Data and sample
We use data from Germany to estimate the effects of resilience on outcomes. Data from the National Educational Panel Study (NEPS) (starting cohort 5, version 12-0-0), a large-scale assessment of educational trajectories in Germany, are suitable for analyzing the research question posed (Blossfeld et al. 2011).Footnote 2 The sample comprises university students from all over Germany who are regularly surveyed from the beginning of their studies on their study programs, developments and grades, but also on a variety of other areas of life. The NEPS is a suitable basis as detailed information on social background and childhood is available in addition to the prospectively collected data on the trajectories, which is crucial to partial out biasing effects or self-selection. The original sample consists of 17,910 students who began their first degree at a university in Germany in the winter term 2010/2011, regardless of nationality, subject or degree. The sample is clustered by universities and subjects (104 public universities, 108 public universities of applied sciences and 49 private universities) (FDZ-LIfBi 2018). The surveys take place regularly (once a year). Students with a degree in teaching as well as students in private educational institutions are overrepresented by the design in order to provide more accurate information for these relatively small populations. Of the 31,082 persons contacted, 17,910 were finally interviewed after checking the target sample affiliation, which corresponds to a participation rate of approximately 58% (wave 1). The realization rate according to AAPOR (2006) is 85.2% (Steinwede and Aust 2012, p. 38). Overall, starting cohort 5 claims to be representative of persons studying in Germany for the first time in 2010/11. As expected, this figure will decline over time as people leave the panel temporarily or permanently (attrition). The data collection is carried out by Infas (Bonn) and the German Centre for Higher Education and Science Research (DZHW, Hanover). For the following analyses, the data are additionally restricted. In all analyses, only students who were not older than 35 years at the beginning of the data collection in wave 1 will be included in order to obtain a “typical” student sample. For example, the processes described in theory cannot be assumed in the same way for students who are retired and are studying for personal development after completing their professional careers. Moreover, in the longitudinal analyses, only data from the first four waves from 2011 to 2014 are used, since beyond that many variables are no longer available and the sample size would be considerably reduced.
Operationalization and variables
The operationalization of the variable resilience is of central importance for all subsequent analyses. A review paper that identifies and assesses 19 different scales in the literature, but does not recommend a particular operationalization, shows how various resilience can be measured (Windle et al. 2011). A solution that is compatible with the existing data is operationalization using the Big Five Inventory. Here, resilience is generated as a binary variable from three metrically scaled inventory items (Ercan 2017). Accordingly, persons are considered resilient if they have a below-average value for neuroticism and above-average values for conscientiousness and extraversion as one study reports (Campbell-Sills et al. 2006, p. 591). The other components of the big Five inventory are not predictive of the resilience status in a multivariate model and are thus not included in the operationalization. The robustness of this form and similar forms (including all five components of the Big Five inventory) has been empirically proven (Waaktaar and Torgersen 2010). The validity of such a procedure is confirmed by other studies that investigate the connection between Big Five and study success. Thus, a meta-study involving 58 individual studies comes to the conclusion that neuroticism and satisfaction with the course of the study correlate negatively and conscientiousness and performance correlate positively (Trapmann et al. 2007). It can therefore be assumed that postulated correlations are stable overall. The resilience status can be calculated for the year 2012, since the inventory items mentioned are measured at that time. This relatively early measurement, even if not directly at the beginning of the study, is helpful, as it provides the temporal order for a better understanding in which causality flows. The respective inventory items are measured on a scale with nine levels between 1.0 and 5.0 in steps of 0.5 points. Descriptively, about 17% of all students are classified as resilient in 2012.
The variable of social origin is particularly important in order to be able to exclude spurious correlations as far as possible. There are various operationalizations for this, but parental status as a continuous variable seems more suitable than, for example, educational level, which was only measured categorically. The average parental ISEI (International Socio-Economic Index of Occupational Status), which is based on the respective occupation, is used. This scale has the advantage that various aspects such as income, educational level and prestige are taken into account, which allows a differentiated measurement of origin (Ganzeboom et al. 1992). If information is available for father and mother (which is the case for 84.1% of all respondents), the highest value is taken. Otherwise, the only available value was used.
After the central explanatory variables, the dependent variables will be explained. For this purpose, three time-variable outcomes were selected for which information is available in the four waves between 2011 and 2014. The first includes the tendency to drop out of the current study (Trautwein et al. 2007). For this purpose, a new quasimetric variable was generated from five different items with four steps each (Likert scaling) (Cronbach alpha > 0.79, depending on the wave). The original items ask, for example, whether the respondent has already seriously considered discontinuing the study. The relevance of the variable is that it estimates the probability that a study will be successfully completed. Students with a high dropout tendency are likely to give in to this sooner or later, which is of considerable importance for the further course of life. The second dependent variable is the self-reported average study grade in the current semester, measured on a scale of 1.0 to 4.0, with higher values representing better grades. It should be noted that this variable, like all others, was collected in the survey and cannot be verified objectively. This variable is used in an unstandardized way, but we control for the field of study. The relevance of the study grade lies in the measurement of general performance. Better grades are by definition an expression of higher academic performance. The last dependent variable is general life satisfaction, which is generated from a total of six individual items (Westermann et al. 1996). Each item is measured on an eleven-point scale between 0 and 10; higher scores represent higher satisfaction (Cronbach alpha > 0.84, depending on the wave). This variable appears relevant to reflect the general quality of life, which is of central interest to the individual. These items account for various sub-dimensions like “How satisfied are you with… your life/your health/your standard of living/your family life?” to give an impression on various forms of overall satisfaction, which can be integrated into a single combined score.
