Introduction

While there is a rich literature on the decision to enrol in higher education, lesser attention has been given to the transition from bachelor’s to master’s programmes (Mullen et al., 2003; English and Umbach, 2016), including in STEM (science, technology, engineering, and math) education. This paper contributes to the study of educational decisions, analysing the intentions of enrolment in master’s education of STEM bachelor students. The study is focused on the students’ decision of pursuing a master-level degree right after graduating bachelor’s studies. Master’s degrees offer graduates of undergraduate education the possibility to acquire deeper or additional skills and knowledge. Master-level programmes engage students in more advanced research methods and independent study while focusing on a narrower specialisation. It is believed that attaining master’s degrees brings significant individual and social benefits (English and Umbach, 2016) and supports scientific and technological advancement (Committee on Enhancing the Master’s degree in the Natural Sciences, 2008). Master’s degree attainment is consistently rising and understanding factors that shape the transition from bachelor’s to master’s is becoming increasingly important (English and Umbach, 2016). A better knowledge of the mechanisms behind this process is of interest to education researchers and practitioners.

According to Hossler and Gallagher’ (1987) 3-phase college-going model, the first stage of the educational choice is the predisposition phase in which students decide on pursuing or not a higher-level education programme. Following theoretical models developed for undergraduate enrolment, this paper is focused on the factors influencing the predisposition to pursue master’s degrees among final-year bachelor students in STEM and non-STEM fields. Early studies on graduate education relied on the idea of persistence or retention of students in the education system (Tinto, 1975; Pascarella and Terenzini, 1980). Developing theoretical models based on the literature on undergraduate college choice, more recent works understand students’ process of choosing to pursue a master’s degree as a new and distinct choice among possible post-graduation alternative options (English and Umbach, 2016).

Education decisions are shaped by a dense mix of mechanisms, including what individuals can do, what they want and conditions that shape individual preferences and intentions (Gambetta, 1987). Such mechanisms embed a wide range of individual, institutional and economic factors (Table 1).

Table 1 Factors influencing educational choices based on literature review.
Full size table

Previous studies found significant heterogeneity among master’s degree students (O’Donnell et al., 2009; Jung and Li, 2021). On the one hand, demographic and background characteristics such as gender, race and age of the students, as well as their area of residence, influence their enrolment in master’s degrees (Perna, 2004; Schapiro et al., 1991; Xu, 2014; Allison and Ralston, 2018; Jung and Lee, 2019). For instance, the underrepresentation of women, as well as racial and ethnic minorities in STEM education has been analysed by numerous studies (Espinosa, 2011; Batsheva and Boards, 2019; McKinney et al., 2021). On the other hand, education decisions on whether or not to continue education are influenced by the expectations regarding academic success. Individuals with higher academic performances tend to take greater risks in this matter and enrol in higher levels of education (Latiesa, 1989; Mullen et al., 2003; Zamfir et al., 2021). According to Bourdieu (1977) theory on the role of cultural and social capital in education, parents’ education influences educational transitions as students with parents having a higher level of education are more likely to graduate at a higher education level (Jiménez and Salas-Velasco, 2000; Mullen et al., 2003; Zamfir et al., 2021). Family income is another factor of influence for educational careers. Students with better economic circumstances are more likely to enrol in higher-level courses (González and Dávila, 1998; DeBacker and Routon, 2021), including master’s degrees (Schapiro et al., 1991). Studies exploring the influence of the level of educational debt on decisions to attend graduate degrees have found mixed results (Schapiro et al., 1991; Weiler, 1994).

Other factors that shape educational choices are related to educational institutions. Quality and other characteristics of the academic environment influence educational choices and decisions in higher education (Kallio, 1995; Zhang, 2005; Zamfir et al., 2021). Educational careers are also shaped by the satisfaction of students with education. In general, the satisfaction of students with higher education is determined both by the perceived institutional performance and the perceived outcomes of institutional performance (Hartman and Schmidt, 1995). Moreover, it has been found that the type of university influences the transition from bachelor’s to master’s as students from research universities are more likely to pursue master’s degrees than those from teaching-oriented universities (Jung and Lee, 2019).

Previous research indicates a significant relationship between the field of study and the decision to pursue a master's education. It has been found that students from the arts stream are less predisposed to enrol in postgraduate studies compared to science stream students (Kong et al., 2015; Jung and Lee, 2019; Zamfir et al., 2021). Considering this difference between STEM and non-STEM students, it would be very important to better understand the variation between the two fields of study.

