The first possible differentiating effect for RSE is the training level. Self-efficacy beliefs stem from mastery experience (Bandura, 2006). Hence, it can be reasoned that the amount of methods training affects RSE, as expanded methods training allows for extended mastery as well as failure experience—a differentiating effect in student self-efficacy. Such experiences of mastery or failure are particularly prevalent in hands-on training settings. Oftentimes, methods training in Bachelor programs is rather theoretical, with hands-on experience increasing in graduate training. Still, any investigation of RSE development and potential differentiation effects should cover all levels of higher education.

First, apart from mastery experience, self-efficacy is affected by epistemological beliefs and the value a person ascribes to the respective domain (Bandura, 2006), namely research. These two factors are probably stressed in undergraduate training already. Once in specific methods classes, but also in subject-matter classes by communicating the relevance of research for theory development. Second, undergraduate research experiences have been employed increasingly over the past years. The heft of this development is mirrored by the recent publication of the Cambridge Handbook of Undergraduate Research (Mieg et al., 2022) which provides an overview of theoretical approaches as well as practice examples in diverse disciplines and from countries across all continents.

Consequently, from their first semester on, students can be exposed to forces that shape their RSE. These forces may increase and specialize as training advances from undergraduate to graduate training. The amount of training might not only affect mean levels of RSE (Livinţi et al., 2021) but also engender differentiation and specification of self-efficacy beliefs regarding the various research tasks and the way they constitute students' self-efficacy in this domain. As a consequence, variation in construct structure might be interpreted as a conceptual change of research in the process of methods education. For example, Rochnia and Radisch (2021) argue that in educational contexts, learning implies change over time (hopefully), both regarding the mean level and construct structure. Thus, measurement non-invariance between training levels does not necessarily indicate low measurement validity (i.e., a bug), but rather a differentiation of a concept due to learning (i.e., a feature; Putnick and Bornstein, 2016; Rochnia and Radisch, 2021). Delineating the amount of measurement (in-) variance might help to understand the intra-individual differentiation of research as a specific domain of self-efficacy across the training level.

A second possible differentiating effect is the research culture in an academic discipline. Although most university training is comprised of research classes, their focus on specific methodology may vary between disciplines, partly due to differences in their genesis and research targets. For example, psychology and educational science are both empirical social sciences and share some overlap (e.g., educational psychology). However, even disciplines that appear akin at first sight may differ regarding their methodological gist. In German academics, a traditional focus on qualitative methods in educational science still holds strong today and qualitative methods are employed and refined in current research, whereas psychological research employs a wider range of rather advanced quantitative methods. Consequently, most German undergraduate methods training in psychology focus almost exclusively on quantitative methods, while methods training in educational science stress qualitative methods and vary much more across universities regarding the extent of undergraduate methods training. These and other potential disciplinary differences may foster different value perceptions and epistemological beliefs regarding research in students, which affect their self-efficacy beliefs (Bandura, 2006; Mason et al., 2013).

One may then ask how broadly RSE can be assessed invariantly across disciplines such as psychology and educational science? Otherwise put, do differences between disciplines evoke differentiated development of RSE in their students?

Disciplines might not only differ in the actual tasks or research methods but also in the emphasis given to a discipline's methods. As a consequence, if methods education is emphasized stronger in one discipline than in another, differences between disciplines might increase over time as students are socialized into the research culture of their respective programs. This might lead to an interaction effect of discipline and time in the program, as socialization and differences in research-specific mastery experiences increase over time.

This study was conducted in Germany with a German-speaking sample.¹ Thus, same-construct measures used for convergent validity estimation should ideally be worded and validated in German so that they provide a legitimate anchor for the ASER items. Unfortunately, validated measures of RSE are missing at the Bachelors' and Masters' levels in the German-speaking realm. Lachmann et al. (2018) developed an instrument that assesses RSE in Ph.D. graduates and targets respective attainments that are not adequate for Bachelors' and Masters' levels (e.g., “I can build cooperations with central researchers in my field”). In younger students, RSE has mostly been included rather as a side note with unvalidated post-hoc items, or with a divergent structural focus (e.g., Gess et al., 2018; Pfeiffer et al., 2018). Still, items from these two studies seem to be the best option in selecting convergent validity items: They do assess RSE specifically and they were developed to assess German students at a pre-doctoral level.

