Introduction

Children who are from socio-economically disadvantaged family backgrounds, but who manage to overcome the setbacks that such a background can entail and who achieve successful learning outcomes are often referred to as being academically resilient (Agasisti et al., 2018, 2021). Even though the definition appears straightforward, the operationalization and measurement of academic resilience has been challenging (e.g., Rudd et al., 2022). Masten (2001) has suggested two criteria to define being resilient: (1) an individual should be exposed to significant risk or adversity, and (2) an individual should at least attain typical development outcomes. A large body of research suggests that socio-economically disadvantaged students are more likely to have parents who work in lower-paid or unskilled professions and often lack home educational, cultural, and material resources. Such students also are more likely to attend schools with fewer educational resources, and a large proportion of the student population speaks another language at home than at school (e.g.,Akiba et al., 2007; OECD, 2013). These adversities are consequences of a disadvantaged family background and make it more difficult to succeed in school.

However, academically resilient students exhibit a strong capacity to overcome the challenges their disadvantaged socio-economic background brings to them and indeed achieve to a decent level of subject knowledge. Furthermore, resilience is typically measured in terms of protective and risk factors that can predict the likelihood of obtaining resilient outcomes. Risk factors are individual or environmental attributes associated with adverse outcomes (e.g., poverty or truancy). In contrast, protective factors provide buffers for individuals to overcome adverse effects and can result in higher than expected educational outcomes (Naglieri et al., 2013). It has been observed that resilient students often display higher self-confidence, perseverance, willingness and capacity to plan, lower anxiety (Martin & Marsh, 2006, 2009; OECD, 2011; Wang et al., 1994), and greater engagement in class and in academic activities (Borman & Overman, 2004; Finn & Rock, 1997). These characteristics — the inferred capacity of positive adaptation despite adversities (Masten, 2014) — enable children to bounce back from adversities and often are used in interventions that seek to foster academic resilience (e.g., Cassidy, 2016; Connor & Davidson, 2003; Lamond et al., 2009; Martin & Marsh, 2006).

The complexity in determining the affective factors of academic resilience is related to the fact that they occur at different levels. Individuals vary in their characteristics and the circumstances in the home, school, neighborhoods, and countries. The interplay of these factors influences individuals’ possibilities to be academically resilient. This means that academic resilience is an intricate construct with multiple sources of impacts. In the theoretical model of academic resilience (Fig. 1), Bronfenbrenner (1999) emphasized that the risk and protective factors operate in integrated levels and contexts across different dynamic mechanisms. Consequently, the development of academic resilience is contingent on environmental circumstances that simultaneously can reduce risk factors and promote protective factors. A factor of great importance is the school environment.

Fig. 1
figure 1

A theoretical model of academic resilience (Bronfenbrenner, 1999; Sandoval-Hernandez & Cortes, 2012)

Full size image

In addition to student-level factors, classroom and school-level factors reported by teachers or school principals (e.g., teacher’s confidence and expectation in student’s academic success; positive student–teacher relations; school emphasis on school success; parental support) may also enhance the likelihood of being academic resilient and, in turn, can predict achievement (e.g.,Borman & Overman, 2004; Cohen, 2013; OECD, 2011). Martin and Marsh (2006) pointed out that, in future studies of academic resilience, multiple sources of data should be simultaneously taken into account in predicting academic resilience. Relevant factors that promote resilience in the different systems that shape positive adaptation, suggested by Bronfenbrenner’s resilience model, were selected in this study to model academic resilience (e.g., Bronfenbrenner, 1999; see also, Masten, 2018). In the current study, academic resilience is viewed as a process of interaction among factors across levels. Different capacities, abilities, and incentives of individuals may lead to diverse actions and coping strategies when facing challenges. The purpose of the current study is to identify a group of academically resilient students through exploring the unobserved heterogeneity among individuals’ protective factors and their reading achievement.

Measuring academic resilience in large-scale assessments

The concept of academic resilience deepens the understanding of disadvantaged students’ performances in various school settings. However, existing approaches to conceptualizing academic resilience may work differently when employed empirically (Ye et al., 2021).