We know that investing in education fosters the accumulation of knowledge and skills, allowing individuals to have access to better job prospects and higher wages (Becker, 1962, 1990; Thomas and Perna, 2004; Paulsen and Toutkoushian, 2008). According to the economics of education, investing in human capital is motivated by expected economic returns (Becker, 1962, 1990; Thomas and Perna, 2004; Paulsen and Toutkoushian, 2008; Menon et al., 2017). Considering the human capital theory, the demand for higher education has been studied in relation to labour market factors such as the level of unemployment (Becker, 1964; Ashenfelter and Ham, 1979; Jiménez and Salas-Velasco, 2000) and expected earnings upon completion of a degree (Kodde, 1986). Expectations regarding monetary and non-monetary returns of education are relevant for educational choices (Altonji et al., 2015). Increased expected earnings have been found to positively influence the enrolment demand, particularly for post-graduate education (Handa and Skolnik, 1975). Jiménez and Salas-Velasco (2000) propose a model of factors determining educational choice, which includes objective and subjective determinants related to the current situation of the students, but also to the future, such as employment prospects and expected income. Additionally, students’ participation in the labour market influences their enrolment in master’s degrees (Zamfir et al., 2021).

Moreover, differences in pursuing master’s degrees in relation to students’ majors are mainly determined by differences in the expected benefits (English and Umbach, 2016; Zamfir et al., 2021). In general, STEM degrees bring higher returns (Burgess, 2016), encouraging individuals to pursue more education in related fields.

Returns to education are influenced by local conditions of the economic environment. Our previous results showed that local labour markets characterised by higher employment in science and technology increase the probability of bachelor students to enrol in master education. In addition, more dynamic local contexts with respect to earnings level, business demography, and business innovation discourage students from prolonging their education career, probably due to higher forgone earnings (Zamfir et al., 2021). From this point of view, it is possible that differences between the intentions of STEM and non-STEM students vary in relation to the local context.

In addition, the theory of skilled-biased technological change (Machin and Van Reenan, 1998) highlights the rising demand for higher skills in more technologically advanced contexts. Previous studies linked various proxies of technological change and innovation with a higher demand for skilled workers (Toner, 2011). Evidence suggests a virtuous circle between the education and skills of the workforce and business innovation capacity, as enterprises with a higher proportion of more skilled workers have a higher probability of introducing new products (Toner, 2011). From this point of view, one could expect contextual factors reflecting the business innovation environment to influence returns to education for STEM and non-STEM students in a different way. Thus, we draw on human capital theory while taking into consideration the business innovation environment for exploring the factors shaping the decision of enrolment in master’s degrees of STEM vs. non-STEM students. Detecting the factors that influence different educational choices of STEM and non-STEM students is useful for identifying effective measures for increasing participation in STEM master education. Promoting STEM education is considered a key element for driving innovation and economic growth worldwide.

Research questions and objective

Integrating human capital theory with concepts of cultural and social capital, Perna (2006) developed a model of college choice based on four layers: individual habitus, community and school context (organisational habitus), higher education context and socio-economic and policy context. Studying transition from undergraduate to graduate education, English and Umbach (2016) revise the approach of Perna (2006) and propose a two layers conceptual model integrating individual and institutional factors that influence students’ decision to pursue graduate education. The individual layer (habitus) includes demographic characteristics, cultural and social capital, academic achievement, supply of resources and expected benefits. The second layer covers features of the undergraduate institution context.

Building on the previous work of Perna (2006) and English and Umbach (2016) regarding the multiple layers of factors that influence educational choices, we take into account individual characteristics, as well as features of institutional and business innovation contexts that shape educational choices. Therefore, the objective of this paper is to propose and test a two-level model for the choice of pursuing a master’s degree that integrates individual and institutional factors with circumstances in the local business innovation environment. First level (individual) includes individual habitus and organisational habitus (higher education institutions of bachelor students, including enrolment in STEM vs. non-STEM fields). The second level (local) reflects local conditions concerning the business innovation environment. The second level recognises that the choice of pursuing master’s degree is influenced by wider forces and conditions that interplay with individual factors and shape individual preferences and intentions. The conceptual model of the current study is presented in Fig. 1.

Fig. 1: Conceptual model of the main determinants for the intention of pursuing a master program.
figure 1

This figure illustrates the two-level model that integrates individual and institutional factors (level 1) with circumstances related to the local business innovation environment (level 2) influencing the choice of pursuing a master’s degree.

Full size image

As has been shown before (Zamfir et al., 2021), STEM bachelor students display a higher propensity to enrol in master’s degrees than other students. In addition, previous results indicate that a more dynamic business innovation local context discourages students to pursue graduate education. By developing the two-level model for the choice of pursuing a master’s degree, this paper is focused on addressing the following research questions:

RQ1: Do STEM vs. non-STEM majors have different effects on intentions of pursuing master education in different local contexts?

RQ2: Does and how does the business innovation context interplay with STEM vs. non-STEM majors in shaping the predisposition of enrolment in master education?