Estimating the discriminant validity of the ASER requires constructs that are theoretically and empirically related to, but still distinct from RSE (Clark and Watson, 2019) to prevent a jangle fallacy.² Those constructs are general self-efficacy (Bandura, 2006) and academic self-concept. While self-concept is rather general and past-informed, self-efficacy is more task-specific and future-oriented. However, both target self-beliefs (Choi, 2005). A further theoretically and empirically positive relation to RSE is given by attitudes toward research (Kahn and Scott, 1997; Livinţi et al., 2021).

Relations between RSE and neuroticism have not yet been tested empirically, but can be reasoned to be negative, as low self-esteem and a heightened sensitivity toward failure (Zhao et al., 2010) impede mastery experience, which is a core requisite for acquiring (research) self-efficacy (Bandura, 2006). Indeed, negative relations with a neuroticism that have been found for academic self-efficacy (Stajkovic et al., 2018) support this hypothesis regarding research self-efficacy.

Research self-efficacy construct validity is also supported by a lack of relation with constructs that are theoretically unrelated to RSE. Correspondingly, several authors report non-significant near-null correlations between agreeableness and academic self-concept (Marsh et al., 2006) or self-efficacy regarding college credits (De Feyter et al., 2012). Theoretically reasoned, the tendency to help others or not to be harsh would not affect a person's confidence in performing research tasks. Thus, based on theoretical reflections and considerations of empirical findings on the relation between agreeableness and other self-efficacies, a lack of relation between RSE and agreeableness would support RSE construct validity.

Test bias regarding gender will be explored in order to delineate whether the complete sample can be used for our analyses. Even today, the ratio of female and male students presumably differs between programs. It is, therefore, important to detect potential test bias before all students are analyzed in a common model. Meta-analytic results on gender mean differences in academic self-concept show an overall effect size of 0.08 favoring male students, specifically regarding mathematics and social-sciences self-efficacy (Huang, 2013). In contrast, in a meta-analysis on the relation specifically between research self-efficacy and gender, Livinţi et al. (2021) report non-significant differences between male and female students. However, those (non-) differences refer to manifest mean scores and can only be validly interpreted in unbiased measurement.

In addition, if MI can be established between female and male students, this indicates that the ASER is generally fit for valid invariant measurement. Following, a lack of MI between certain groups, as tested under the differentiation hypothesis, can be attributed to real group differences in measurement and probably does not stem from the weakness of the instrument.

H1—Convergent validity: Satisfactory fit of a comprehensive measurement model of RSE including items from different RSE measures is expected.

H2a–e—Discriminant validity: RSE will display moderate to high correlations with general self-efficacy, academic self-concept, research attitudes, and neuroticism (negative relation). The relation between agreeableness and RSE will be non-significant.

H3—Construct validity: Convergent validity coefficients will be higher than discriminant validity coefficients.

H4—Test bias: MI is expected between female and male students. Results will indicate whether further analyses should be conducted for both genders together or separately.

H5: Limited measurement invariance is expected between different training levels.

H6: Limited measurement invariance is expected between psychology and educational science students.

H7: An interaction effect of training level and discipline on the level of MI is expected.

The two-factor structure reported by Klieme (2021) could not be confirmed in this study. Here, the ASER items were better represented by a non-hierarchical three-factor model [χ(149)2 = 711.746, p = 0.000, CFI = 0.911, RMSEA = 0.083, SRMR = 0.048] which can be inspected in Table 1. This overall only moderate fit might be explained by non-invariance between different sub-groups, which renders a very good model fit impossible when the whole sample is analyzed together. Thus, differentiating effects between certain groups are still worthwhile to be analyzed, even if the overall model fit with data comprising all these potentially heterogenous groups indicates some issues. Consequences hereof will be addressed in the following MI analyses and the discussion. The three RSE sub-factors refer to Theoretical Aspects of research (nine items, e.g., “Generate researchable questions” or “Write a discussion section for my thesis”), Technical aspects (seven items, e.g., “Collect data” or “Know which data analysis method to use”), and Interpretation (three items, e.g., “Interpret and understand the results of my research”). The 12 convergent validation items were best represented by individual factors corresponding to their origin of publication (Gess et al., 2018; Pfeiffer et al., 2018, respectively).