Studies on academic resilience demonstrate great variability in how the construct is measured. In turn, this results in a variability in the inferences drawn about resilience. In a systematic review of studies on academic resilience with quantitative data, Rudd et al. (2021) identified three ways of measuring academic resilience: (a) the definition-driven approach, which identifies resilient individuals based on pre-determined criteria of risk and resilience; (b) a process-driven approach, which investigates the interaction of both protective and risk factors on different achievement outcomes; and (c) a latent construct approach, which yields a resilience score based on the individuals’ capacity for resilience. Typically, studies that follow the definition-driven approach operationalize academic resilience using two key components: students’ social background and educational outcomes. Such approach has been used when examining academic resilience within OECD’s PISA as well as IEA’s TIMSS and PIRLS which have enhanced the narratives in this research area (e.g., Agasisti, et al., 2021; Cordero & Mateos-Romero, 2021). In the OECD studies, for example, identification of the academically resilient children has relied on predetermined cutoff values in socio-economic and cultural status (ESCS) and student achievement in the core subject (e.g., OECD, 2011; OECD, 2019). The typical criterion for being academically resilient is to be in the bottom quarter (≤ 25%) of the ESCS scale in their own country, but to score in the top quarter (≥ 75%) in that country. On average, the share of resilient students is 11.3%, ranging between 10 and 15% among the participating countries (OECD, 2019). However, the applied cutoffs imply a somewhat narrow view of what it means to be academically resilient. It also differs from the conceptual definition stating that resilient students should have at least normal development regarding achievement (Goldstein & Brooks, 2013) since resilient students, as operationalized in PISA, need to score in the top quarter. The approach overlooks other adversities and places students in the same pool purely depending on the SES disadvantage. Moreover, in this approach, the likelihood of misclassification is substantia. For example, neither a very disadvantaged (< 10th percentile) student who performs in the 50–70th achievement percentile, nor a student with family SES just above the 25 lowest per cent and performs among the top 10% would be characterized as resilient. Therefore, a more nuanced understanding of what it means to be resilient is required. Ye et al. (2021) have clearly illustrated the problems associated with predetermined criteria when defining academic resilience. They demonstrated how the composition of academically resilient students is affected by the cutoff values of SES and the indicators used to measure SES, as these depend on the country’s socio-economic circumstances. Most importantly, the distribution of gender and ethnicity varied dramatically depending on the criteria and measures used. Rudd et al. (2022) also showed differing prevalence rates, ranging from 5 to 40%, could depend on the nature of the operationalization used by the researcher to identify resilient students.

Another reason for improving the measurement of academic resilience is that the part positive adaptation of the resilience concept is composed of factors other than achievement scores. Being resilient is not a singular trait, but rather a chained action where many intertwined processes and contexts operate. The capacity of a resilient child depends on what “resources and supports available to the child through many processes, both within the child and between the child and the many systems the child interacts with” (Masten, 2018, p. 16; see also, Bronfenbrenner, 1999). Hence, rather than examining the relationship between protective and risk factors — such as levels of perseverance, confidence, and school climate and students’ belongingness to the 25–75% group (i.e., academically resilient group) — these factors should be embedded in the identification of the resilient group (Rudd et al., 2021; Tudor & Spray, 2018). Therefore, we suggest that resilient student groups should be modeled based on a set of indicators derived from theory, and not only achievement and socio-economic status. We hypothesize that several personal and environmental factors act as protections to compensate an individual’s disadvantage (see Fig. 1). These factors encompass both individual aspects, such as a more positive self-concept and motivation, as well as relational and environmental attributes, such as a sense of belonging and a sense of safety and positive relationships with teachers and friends (e.g., Masten, 2014, 2018). One challenge lies in collecting sufficient data to explicitly test such hypothesis. The international large-scale assessments (ILSA) collected data to monitor trends in student achievement in mathematics, science, and reading, and to facilitate examinations of the determinants of such trends across education systems around the world (e.g., Nilsen et al., 2022). Given the general purposes of the ILSA’s, existing data may not be specific and adequate for studying academic resilience. However, the measures of achievement and social background have been used to define academic resilience (e.g., Agasisti et al., 2018), and several factors are also useful for operationalizing academic resilience, not least since they are well in line with the Bronfenbrenner’s theoretical model. For example, the experience of belongingness in school is fundamental to successful learning. Thus, the school’s ability to create a sense of belonging is a vital aspect in promoting resilience (Cohen, 2013; Taub & Pearrow, 2013). On the other hand, the experience of bullying is detrimental for learning (e.g., Olweus, 2010). Equally, supportive teachers are beneficial for both student learning and the promotion of resilience (e.g., OECD, 2011). Confidence is closely related to students’ self-concept in school subjects and often has a positive relation with achievement. Having a positive self-concept is also related to academic resilience (e.g., Goldstein & Brooks, 2013; Masten, 2018). Interest and passion are essential aspects of grit, which has been shown to correspond well to the resilience construct (Thorsen et al., 2021).