Data and methods

In order to empirically test the proposed model, we integrated data from various administrative and statistical sources for a sample of students enroled in ten public universities located in Romania (out of a total of 54 public universities). In 2019, intentions of pursuing a master’s degree in the next school year (2020/2021) were collected via a questionnaire-based survey from 502 students enroled in their final year of bachelor studies (age M = 22.08, SD = 1.185; 273 (54.4%) men, 229 (45.6%) women) in various fields of education (medical studies, sports, military and defence have not been covered). Fields of study have been registered for each bachelor student, and two categories have been constructed: STEM and non-STEM fields. The final sample included 250 STEM students and 252 non-STEM students. For university quality, performance scores of higher education institutions have been retrieved from the national university ranking for 2019 (Guo et al., 2023). Also, individual and background characteristics (academic performance, parents’ education, subjective family income, gender, age, area of residence, employment status, working experience), as well as subjective expectations regarding economic benefits anticipated upon the completion of a master’s degree have been collected from students.

The second level of the indicators includes data on the economic context at local level, more specifically on the business innovation environment. Regional-level data on enterprises introducing product and/or process innovations have been retrieved from statistical sources (Romanian National Institute for Statistics). List and description of indicators and data sources are presented in the appendix.

Aiming to identify the main determinants of students’ intention in pursuing a master’s degree in the next year, we used a multilevel logistic regression analysis based on the hierarchical nature of the data (individuals from different universities that are placed in different regions), which includes the individual-level variables, and then explores whether university-level factors together with regional level indicators are significantly associated with the intention of pursuing a master programme.

In order to analyse the between-region variation while taking into account the influence of individual characteristics in the overall intention to enrol to a master programme, different types of two-level models were used. The general approach of constructing the models is presented in Fig. 2. The first stage was to estimate a baseline random intercept model with no explanatory variables. This step is needed for establishing whether a multi-level approach is appropriate. The null or empty two-level model with only an intercept and region effects has the following form:

$$\log \left(\frac{{\pi }_{{ij}}}{1-{\pi }_{{ij}}}\right)={\beta }_{0}+{u}_{0j}$$
Fig. 2: The general scheme of two-level models for the intention of pursuing a master program.
figure 2

This figure presents the step-by-step approach of the statistical analysis, outlining different types of the estimated two-level models.

Full size image

The intercept \({\beta }_{0}\) is shared by all regions, while the random effect \({u}_{0j}\) is specific to region j. The random effect is assumed to follow a normal distribution with variance \({\sigma }_{{uo}}^{2}\). The baseline random intercept model with no explanatory variables was estimated using maximum likelihood estimation using adaptive quadrature. The log odds are the logarithm of the odds (i.e. the ratio between a probability value (Phi) and its complementary).

The second stage was to develop a model with first-level variables (i.e. individual-level) in order to test the impact of individual characteristics:

$$\log \left(\frac{{\pi }_{{ij}}}{1-{\pi }_{{ij}}}\right)={\beta }_{0}+{{\beta }_{1}\cdot {X}_{{ij}}+u}_{j}$$

β0 is interpreted as the log-odds that y = 1 when x = 0 and u = 0 and is referred to as the overall intercept in the linear relationship between the log-odds and x. If we take the exponential of β0, exp(β0), we obtain the odds that y = 1 for x = 0 and u = 0. As in the single-level model, β1 is the effect of a 1-unit change in x on the log-odds that y = 1, but it is now the effect of x after adjusting for (or holding constant) the group effect u. If we are holding u constant, then we are looking at the effect of x for individuals within the same group, so β1 is usually referred to as a cluster-specific effect. Exp(β1) can be interpreted as an odds ratio, comparing the odds that y = 1 for two individuals (in the same group) with x-values spaced 1 unit apart.

While β0 is the overall intercept in the linear relationship between the log-odds and x, the intercept for a given group j is (β1 + uj), which will be higher or lower than the overall intercept depending on whether uj is greater or less than zero. Therefore, uj is the group (random) effect, group residual, or level 2 residual. The response probabilities \({\pi }_{{ij}}\) can be expressed as follows:

$${\pi }_{{ij}}=\frac{\exp ({\beta }_{0}+{\beta }_{1}\cdot {x}_{{ij}}+{u}_{j})}{1+\exp ({\beta }_{0}+{\beta }_{1}\cdot {x}_{{ij}}+{u}_{j})}$$

At the second level, there will be added contextual factors to the model. In the third step, the logit random intercept model specification, including both individual-level explanatory variables, as well as region-level explanatory variables, is the following:

$$\log \left(\frac{{\pi }_{{ij}}}{1-{\pi }_{{ij}}}\right)={\beta }_{0}+{{\beta }_{1}\cdot {X}_{{ij}}+{\beta }_{2}\cdot {X}_{j}+u}_{j}$$

where: \({\beta }_{0}\) is the overall intercept, \({\beta }_{1}\) is the cluster-specific effect, \({\beta }_{2}\) is the contextual effect, Xij is the vector containing individual-level explanatory variables, Xj is the vector containing region-level explanatory variables, and uj is the group (random) effect. The log odds are the logarithm of the odds (i.e. the ratio between a probability value (Phi) and its complementary).