Construct validity was investigated by a comprehensive model, comprising a measurement and a structure model (see Figure 1). RSE was modeled as a second-order factor with five sub-factors: three ASER sub-factors and one sub-factor for each set of validation items (Gess et al., 2018; Pfeiffer et al., 2018). In the structure model, discriminant constructs were modeled as one-factorial. One exception is attitudes toward research, which shows a hierarchical structure with a positive and negative attitudes sub-factor. Model fit was again moderate to poor [χ(429)2 = 4,539.186, p < 0.001, CFI = 0.838, RMSEA_CI95% = (0.056; 0.058), SRMR = 0.059]. Still, the loadings of the five sub-factors on the RSE super factor were all similar in size (see Figure 1). Thus, the ASER items represent RSE similarly to the validation items (H1a). Latent factor correlations between the ASER and each validation factor 1 and 2 were 0.92 (p < 0.001) and 0.86 (p < 0.001), respectively (for factor naming see Figure 1).

The relations in the nomological net hypothesized in H2 were confirmed, supporting the positioning of RSE in the nomological net. Validity coefficients are shown in Figure 1. Corresponding to H3, correlations supporting convergent validity were stronger than those supporting discriminant validity.

Measurement bias was tested through MI analyses using MGCFA. MI Interpretation was based on differences in fit statistics between the increasingly constrained models, with cut-offs of ΔCFI < 0.01, ΔRMSEA < 0.015, and ΔSRMR < 0.015 as suggested by Chen (2007). In addition, chi-square difference tests were considered in case of unclear results. However, since they tend to be overly conservative (Putnick and Bornstein, 2016), they were included with caution.

Subjects who identified as non-binary were excluded, because this group was too small (N = 23), as were seven subjects who did not indicate any gender. For the remaining 524 participants (183 male students and 341 female students), the ASER three-factor model displayed satisfactory fit in the MGCFA configural model [χ(149)2 = 699.938, p < 0.001, CFI = 0.908, RMSEA = 0.084, SRMR = 0.049]. Strict measurement invariance between female and male students was found, meaning that both groups displayed non-invariant factor loadings, item intercepts, and residual variances (for model statistics see Supplementary material B). Thus, the latent factor scores can be compared between genders (see Table 2). Differences in mean factor scores between female and male students were significant for the Technical Aspects factor only (p = 0.016) with manifest mean values of M = 5.25 (female students) vs. M = 5.60 (male students). The mean factor score differences were non-significant for the Theoretical Aspects factor and the Interpretation factor.

Overall strict MI was found between psychology and educational science students, that is non-invariant factor loadings, items intercepts, and residual variances. For the sub-factors, MI was scalar (equal factor means) for the Theoretical Aspects and strict for the Interpretation factor. Thus, means can be compared between both disciplines on these two sub-factors (see Table 2). For the Technical Aspects factors, configural MI indicated that the latent meaning of this factor is different for psychology and educational science students due to differences in factor loadings. Manifest scores should, thus, not be used for group comparison. The mean factor scores were significantly different for Theoretical Aspects (p = 0.017, manifest M = 6.13 and M = 5.79, respectively) but not for Interpretation (manifest M = 6.30 for both groups). Concluding, MI results between both disciplines varied for the three sub-factors.

Interaction effects occurred between discipline and training levels. When separating Bachelor and Master students, MI between disciplines changed with the training level: it decreased for the Theoretical Aspects factor but augmented for the Technical Aspects factor (see Supplementary material B5, B6). Viewed from the other angle, MI between Bachelor and Master was lower in the psychology than in the educational science sub-sample. This means that the differentiating effect of discipline on RSE latent structure is moderated by training level, or else, the differentiating effect of training level is stronger in psychology students.

The results of MI analyses between psychology and educational science students were mixed. At least scalar MI was found for the Theoretical Aspects and the Interpretation factor, but not for the Technical Aspects factor. Thus, RSE appears to constitute slight differences in meaning across disciplines, which was especially apparent in research tasks like processes of data collection or knowledge of appropriate analysis methods.

Concluding, while RSE should be of importance in almost every academic program, the specific research and/or training cultures apparently differ between academic disciplines. The difference in mean scores on the Theoretical Aspects factor also underscores the importance of considering disciplines when researching RSE. It seems that differences do, indeed, occur both on structural and mean score levels, as well as regarding a variety of research tasks.