In order to select all control variables, it is useful to integrate the following analyses into a causal theoretical framework. This means that the quality of empirical research can be improved if the intention to explain causality is explicitly stated (Hernán 2018). (Pearl 2009).Based on the framework as proposed by Pearl, to rule out spurious correlations it is necessary to include all variables as controls that influence both the treatment as well as the outcome simultaneously. It is important to note that it is not the choice of statistical method for identifying causal effects that is decisive, but the selection of the relevant control variables. All those that influence the independent and dependent variables simultaneously must be selected. The following control variables are used for this purpose: Gender, age, place of birth (West Germany, East Germany, abroad), migration status, whether the study was taken up at a university or at a university of applied sciences, field of study (pre-grouped into six categories by the NEPS due to data protection regulations), highest parental ISEI, highest parental educational level, number of siblings, years spent in kindergarten, death of father, death of mother, age of both parents in 2011 and student loan eligibility. The logic of this selection is explained using the example of the parental ISEI: A higher social status goes hand in hand with greater resilience (Schoon 2006). At the same time, social status has an impact on the study situation, since parents with a high social status can provide more financial resources and thus have a positive influence on the housing and living situation during the study period as well as on life satisfaction. If the social status were not included, a spurious correlation between resilience and satisfaction could arise at this point. Most control variables are time constant, such as gender or highest parental education. The only time varying control variables are the field of study and whether the respondent lives with her family at the time of survey). While we are confident that most relevant confounding factors are accounted for, since only observational data are available, it is usually not possible to rule out all confounding. Consequently, we cannot argue that the results represent pure causal effects and regard them as associations instead. The reader can decide whether he or she believes how well these numbers represent causal effects or whether spurious influence might still be present that are not accounted for. All variables of the analyses are presented in Table 1 for a quick overview.
Table 1 Overview of all variables of the study Strategy of analysis
Since all outcome variables were measured at several points in time (panel design), longitudinal models can be applied. It should be noted that the central explanatory variable, the status of resilience, was measured in 2012 and is therefore constant over time. Technically, the following analyses are calculated as multilevel growth-curve models. This estimation procedure is mandatory, as otherwise standard errors would be incorrectly calculated based on correlating observations by the same person. Growth-curve models are chosen since the respondents were surveyed annually and within a time frame of a few weeks. Since there is little variation in the time of survey, other approaches like survival models cannot reveal further insight since these models require a finer-grained time of survey, for example, monthly or better even weekly with a large variation for all participants. Therefore growth-curve was selected, which give very similar results to panel regression models with random effects. Linear models are used for all dependent variables because the variables are scaled continuously (OLS regression technique). The models are built step by step: The first model contains only the explanatory variable resilience, the wave dummy, and their interactions to allow for most flexible estimation of effects for each wave. This form of parameterization is rather uncommon for growth-curve models, however, beneficial for the present data since only a small number of time points is available and all participants were surveyed at the same points in time (Rabe-Hesketh and Skrondal 2012). This design has some beneficial properties: due to the interaction effects, the change or growth of the outcomes is modeled independently for each point in time so that changes in all directions are mapped flexibly. Furthermore, in contrast to regular Growth-curve models, we do not have to make assumptions about the mathematical shape of the curve, for example, linear or exponential. The results are therefore almost free of model assumptions. Hence, this approach adds further flexibility and looses artificial constraints. The second model adds all control variables. For mathematical precision, the model equation for the models without control variables is shown below. The last two terms describe the person-specific (j) and observation-specific (ij) error terms. The variable resilience varies only between persons but not within a person, while the wave dummies vary only within a person but not between persons.
$$\begin{array}{c}{Y}_{ij}={\beta }_{0}+{\beta }_{1}{Resil}_{j}+{\beta }_{2}W{1}_{i}+{\beta }_{3}W{2}_{i}+{\beta }_{4}W{3}_{i}+{\beta }_{5}W{4}_{i}\\ +{\beta }_{6}{Resil}_{j}\times W{1}_{i}+{\beta }_{7}{Resil}_{j}\times W{2}_{i}+{\beta }_{8}{Resil}_{j}\times W{3}_{i}+{\beta }_{9}{Resil}_{j}\times W{4}_{i}+{u}_{j}+{\varepsilon }_{ij}\end{array}$$
We argue that, given the current data, this model design is preferable to inspect the temporal development of outcomes in interaction with the resilience status. Furthermore, graphs can be produced that allow an intuitive understanding of the developments, which is beneficial for interpretation and discussion. All analyses are carried out using Stata 16.1. Missing information was imputed using Multiple imputation by chained equations (MICE; 40 imputations after a burn-in sequence of 80) (Allison 2001; Azur et al. 2011). Typical diagnostic tests such as convergence of imputation models were examined. Graphs are generated using the software package mimrgns (Klein 2014).