Additionally, we test interaction effects exploring the possibility of the effect of one independent variable on the outcome to vary with the value of another explanatory variable. An interaction between a level 1 variable and a level 2 variable is called a ‘cross-level interaction’. Furthermore, it was worth to test if the effect of contextual regional factors on the decision of applying to a master programme depends on level 1 characteristics. Therefore, in the fourth step, we have estimated random intercept models with cross-level interactions.

$$\log \left(\frac{{\pi }_{{ij}}}{1-{\pi }_{{ij}}}\right)={\beta }_{0}+{{\beta }_{1}\cdot {X}_{{ij}}+{\beta }_{2}\cdot {X}_{j}+{\beta }_{3}\cdot {X}_{{ij}}\cdot {X}_{2j}+u}_{j}$$

In the random intercept models, the model intercept varies randomly across regions and the main assumption was that the coefficients of all explanatory variables are fixed across regions.

In the last stage, assuming that the decision of applying for a master's degree could vary across regions depending on the field of study—STEM (science, technology, engineering, math) and non-STEM, random slope models have been estimated, allowing for both the intercept and the coefficient of field of study to vary randomly across regions, also including cross-level interactions.

In a random slope model, a group-level random term uj has been included as a linear predictor of the model.

$$\log \left(\frac{{\pi }_{{ij}}}{1-{\pi }_{{ij}}}\right)={\beta }_{0}+{{\beta }_{1}\cdot {X}_{{ij}}+{\beta }_{2}\cdot {X}_{j}+{\beta }_{3}\cdot {X}_{{ij}}\cdot {u}_{1j}+u}_{0j}$$

\({X}_{{ij}}\cdot {u}_{1j}\) is a new term to the model, 0 is the subscript for the intercept residual, random effects \({u}_{1j},\,{u}_{0j}\) are normally distributed with the variances \({\sigma }_{u1}^{2},\) \({\sigma }_{u0}^{2}\) and the covariance \({\sigma }_{u01}\).

The extension from random intercepts to random slopes has introduced two new parameters to the model—\({\sigma }_{u1}^{2}\) and \({\sigma }_{u01}\) carrying out a test of the null hypothesis that both are equal to zero.

Also, regions showing an above-average positive relationship between the field of study and master enrolment intention will have \({u}_{1j}\, > \,0\), while regions with a below-average positive (or possibly negative) relationship between the field of study and master enrolment intention will have \({u}_{1j}\, < \,0\).

In order to test whether the effect of STEM compared with non-STEM fields varies across regions, a likelihood ratio test was applied, taking into account the difference in the log-likelihood values between the model with and without the random slope on STEM.

Empirical results and discussion

Results of the estimated models

Concerning the predisposition phase in which students decide on pursuing a higher-level education programme, 53.6% of the total students report that they have the intention to enrol in a master's programme in the next year. Such intentions are more prevalent among STEM students (62%) than non-STEM students (45.2%). On the other hand, the Kruskal–Wallis test indicates statistically significant differences in this intention among students from different regions or universities. The empirical results indicate that students from the Bucharest region display the highest propensity to enrol in master's education.

The two-level model was used to allow for correlation between master enrolment intentions of individuals in the same region and to explore the extent of between-regions variation. The empirical results of the random intercept model with no explanatory variables revealed that the multilevel approach is suitable and estimated the log-odds of enrolment for an ‘average’ region to the value of 0.145. With a standard error of 0.086, we estimate the between-region variation of the log-odds of enrolment in a master programme at 0.094%. Also, the empirical results of the Wald test proved that there is a statistically significant variation between regions regarding the share of those applying for a master's programme. Based on the value of the between-regions variance (0.094), the variance partition coefficient (VPC) highlighted that 2.77% of the residual variation in the propensity of enrolment in a master programme is attributable to unobserved regional characteristics, indicating that almost 3% of the variance in applying to a master programme can be attributed to differences between regions. When exploring the university characteristics, the empirical results revealed the between-university variance of the log-odds of enrolment in a master programme estimated as 0.81 with a standard error of 0.50, and the Wald test pointed out a significant variation between universities in the proportion of those applying for a master programme. Based on the value of the between-university variance (0.81), the variance partition coefficient (VPC = 19.75%) indicated that almost 20% of the variance in applying to a master programme can be attributed to differences between universities. Thus, we have developed a two-step iterative procedure, building firstly a model only with individual level characteristics and then incorporating both level 1 and level 2 indicators, as well as interaction effects.