It would be interesting to investigate the reasons for these disciplinary differences in the perception of RSE. Do the main research practices differ across disciplines, and are they communicated differently to the students? Is the nature or the role of research perceived differently by students? Are research methods regarded and taught differently? In addition to psychology and educational science that were investigated in this study, taking into account further disciplines even beyond the social sciences might be insightful with respect to the investigation of disciplinary differentiation effects on RSE development.

Measurement invariance analyses between training levels, again, yielded mixed results. MI was scalar for the Technical Aspects and Interpretation factors, and metric for the Theoretical Aspects factor. Local structural equation modeling provided a more detailed insight due to the continuous moderation of model parameters: Significant variations in fit statistics, item loadings, and item residuals across semesters were revealed. These findings indicate a change in the meaning of the latent construct of “research self-efficacy” (Molenaar et al., 2010). Most importantly, differences in model fit may indicate that not even configural MI (relating to the basic construct structure) can be validly assumed across all training levels: not only the meaning of each latent factor but also the amount and segmentation of RSE factors might differentiate with time at university.

Concluding, RSE changes with the training level on two levels: First, change in mean values can be reasoned to stem from increasing mastery/failure experience (Bandura, 2006). Second, factor differentiation can be attributed to conceptual change in how research is regarded (Rochnia and Radisch, 2021).

For an understanding of RSE development covering the whole qualification period of junior researchers, future studies might extend the investigation of the differentiation hypothesis to the doctoral level. However, different from other countries, doctoral qualification in Germany is not standardized: while there are a few distinguished doctoral training programs, most PhD candidates complete their PhD research within the context of common employment as assistant researchers at universities without ongoing formal methods education. Thus, Bachelor's and Master's programs are more similar to each other regarding method training, since this is where coursework and institutionalized learning takes place. The ASER was developed to assess RSE in Bachelor and Master students as RSE measurement for this level was missing. RSE at the doctoral level can already be validly assessed by a questionnaire by Lachmann et al. (2018). This measure comprises research tasks that are relevant on the PhD level, but not yet for Bachelor and Master students (e.g., “I can build cooperations with central researchers in my field”). Delineating whether and how these two measures can be employed to capture RSE development over the whole qualification span would be an interesting endeavor for future RSE research.

Four independent MI analyses were conducted to investigate dichotomous interaction effects of training level and discipline on RSE structure. Results indicate that there is an interaction effect: Invariance seems to change with program level, especially so for psychology students. The data support the notion that psychology students at the master level perceive research differently than the other sub-groups. One reason might be differences in emphases on methods education between the disciplines. Presumably, methods education is more emphasized in psychology than in educational science. Even for programs such as school or clinical psychology, where most students plan a career in applied settings, American graduate education aims to also strengthen students' scientific competencies by explicitly implementing the so-called “scientist practitioner approach” (Jones and Mehr, 2007). Klieme et al. (2020) called for a similar focus in educational science programs through research-based learning.

Especially based on the fact that measurement variance increases with the program level, it can be reasoned that these differences are less due to primal interests that influence the choice of program, but due to socialization processes once a program is studied. Future research might analyze both differences in research practices (methodology) as well as the research training environment.

The methodical implications of our endeavor to take a non-traditional perspective on measurement (non-)invariance are the basis for future research on any differentiation of latent constructs. If test developers seriously begin to consider non-invariance in measurement as a feature, strategies are needed to deal with naturally resulting poor model fit in the configural baseline model, and with other procedures that have been employed to judge latent construct measurement under a perspective of desired invariance. Common procedures and cut-off values employed in current MI analyses do the concept of MI as a feature injustice and need to be differentiated as well. In this regard, Rochnia and Radisch (2021) revive the AGB Typology by Golembiewski et al. (1976) that systematizes levels of change in measurement across time. By this means, Rochnia and Radisch (2021) refrain from considering a lack of invariance generally as a bug but emphasize that considering different types of change in measurement may hold relevant information in educational settings. Regarding concrete analysis methods for MI, LSEM (Hildebrandt et al., 2009), and moderated non-linear factor analysis (Bauer, 2017) allow the ability to test changes in the model parameter as effects of certain moderating variables, like semesters at university. However, while both approaches are well fit to investigate changes in meanings and emphases of latent constructs, they are not (yet) able to detect systemic changes in the factor structure between groups. Meaning, both approaches are fit to investigate parameter changes in the same model but are not fit to investigate changes in the model structure (e.g., from a two to a three-factor model). Furthermore, a developmental differentiation of RSE may constitute itself in a growing, more complex factor structure contingent on relevant moderating variables. This will render the test of invariant covariance matrices as suggested by Vandenberg and Lance (2000) relevant again—a step in MI analysis that has been omitted in most MI practice recently (Putnick and Bornstein, 2016). In addition, exploratory methods for structure analyses might be needed that can depict changes in the construct structure across a certain moderator.