The Swedish case

For the purposes of the current study, Sweden provides an interesting example for the investigation of resilience. The Swedish school system has implemented several reform changes in the last three decades. As a consequence of these changes, the Swedish school system has gone from being one of the most egalitarian school systems with the smallest variation in school achievement and the SES gap, to a system with a severe decline in educational equity and an intensified school segregation (e.g., Gustafsson and Yang Hansen, 2018). The new landscape of the Swedish school system is characterized by decentralization, deregulation, privatization, and marketization (Blossing & Söderström, 2014; Lundahl, 2016). The strong voucher system, together with a free-school choice, has made the school-market appealing to private actors (school companies) that make huge profits by attracting high-performing students from more advantageous backgrounds, leading to increased school segregation. Evidence now indicates that these system changes have weakened the ability of schools to fulfill their compensatory mission, especially in relation to socioeconomically vulnerable students. The challenges in providing equal access of teaching resources and learning opportunity have further increased school segregation (Hansson & Gustafsson, 2016). However, in the growing body of disadvantaged students, there is a group of students who manage to beat the odds with their personal and contextual assets. The proportion of resilient students in Sweden does not stand out in international comparison according to previous reports (e.g., OECD, 2019), but a fair share of students is characterized as academically resilient. Therefore, it is critical to identify the assets and mechanisms that resilient children rely upon when facing the adversities in academic contexts, since knowledge of this sort has potential to reverse the trend of increasing school segregation. It may not be cognitive ability that is most important in the identification of resilient students, but the emotional intelligence, self-awareness, self-regulation, motivation, empathy, and social skills also matter. The present study used the indicators of some of these aspects to enhance the identification of resilience students.

Aim

The concept of resilience comprises two key components — disadvantage and positive adaptation. In the present study, disadvantage is a necessary condition for modeling resilience. Thus, the first step in the analysis was to define disadvantaged students based on a social background index. Rather than using fixed or absolute thresholds across countries, we adopted different thresholds that were relative for each of the countries tested. In an absolute sense, it is not reasonable to argue that 33% of the students, in, for example, Sweden or Norway, are classified as disadvantaged. However, we consider being disadvantaged relative to context and circumstance — and we provide three different approaches to categorize disadvantage.

The current study builds on a latent construct approach (i.e., latent profile analysis, LPA) that includes protective and risk factors to infer academic resilience. This allows for a more comprehensive understanding of the construct, as well as the ability to identify individuals who may be more resilient than others. LPA is a person-oriented approach where the emphasis is on the profile of an individual’s characteristics captured by the variables in the analysis. LPA postulates that there is unobserved heterogeneity among individuals’ characteristics, and that an underlying latent categorical variable can be identified to classify individuals into discrete, mutually exclusive latent profiles (e.g., resilient vs non-resilient, Muthén, 2008). Within each profile, individuals are homogenous in their characteristics described by the latent profile indicators. The latent profile membership of individuals is thus initially unknown, but can be inferred from individuals’ response patterns to a set of observed indicators, such as students’ attitude towards reading, reading achievement, and sense of belonging. Two research questions were specified:

  • Is it possible to characterize a group as resilient among disadvantaged students, using different cutoffs (25%, 33%, between 25 and 33%) for disadvantage?