Table 2 reports the results of the random intercept models that only include the individual level variables, namely student-level predictors for model I, as well as university-level predictors for model II. The empirical results for the individual level variables pointed out the lack of significance for the association between gender, age, subjective income, expected full-time wage for a person who graduated a master degree or perceived share of unemployed with master’s degree. On the other hand, variables such as average grade, higher education of the father, seniority, or the type of working contract significantly influence the decision of applying to a master programme. Those with higher grades are significantly more inclined to apply for a master's degree, and so are students whose fathers graduated from higher education and students who work on a full-time contract. Additionally, students from rural areas and those with longer working experience have a lower propensity for enroling in master’s degrees. Perceived benefits associated with the graduation of a master's education in terms of wages and employment have no significant influence on the intention of enrolment.

Table 2 Multilevel mixed-effects logistic regression models of the students’ enrolment decision to a master programme.
Full size table

Considering the variation induced by university-level factors, Model II shows that students enroled in non-STEM fields are less inclined to apply for a master's programme compared with those studying STEM disciplines, but the coefficient suffers from a lack of statistical significance at this point. Also, in the case of this model, the empirical results did not support the hypothesis that students decide to enroll in a master's program influenced by university performance.

Table 3 reports the results of the random intercept models that include the individual level variables (students and university characteristics) for model I and also region-level predictors for models II and III, together with the cross-level interaction terms. The findings of the random intercept model incorporating the individual level characteristics revealed that STEM students have a significantly higher predisposition to enrol to master education than non-STEM students. Also, the empirical findings indicate the positive influence of the university performance on the decision of students to enrol in master programmes. With respect to the individual factors introduced in the model, the results confirm that students with higher grades, those working full time and those whose fathers graduated higher education are more likely to apply for a master’s degree. On the other hand, students who have been employed for a longer period are less likely to pursue a graduate degree. The statistical significance of the above-mentioned individual variables preserved in all models. On the other hand, the intention to apply to a master’s programme remains unaffected by gender, age, residence area, subjective income, perceptions of the full-time wage of people with a master’s degree, or the percentage of unemployed people with a master’s degree.

Table 3 Multilevel mixed-effects logistic regression models (random intercept models) of the students’ enrolment decision to a master programme.
Full size table

Analysis of residual level-2 region effects (with only the individual characteristics-model I) supports the hypothesis that there are important differences among regions. North-West and Soth-West are the regions with the lowest probability of applying to a master programme (largest negative values of uj) for which the confidence intervals do not overlap with 0, indicating that they have a significantly lower probability of enrolment than the average region. At the upper end, Bucharest-Ilfov and Centre are the regions with intervals that do not overlap with 0 with the highest response probability (largest positive values of uj), indicating a significantly higher probability of applying to a master programme compared with the region average.

Considering the regional context, model II indicates that the percentage of businesses that introduce product innovations has a detrimental effect on the choice to apply to a master programme. It appears that a more innovative corporate environment might provide alternative incentives for students making them less likely to enrol in master’s programmes.

In addition, model III analyses cross-level interactions between the field of study and the intensity of business innovation. Results show that STEM students are more impacted by the share of enterprises introducing product innovations than non-STEM students when deciding whether or not to apply to graduate school. On the other hand, the percentage of enterprises that introduce process innovations and enterprises that introduce product and process innovations exhibit no influence on the intention to pursue a master’s program.

Until now, we have found that the intention of enrolment in a master programme depends on several student, university, and regional innovation characteristics and this was achieved by allowing the models intercept to vary randomly across communities in random intercept models. We have assumed that the effects of individual characteristics are the same in each region, i.e. the coefficients of explanatory variables are fixed across regions.

In this stage, we extend the random intercept model, allowing both the intercept and the coefficient of one of the explanatory variables to vary randomly across regions, making the assumption that the probability to apply to a master programme could vary from region to region depending on STEM vs. non-STEM majors. Random slope models with both individual variables and regional level variables together with cross-level interaction have been estimated and the empirical results are presented in Table 4.

Table 4 Multilevel mixed-effects logistic regression models (random slope models) of the enrolment decision of students to a master programme.
Full size table

Thus, the model of master’s programme enrolment intentions was updated to compensate for variances in STEM vs. non-STEM disparities among regions. It is assumed that the only variation in the association between the field of study and region is in the difference between STEM and non-STEM in this model, which allows for various probabilities of master enrolment for different fields of study (as in the random intercept model above). It is calculated for Model I with just level 1 individual variables that the coefficient of the field of education (the difference between STEM and non-STEM) is 0.835 + u1j in the corresponding region j. STEM field has a random coefficient, which suggests that the variance between regions relies on the field of education (STEM vs. non-STEM). Master programme enrolment chances differ between STEM and non-STEM fields by 0.062 and 1.29, with the intercept variance interpreted as the between-region variation in log-odds, respectively. As a result of the negative intercept-slope covariance estimate, it may be concluded that regions with an above-average likelihood of master’s degree enrolment (intercept residual u1j > 0, slope residual u1j < 0) are likely to have lower-than-average impact on STEM field. Based on the LR test, where the null hypothesis of no region variation in the difference between STEM and non-STEM students was tested, we may infer that the gap between the two educational fields does indeed change between regions. The difference between communities is now calculated as follows:

