learning German, while the result for the world speakers of the immigrants’ natives language is mixed. Visa requirements decrease the incentives to learn German abroad in the ‘Restricted Move’ countries in line with the certainty-of-investment mechanism discussed above. Overall, the OLS specification allows gaining further insights into the role of heterogeneities across countries. At the same time, the main results related to the migration variables are mostly unaffected.

5.2 Within-between estimations

To further explore between-country differences, we complement our fixed-effects models with a random-effects (RE) specification. We use Bell and Jones’ (2014) re-formulation of an RE estimator proposed by Mundlak (1978). Instead of removing all between country variance from the data, the RE estimator uses it to identify the between coefficients (subscript B). As with the FE estimator, the within coefficients (subscript W) are only identified by within-country variation. For exam participation abroad, we estimate

and for course participation in Germany

where Z stands for I m m i g and S t u d I m m i g and bars indicate averages.^[22] This within-between specification also disentangles the interaction between immigration (and analogously for student immigration) and migrant stocks into three components: One component that is driven by the within variation of both variables ( WW) and two additional components, each of which is driven by the within variation of one but by the between variation of the other variable ( WB and BW). For example, the within-between interaction “Immigration (B) × Migrant Stocks (W)” can be interpreted as the change in the effect of the country-average of immigration on language learning that is brought about by growing (or shrinking) migrant stocks. A fourth interaction term is identified by the between variation of both variables ( BB), but this term has no equivalent in the FE estimation, where it would have been captured by the institute-fixed or country-fixed effects. u i , u j , u t , u j t and u i j t are error terms (random effects). These specifications are extensions of Bell and Jones’ (2014) equation (12). We add the interaction terms and the additional error term u i to account for our nested geographical structure of institutes within countries.

We would expect most of the coefficients that are identified by within variation to be the same in our RE and FE estimations, because the regression models are very similar. The two differences are the disentangled interaction terms discussed above and the different error structure. In addition to the FE models’ idiosyncratic error term, u i j t or u j t , the RE models allow for random errors on the levels of institutes, countries, and years. When looking at Table 6, we see our expectation of identical within coefficients largely confirmed. In particular, we find highly significant within-coefficients for general immigration and student immigration in the ‘Abroad, Free Move’ specification and in the ‘Germany, Restricted Move’ specification (cf. also Tables 3 and 4). This also holds for the within-coefficient of migrant stock in the ‘Abroad, Restricted Move’ specification.

The within-between model adds several insights. First and in addition to the significant within-coefficient of immigration and student immigration discussed above, the between-coefficients of both, immigration and student immigration, are large and significant in both the ‘Abroad’ and the ‘Germany’ specification for ‘Restricted Move’ countries. This implies that countries that, averaged across the entire period of observation, sent more students to Germany also showed more language learning activity, both abroad and in Germany. There are at least three potential explanations for this result. First, in the long term, (student) migration may lead to increased interest in the German language and culture abroad. Second, strong cultural links may not be the consequence but the cause of both language learning activities and migration. While time-invariant cultural links are controlled out of our within coefficients, they do enter into the between-coefficients. Third, (student) migrants may learn German several years before or after staying in Germany. As discussed above, this effect would not be picked up by our unlagged within-coefficient, but may show up in the between-coefficient.

Second, we see a significant within-between-coefficient for the interaction of immigration and student immigration, respectively, with the migrant stock in the ‘Abroad, Free Move’ specification (and partially in the ‘Germany, Restricted Move’ one). In countries with, on average, a larger migrant stock, a change in immigration is positively associated with language learning, while the opposite holds for a change in student immigration. The incentive effects thus diverge. Minority language concentration decreases students’ incentives to learn the language in the home country before migration. This could indicate that they rely more on help from fellow citizens at the beginning of their stay in Germany. On the contrary, immigrants with a likely larger labour-market interest increase their language learning effort in the presence of a larger stock of migrants. One reason could be that the fellow citizens make them aware of the importance of language skills.

Third, the random effects model allows us to explicitly compare the variances of its various error terms. The comparison shows that most of the variance that is not attributed to our explanatory variables can be attributed to time-invariant factors on the city level in the ‘Abroad’ estimations and on the country level in both estimations. The variances of the respective error terms are larger than the variances of the other entity-level error terms and also larger than the residual variance in all but the ‘Germany, Restricted Move’ sample. Fixed effects on the year-level do not seem to play a role as the variance of the respective error term is very small in all estimations.