  • How do the levels of the protective factors differ across the disadvantaged SES groups? What characterizes the resilient student group?

Data and method

Data come from the large-scale international assessment Progress in International Reading Literacy Study (PIRLS) 2016, carried out in around 50 countries worldwide by the International Association for the Evaluation of Educational Achievement (IEA) (Mullis et al., 2017). The current study explicitly focused on the Swedish school system, using data from all countries in a sensitivity analysis. Sample sizes vary between around 4000 and over 10,000 grade 4 students across the countries in this study (see Appendix 2). In addition to a reading test, sampled individuals also completed an extensive Student Questionnaire (StQ), by which rich information about family background and personal characteristics relating to different aspects of learning and schooling was collected. Moreover, a home questionnaire (HQ) was administered to students’ parents (Foy, 2018).

Variables

The basic idea of the study was to identify homogenous groups of students based on antecedents of academic resilience using a modeling approach. In line with previous research and theory, several variables were selected to model student groups with different degrees of academic resilience. While disadvantage was defined by the fewest home resources, school belongingness and confidence in reading, and subject-specific confidence were also found to be stable predictors of academic resilience (see e.g., Rudd et al., 2022).

Disadvantage — student educational background

Students’ social background is one of the key components of academic resilience. IEA constructed the Home Resources for Learning (HRL) scale based on students’ and parents’ responses concerning the availability of (1) number of books at home, (2) number of study supports, (3) number of children’s books, (4) highest education of either parent, and (5) highest level of education of either parent. The HRL variable was computed using an IRT-model and placed on a continuous scale in the same fashion as the other variables we are using in the study. We used this measure to operationalize student adversity by categorizing the 25% and 33% students with the fewest home resources as well as the 8% students in between, as disadvantaged.

In order to more fully understand the nature of resilience, three different categorizations of disadvantage were made, based on the variable Students Home Educational Resources (HomeRes). In line with previous research on academic resilience, three sub-samples were thus created to shed light on potential differences in the number of resilient students.

  1. 1.

    The lowest 33% cutoff is used to select the 33% students with the fewest home educational resources as our first sample.

  2. 2.

    The lowest 25% cutoff is used to select the 25% most disadvantaged with respect to home educational resources.

  3. 3.

    Students between 25 and 33% are selected as an at-risk group with higher resources levels than 25% of the students but only up to the 33rd percentile. These students could potentially differ from the rest of the first two student groups, but have not been focused on in previous research.

Indicators of academic resilience

Reading achievement (Readach)

Students’ overall achievement in reading is represented by five variables in PIRLS. These are so called plausible IRT-values (Item Response Theory) by which all students’ results are placed on a common scale although they have not taken all test-items. In IRT scaling, every student’s score is calculated through taking student ability and test-item difficulty into account (Embretson & Reise, 2000; Rutkowski et al., 2010; Wu, 2005). The international mean value in PIRLS is set to 500 points with a standard deviation of 100. Typically, all five plausible values should be used for separate runs (e.g., Jerrim et al., 2017; Laukaiyte & Wiberg, 2017; Osman et al., 2019), and a single estimate shall be obtained by averaging the resulting statistic. We ran analyses for each plausible value and present averages for these runs in the results section (Table 3).

Students’ sense of school belonging (Belong)

The Students’ Sense of School Belonging (SSB) scale was created based on students’ degree of agreement with the seven statements concerning whether they liked being in school, feeling safe in school, that teachers are fair and so forth. Students perceived their belongingness on a 4-point scale ranging “Agree a lot” to “Disagree a lot.”

Student bullying (Bully)

The bullying scale was created based on students’ responses to how often they experienced 8 different bullying behaviors. Students should consider whether they experienced that other students made fun of, threatened, spread lies, left them out of activities etc. during the school year. Items reflected in the bullying index also reflect school safety. Responses ranged from “Never” to “At least once a week.”