$${{\rm{Var}}}\left({u}_{{oj}}+{u}_{1j}{x}_{{ij}}\right)={\mathrm{var}}\left({u}_{{oj}}\right)+2{x}_{{ij}}{\mathrm{cov}}\left({u}_{{oj}},{u}_{1j}\right)+{x}_{{ij}}^{2}{\mathrm{var}}\left({u}_{1j}\right)={\sigma}_{u0}^{2}+2{\sigma }_{01{x}_{{ij}}}+{\sigma}_{u1}^{2}{x}_{{ij}}^{2}$$

=0.062–0.215\({x}_{{ij}}\) + 1.29\({x}_{{ij}}^{2}\) which because STEM students (xij) can only take values of 0 and 1, simplifies to: 0.062 for STEM = 0 and 1.137 for STEM = 1. Therefore, between-region differences in the intention of applying to a masters' programme are greater for STEM students, while regional variation for non-STEM students is lower.

As in the case of random intercept models, STEM students have higher propensity for enroling in masters' education. Also, the intention of applying to a master's program is influenced by academic performance, the higher education of the father, work seniority, and a full-time working contract (Model I), and the significance of these variables is preserved in all models.

Adding and testing the region variables show that the share of enterprises introducing product innovations negatively impacts the decision of applying to a master programme (Model II). Thus, regions with higher shares of innovative enterprises are characterised by a lower propensity of students to enrol in masters' education. A more innovative business sector discourages Romanian students to prolong their educational careers, offering attractive incentives to enter labour market after bachelor’s degree.

Moreover, the effect of the proportion of enterprises introducing product innovations affects the decision of applying for a masters' programme more effectively for STEM students compared with non-STEM students (Model II). For an increase in the proportion of enterprises with product innovation, the effect positively depends on the field of education and the effect will be higher for STEM students. So, the influence of a more innovative business context is higher among STEM students than non-STEM students. This cross-level interaction between the major of the students and the innovation intensity from the regional level indicates why and how the differences between STEM and non-STEM students in terms of master enrolment are not similar across regions.

Discussion of the results

Choice of enrolment in master education

This study supports recent educational choice models that include along with the insights from the human capital theory, cultural and social capital embedded in individual characteristics and background (Perna, 2006; English and Umbach, 2016). Thus, our results confirm the influence of various individual-level factors shaping the decision of prolonging the educational career. Similar to the findings of other scholars (Latiesa, 1989; Mullen et al., 2003; English and Umbach, 2016; Zamfir et al., 2021), we found that students with higher academic performances are more interested to continue their education with a master degree, suggesting that the grades’ level influences perceived academic self-efficacy and expectations regarding future academic success, encouraging or discouraging students to continue their education. With respect to cultural capital, our results point out that higher education attainment of the father is associated with a higher propensity of enrolment in a master's programme. This is according to the theory of Bourdieu (1977) linking educational success to the possession of embodied cultural capital, which determines cultural and social reproduction across generations. With respect to institutional factors, we found that the performance of universities influences the intentions of students to pursue master programmes. It seems that students in universities with higher performance are more satisfied and more inclined to enrol in master programmes.

Furthermore, our results provide evidence supporting the influence of predictors derived from the human capital theory, with some particular aspects that appear in relation to master education. First, our expectations concerning the influence of traditional predictors, such as the perceived benefits upon completion of the education programme were not confirmed. Instead, variables related to the individual demand for specialised skills were found to significantly influence the predisposition of bachelor students to enrol in master education. We found that students working full-time are more inclined to apply for a master's programme than those not working. As attending master programmes is a way of acquiring specialised skills, those working are those who can use such skills and benefit from them. The demand for specialised skills is higher among those who work full time, suggesting that students who work expect higher returns from pursuing a master’s degree than those not working. The latter are less likely to take the risk of accumulating more education for future gains than those who are full-time employed. Master programmes are seen to bring higher returns for insiders on the labour market, rather than for outsiders. This is consistent with the theory that “insiders” often enjoy better employment opportunities than the “outsiders” (Lindbeck and Snower, 2001), allowing them to benefit more from continuing their education at master-level. On the other hand, seniority is associated with a decrease in the predisposition of applying to a master programme. As individuals accumulate experience in the workplace, they are no longer in demand for acquiring specialised skills. So, we can consider that master programmes are seen as providing specialised skills required on the labour market, necessary for those who didn’t acquire such skills by longer working experience.