The identified variance in construct structure and constitution may explain previously reported heterogeneous results. Factorial differences between different measures of RSE have been pointed out repeatedly (e.g., Forester et al., 2004; Livinţi et al., 2021). However, an explanation for these differences is lacking as of yet, but might be achieved by considering our differentiation hypothesis and examining the respective databases used for instrument development and factor analysis. The empirical structure of the Self-efficacy in Research Measure (Phillips and Russell, 1994) was based on the data from 219 doctoral students in counseling psychology. In contrast, the structure of the Research Self-Efficacy Scale (RSES, Bieschke et al., 1996) resulted from a factor analysis with the data from 177 doctoral students enrolled in various programs, namely biological sciences (32%), social sciences (28%), humanities (23%), and physical sciences (17%). Since our results suggest that students' academic discipline affects the RSE factor structure, the heterogeneous results from these two studies might be reconciled by considering disciplinary differentiation effects. This could be achieved by re-analyzing the factor structure of the RSES separated by discipline. A more feasible approach might be to systematically recognize and investigate the effects of the disciplines that participants are sampled from in future studies.

Furthermore, inconsistencies in the relations of RSE to other constructs have been pointed out regarding gender (Livinţi et al., 2021), research training environment, year in the program, interest in research, and research outcome expectations (Bieschke, 2006). Again, suchlike heterogeneous results may be reconciled by taking a closer look at sample characteristics that affect construct structure and meaning. Specifically, if the meaning of a latent construct, namely RSE, is inconsistent across samples, empirical relations to neighboring constructs may be inconsistent as well. In their recent meta-analysis, Livinţi et al. (2021) point out that student samples in RSE research are pretty heterogeneous, as participants are sampled from different training levels (undergraduate vs. graduate training) and from various disciplines such as counseling psychology, education, STEM education, and even law. Following, inconsistencies in construct relations may be explained under the differentiation hypothesis, since we found differentiation effects on the RSE structure for both of these dimensions, training level and discipline. The exemplary analysis of two studies on the relation between RSE and training environment does, indeed, reveal differences in sample characteristics. A non-significant relation reported by Phillips et al. (2004) is based on a sample of 84 individuals who had already accomplished their full doctoral degree (85%) or all requirements but the thesis. In contrast, a significant relation reported by Kahn (2001) is based on a sample of 149 doctoral students of which 50% were still in their first 2 years of training.

Concluding, in the context of theory development, the lack of MI in RSE measurement should caution us to compare findings on RSE structure, mean scores, or relations to other variables that stem from the data on students with different backgrounds. Similarly, practical implications drawn from empirical RSE research should be given very cautiously with regard to whether the targeted group that implications are drawn for is well-enough represented by the empirical sample. Therefore, future research should give more thought to considering potential sub-groups that have gone unregarded as of yet in order to reveal the differentiated picture of RSE that is needed. To do so, differentiated analyses are called for that are fit to model a lack of MI as a feature.

Considering MI between training levels will foster a process-oriented theory of RSE structure and development. The differentiation hypothesis that we propose in this article might be one unifying factor in RSE theory development. Concrete characteristics that affect RSE mean scores as well as structural differentiation can be deduced from previous research (e.g., Livinţi et al., 2021) and social cognitive career theory (Lent et al., 1994). They constitute person factors like interest and environmental factors like aspects of research training as antecedents to RSE development. Livinţi et al. (2021) identified these antecedents of RSE in the context of social cognitive career theory. The studies included in their meta-analysis, however, exhibit two weak spots: first, they stem from cross-sectional data. Second, they analyze RSE total scores only, without taking into account structural differentiation.