5.3 Causality issues

Above, we present evidence of a positive association between language learning and several migration-related variables. For a causal interpretation, the issue of reverse causality needs to be addressed. In this section, we provide arguments why we think that this association might be driven by a causal effect of those variables on language learning and complement this by an instrumental-variable (IV) exercise.

With respect to language learning in Germany, a reverse causal effect of course participation on our migration-related variables is unlikely, because language learners are already located in Germany by definition. Theoretically, the availability of post-migration language courses might affect the decision whether and where to migrate. However, since courses that teach migrants the language of the destination country are likely to be available in any potential destination, the effect of availability on migration can be neglected in our application.

With respect to language learning outside of Germany, three channels of reverse causality could be particularly relevant: the opening and closing of institutes, the participant recruitment and capacity planning of the institutes, and changes in the individual motivation of participants. First, the opening and closing of institutes may be motivated by the presence of migration flows. This channel of causality would affect the results of our estimations if these estimations “compared” participation in cities in which language courses or exams were offered with zero participation in cities where this was not the case. However, our dataset only includes observations for city-year combinations where language courses were offered and is, thus, not susceptible to the endogenous opening and closing of institutes.^[23] Second, the presence of large migration flows from a particular country may motivate institutes in that country to advertise more heavily to “capture” a larger share of the outgoing migrants. While we cannot measure the intensity of advertising of individual institutes, there is no indication that the institutes follow such a strategy. In our conversations with officials at the Goethe institutes, they repeatedly stated that they attempt to adjust to local demand rather than to actively encourage outgoing migrants to participate in courses. Third, the language learning experience at the Goethe institutes may motivate individuals, who initially take the course for non-migration related reasons, to move to Germany. While this may be a good description of the experience of a small number of language learners, the migration choice literature seems to agree that the key determinants of migration decisions are others, income differentials and migration policies in particular (Bertoli and Fernández-Huertas Moraga 2015; Grogger and Hanson 2011; Ortega and Peri 2013).

In the next step, we want to address this in a somewhat more systematic way. We are fully aware that reverse causality issues and endogeneity issues more generally affect several of our main migration variables. Also trade, which we include as a control variable, likely suffers from this problem as Ginsburgh et al. (2017) have demonstrated. Trade can affect language learning, but better language skills can also foster trade. To overcome this problem, they implement an IV strategy based on bilateral trade shares weighted by trading partners’ respective ratios of native speakers of the target language. As we are not able to do a full-fledged IV analysis taking all possible endogeneities into account, we only consider the relation of immigration and language learning in more detail.^[24] This exercise can therefore only give a first indication; at the same time, it complements the arguments brought forward above.

Our instruments are based on the sectoral composition, geographic distance and tertiary education shares. The first stage of the 2-stage least squares (2SLS) estimation works well if we take the F-statistics as guide for the joint significance of the instruments (see the notes in Tables 7 and 8). Results of the second stage confirm our main FE-results: In the ‘Abroad’ specification in Table 7, the IV- and the FE-results are very close. In particular, we find that immigration affects language learning abroad in the ‘Free Move’ countries and does not do so in the ‘Restricted Move’ countries. We also find similar IV- and FE-results in the ‘Germany’ specification (Table 8). For the ‘Restricted Move’ countries, the results might require some further thoughts given the relatively large second-stage immigration coefficient.

As already noted, we see this as an exercise and not a full-fledged analysis. Two caveats are particularly important to keep in mind: First, the exogeneity of the instruments can be discussed. There are arguments in its favour: A larger share of the agricultural sector makes the economy more dependent on the weather and more generally on climatic conditions. Adverse developments might push individuals out of their home country without directly affecting their language learning incentives. In a similar vain, given a migration intention, geographic distance might affect the choice of the destination country and, only after that choice is made, the language learning incentives. Tertiary enrolment shares could be expected to affect immigration plans in particular for ‘Restricted Move’ countries as a higher education level might enable migrants to benefit from preferential visa regulations. We acknowledge that the exogeneity argument is weaker here as a higher education level could also directly affect language learning incentives for some individuals. Second, other variables might also be endogenous. We already mentioned trade. Student immigration could also be concerned in a similar way as immigration, even though one could argue that the endogeneity issue might be somewhat less relevant. Learning German in a Goethe institute by young people can of course be triggered by the wish to complement the language education at school and reflect a true interest in the language. Still it is more likely that this is done in preparation of a future stay in Germany possibly for studying purposes.

Overall, we see the different arguments brought forward in this section as providing some indications about a causal relation; in light of the caveats, however, we are careful to abstain from seeing this as anything more than that.