Regional variation in the influence of STEM and non-STEM majors on the choice of enrolment in master education

First, the current study confirms that predisposition for pursuing master’s degrees is influenced by the field of education. We found that, generally, STEM students are more interested in master’s degrees than non-STEM students. It suggests that both the demand for highly specialised skills and advantages obtained by master graduates vary in relation to STEM vs. non-STEM fields (Lee et al., 2020).

On the other hand, we find evidence indicating significant regional variability in the intentions of enroling in master programmes. This is in line with the idea that returns to education vary across regions within a country (Backman, 2013) as the gains from education are determined by the local labour market (Combes et al., 2008). According to the new economic geography theory (Krugman, 1991), core regions provide higher returns than other regions. Our results confirm this perspective and show that intentions of enrolment in master education are higher in large regions with intensive economic activity, such as the Bucharest-Ilfov region, which includes the capital of the country.

More importantly, our results show that the local context is not independent of the field of education when shaping students’ educational choices. There is a significant regional variation in the difference between STEM and non-STEM students. According to our findings, between-region differences in the intention of applying to a master programme vary in relation with the field of education, showing the complex way in which local conditions interplay with graduating STEM vs. non-STEM fields. So, differences between the propensity of STEM and non-STEM students to enrol in master education varies across local contexts.

More exactly, between-region differences in the intention of applying to a master's programme are greater for STEM students than for non-STEM students. This is consistent with the concept of constrained choice, reflecting the way structural factors interplay with individual decision-making and influence the educational pathways of students (Kurlaender and Hibel, 2018). So, our results suggest that decisions of STEM students in relation to master education are more sensible to the local context factors, while choices of non-STEM students are more invariant across regions. As a result, the difference between the intentions of STEM and non-STEM students varies in relation to the local context, further confirming the theory that core regions provide higher returns to the accumulation of master education than other regions.

The role of the business innovation context in shaping predisposition for master education of STEM and non-STEM students

The empirical results confirmed the influence of local-level factors and showed that the intensity of the business innovation is of relevance. We found evidence that more innovative business contexts are less conducive to the enrolment of students in master programmes. Our results suggest that students living in regions with more developed business innovation display a lower propensity to pursue master programmes, probably due to higher foregone earnings. From this point of view, more innovative economic environments act as a pull mechanism for bachelor graduates, preventing them to prolong their educational careers.

Moreover, innovative business environments discourage more STEM students from further continuing their education. These students are probably those who expect higher immediate earnings in enterprises introducing product innovations, preventing them from further enrol in master education. On the other hand, non-STEM students are less influenced by the regional innovation conditions and their intentions to pursue master education are more invariant across regions. Our findings suggest that forgone and future earnings of non-STEM students are more similar across regions with different levels of business innovation intensity. This suggests that immediate earnings available to non-STEM graduates are less fuelled by the local innovation ecosystem.

So, in general, after controlling for various individual and institutional factors, STEM students are more interested in pursuing master's education than non-STEM students, suggesting that they anticipate higher returns when accumulating master-level education. However, STEM students’ intentions are more influenced by the regional innovation conditions as they are more discouraged to enrol in master education than other students by more innovative business contexts. It seems that such contexts provide immediate attractive incentives for STEM bachelor graduates, preventing them from further investing in master's education. Individuals decide to further accumulate human capital as long as they anticipate that future additional earnings are higher than the direct and indirect costs of continuing education. From this point of view, our results suggest that STEM bachelor graduates in innovative business contexts are discouraged from continuing formal education due to higher immediate earnings. On the other hand, enterprises with a higher propensity to innovation are also the ones providing more employer-funded training (Toner et al., 2004). This could represent an alternative way to acquire specialised skills, replacing the demand for master's education.

The proposed model should be further tested on more comprehensive sets of data, covering more variate educational and economic contexts. Regarding the relevance of our findings to other contexts, one has to consider the economic, cultural, and educational diversity of different regions. For instance, countries with more developed innovative ecosystems and strong links between education and industry may display a more positive influence of business innovation on participation in master education. On the other hand, the influence could differ in regions where the innovation ecosystem is developing or educational systems are less connected with the industry. Acknowledging the study’s geographic limitation is important. While the results provide valuable insights into the relationship between master education and regional innovation conditions in Romania, these findings may not be relevant to other contexts without considering differences in education and economic systems. This limitation points to the need for localised studies or comparative research addressing similar research questions. Future research that can address such limitations and explore the model’s relevance in varied contexts can include cross-country comparative studies.