Our results demonstrate the importance of structural differentiation, not only based on main effects but also on complex interaction, for example between the discipline studied and time in the program. Future research should address such potential interaction effects of discipline and the amount of methods training on factor model parameters in more detail. In a way, this interaction describes the culture and aspiration under which research is taught and addressed within a university program. Are students introduced to research methods and research matters in their discipline? Are they given the opportunity for their own (mastery) experience? Do faculty model scientific behavior and attitudes, and do they, thereby, socialize students into a (discipline) specific way to conceptualize and approach research? The theory of the research training environment (Gelso, 1993) stems from an investigation of the science practitioner teaching in U.S. American psychology programs and incorporates those questions.

Indeed, previous researchers have identified these factors of the research training environment as a predictor of RSE in cross-sectional studies (e.g., Kahn and Schlosser, 2010). One study explored the longitudinal effect of the research training environment on changes in RSE after 1 year in graduate training: Kahn (2000) found effects for one training aspect, namely the student–mentor relation. In addition to Kahn, also Livinţi et al. (2021) conclude from their meta-analysis the need for further longitudinal studies. Analyzing students' research training environment might serve as an explanation for the interaction effect of discipline and training level on RSE differentiation. Longitudinal studies can in addition give insight into the intra-individual development of RSE, as well as delineate additional causal effects of possible predictors beyond the training environment (such as general self-efficacy, interest, or other person variables). Understanding causes, interactions, and onsets for this intra-individual change will help refine a theory of RSE development regarding both RSE scores and structural changes.

A fine-grained investigation of structural changes can help clarify how students think about research and how this affects their self-efficacy at different levels. This might help to enhance and customize methods education to the “zone of proximal development.” Again, longitudinal studies may investigate these intra-personal structural changes in RSE as an effect of a specific person and environmental factors.

Developing an understanding of RSE in academic education can serve several goals in higher education practice. On an individual scale, considering students' RSE can support our understanding of their individual needs regarding methods education. Specifically, insights from longitudinal studies will enable educators, faculty, and mentors to identify said “zones of proximal development” in regard to specific research tasks that students should be exposed to. This way, they can tailor methods education to students' needs, focusing on tasks in which students perceive low self-efficacy. In addition, students might use the differentiating facets of RSE for the self-assessment of their own development (Forester et al., 2004), especially in self-regulated learning settings. This can help them delineate what research tasks they should acquaint themselves with next. Hereby, focusing on self-efficacy development, beyond mere research knowledge, will draw both students' and educators' attention toward learning settings that deliberately incorporate its facilitation. Specifically, such settings should emphasize the main factors that influence self-efficacy beliefs. This is to facilitate mastery experiences through student activity and be aware of the communicated perspective on the value and epistemological nature of research. Our results show that the latent structure of RSE differs between students based on their disciplinary socialization. Thus, zones of proximal development are not only contingent on a student's training level, but also on other context factors that need to be considered in individual mentoring. Our results indicate that academic discipline is such a context factor, but systematic investigations of the longitudinal data on individual RSE development are needed to specify and confirm differentiating factors that operate on an individual level.

More generally, including RSE in higher education research may enhance our understanding of the psychological processes that underly successful academic development. Longitudinal studies should, thus, delineate specific factors in the training environment that prove promotive of RSE. Prospectively, these factors will enable faculty to foster students' RSE, following evidence-based reasoning. However, considering the differentiation hypothesis, RSE sensitive practices in methods training can never be simply carried over from one setting to another. Following that, university didactics need to develop approaches to fostering RSE that are customized to different disciplines and training levels. These customized approaches can only be delineated through longitudinal studies that systematically investigate possible differentiation effects that operate on the group level.

Once the role of RSE in higher educational processes is delineated, its assessment might be an asset to the evaluation of research training settings on a broader scale, such as settings that employ research-based learning in Europe (Wessels et al., 2021), or the scientist–practitioner approach in the U.S. (for a description see Jones and Mehr, 2007). Their focus on student activity (Huber, 2014) can be reasoned to particularly enable mastery experience. Therefore, research self-efficacy should be regarded as a core outcome of such student-active learning environments and should, thus, be considered in program evaluation.

Taken together, the findings reported here can add to the conclusion that further investigation of RSE differentiation is needed, worthwhile, and beneficial. Understanding causes and onsets for change in the RSE structure will help refine both, a theory of RSE development and university methods training.