Final reflections and policy implications

Educational decisions are an important topic of study for education research. The results of this study are in line with recent educational choice models that include individual, institutional and economic characteristics among factors shaping decisions of enrolment in master education (Perna, 2006; English and Umbach, 2016). Consistent with Bourdieu's (1977) theory, we found that parent education and academic performances predict the propensity towards master education, confirming the results of previous studies (González and Dávila, 1998; Latiesa, 1989; Jiménez and Salas-Velasco, 2000; Mullen et al., 2003; Perna, 2004; Xu, 2014). Also, according to our results, university performances shape the intentions of students to pursue master’s degrees, supporting conclusions of other studies (Schapiro et al., 1991; Hartman and Schmidt, 1995; Kallio, 1995; Zhang, 2005). In addition, we provide evidence that the expected benefits of a master's education are higher among full-time workers, especially at the beginning of their careers.

On the other hand, our study complements the literature in the field of master’s degree attainment in STEM education. With respect to differences across fields of study, we show that STEM students are more interested in master’s degrees than non-STEM students. This research covers a knowledge gap related to the extent to which differences between STEM and non-STEM majors are influenced by local contexts and economic conditions. More exactly, the proposed model and our empirical results provide a better understanding of the local context's influence on the educational choices of STEM and non-STEM bachelor students. By employing a multi-level model, we confirm that educational choices are shaped by a dense combination of factors, including individual, university and local level factors. Moreover, we show that the influence of local circumstances depends on the individual-level factors. Local conditions regarding business innovation influence educational choices differently in relation to the field of education. STEM students’ propensity to enrol in master’s degrees is more influenced by the innovation environment than other students.

Understanding how individual, institutional, and contextual factors influence the intention to pursue master’s degrees can be beneficial for improving STEM master’s level programmes efficacy. Our study allows the formulation of several recommendations and implications for STEM higher education policy and practice, including for the widening participation agenda.

First, our study confirms that, in general, the propensity to a linear transition from bachelor to master’s degrees is higher for specific groups of students, such as students with better academic performances, those from families with higher educational attainment, students working full time and those at the beginning of their working history. From this perspective, universities need to find mechanisms for enhancing the access of students with lower grades or from less educated families to master-level education and to adapt the way such programmes are delivered to the needs of working students in early career stages. In particular, financial support schemes could be beneficial for students interested to pursue master's education but are discouraged by the opportunity cost of not entering the labour market immediately. Such support schemes would include scholarships and special loan conditions available for students from less advantaged backgrounds. Moreover, master programmes are expected to provide specialised skills that are used in the workplace. Thus, universities need to enhance their link with the world of work and design master programmes that are closer to the skills demands of the companies. Educational institutions should revise curricula to increase their relevance and strengthen partnerships with the industry to provide flexible, relevant learning opportunities and practical work experience for students, matching education with the demands of business environments. For instance, collaborative programmes between businesses and universities could provide opportunities for students to acquire practical experience through work-based learning while still enabling students to pursue master-level education. Such an approach combines the benefits of immediate job placement with ongoing education. Also, by improving the overall quality of their educational process, universities can expect to retain more bachelor graduates in their master's programmes.

Second, our results show that STEM bachelor graduates anticipate unattractive net benefits from pursuing master's education in more innovative business contexts, probably due to higher forgone earnings. This conclusion is consistent with the idea that highly innovative local contexts attract highly skilled people and talents to a greater extent (Toner, 2011). In such dynamic innovation landscapes, STEM educational institutions need to strengthen their synergy with the local business sector and improve opportunities for master students to work while studying. Thus, providing incentives and developing collaborative structures between universities and the business environment could balance immediate job opportunities with increased long-term returns of continued education for STEM graduates in innovative regional contexts. For instance, tax incentives granted to enterprises that stimulate employees to pursue master's education through funding or paid leave could help mitigate the trade-off STEM graduates face between full-time immediate employment in innovative enterprises and continuing education. In addition, more flexible learning pathways would allow STEM students to engage with the industry while pursuing master's studies. This could include part-time study options, industry placements as part of the learning programmes, or projects in collaboration with local businesses. In particular, expanding dual education within STEM master programmes would be beneficial for retaining bachelor graduates and improving the capacity of STEM education to respond to local skills demand. In addition, career guidance services should be enhanced to better inform undergraduate students about the long-term returns of master's education versus the immediate benefits of entering employment. This guidance should be tailored to the specific context of the students’ major and regional innovation environment.

We conclude that human capital theory continues to provide a valuable framework for understanding educational choices, especially in the case of STEM fields. Future research will focus on longitudinal studies to track STEM and non-STEM graduates’ long-term career outcomes. This would offer insights into the relevance of masters' education for the skills demands of local industries. Comparative studies across different regional innovation ecosystems can also shed light on how specific local conditions allow graduates of various fields of education to benefit from pursuing masters' education. From the methodological perspective, this study highlights the importance of robust approaches to understand the complex dynamics between education, career choices and local economic contexts. Future research should consider mixed methods designs that combine quantitative analysis with qualitative insights from students, educational institutions and industry stakeholders. This would offer an in-depth understanding of graduates’ motivations, barriers, and opportunities.