Confidence in reading (ConfRead)

Items were to be considered on a 4-point scale ranging from “Agree a lot” to “Disagree a lot” and examples of statements were “I usually do well in reading,” “Reading is easy for me,” and “I am just not good at reading.”

Reading engagement (Engage)

This index was computed by means of nine questions where students were to perceive their teacher and the instruction. The construct captures the instructional quality of teachers and is likely to be subjectively perceived. Nine items were used, three examples of these are “I know what my teacher expects from me,” “My teacher is easy to understand,” “I am interested in what my teacher says.” Items were to be considered on a 4-point scale ranging from “Agree a lot” to “Disagree a lot.”

Reading interest (LikeRead)

We included a scale capturing students’ liking of reading. Students perceived their liking of reading on a 4-point scale ranging Agree a lot to Disagree a lot. Ten items were used to construct the scale. Item examples are “I enjoy reading,” “I would be happy if someone gave me a book as a present,” and “I read for fun.”

In Table 1, descriptive statistics for the samples we used are presented for the variables that will be used to model resilience.

Table 1 Descriptive statistics of the indicators of academic resilience for three disadvantaged groups in Sweden
Full size table

As can be observed, the groups categorized as being disadvantaged have lower average reading achievement than the national mean. However, the sample between 25 and 33% with students at risk has an average achievement close to the national mean. As regards the other factors, mean values are similar for the disadvantaged groups in relation to national means.

Method of analysis

A latent profile is determined by patterns of conditional probabilities giving the profile membership. The conditional probabilities indicate the likelihood for a student within a given latent profile giving a particular response on an observed measure. If the groups are distinctly different from each other, and the probability approaches 1, it is very likely that a student belongs to a certain group. By examining the pattern of responses for all items, it is possible to obtain an overview of the nature of each latent profile. In the present analysis, the LPA provides a means to classify disadvantaged students in different groups according to their attitude towards reading, reading achievement, and sense of belonging.

To determine the number of profiles, Bayesian information criteria (BIC) and entropy values were compared across models with different profile solutions (Schwartz, 1978). The model solution with the lowest BIC was usually preferred. The Vuong-Lo-Mendell-Rubin likelihood ratio test (LRT, Lo et al., 2001) and parametric bootstrapped likelihood ratio test compare the log likelihood differences of the model with k classes with the one with k-1 classes. A significant p-value for the k-1 (H0) versus k classes indicates that the solution improved with k classes, while a non-significant value suggests that the k-1-class solution should be retained. Based on findings of a simulation study by Nylund et al. (2007), a significant improvement in log likelihood difference implies that the k-class model fits the data better than the model with one less class. As noted however, the BIC-value frequently performs better than the LRT-test in terms of deciding on the number of classes (Nylund et al., 2007). A sequence of LPA models was specified, starting from 1-profile model, and successively adding the number of latent profiles based on the variables presented in Table 1. We typically ran analyses with 1–5 classes. Generally, a 2-, 3-, or 4-class solution was preferred. We plotted the BIC statistics in order to graphically identify the most appropriate solution in terms of fit. BIC statistics for Sweden are presented in Appendix 1.

It should be noted that the substantive meaningfulness of the latent profiles is an important factor to consider when determining the number of latent classes. If the latent profile size is small, a substantive rationale based on previous research and theories must be offered to support the inclusion of the class. However, the extremely small group may be a statistical artifact that simply reflects the measurement character in the sample (Samuelsen & Dayton, 2010). Analyses were carried out using IBM SPSS (statistical package for the social sciences) and Mplus 8.4 (Muthén & Muthén 1998-2017).

Results

In the following sections, the results for Sweden will be presented in more detail. We selected a set of countries for further scrutiny for our investigation of the share of resilient students. The results for all countries are presented in Appendix 2.

Modeling resilience

The results suggested that a four-profile solution was preferred for the Swedish samples. We noted that BIC kept decreasing with the introduction of more classes. However, the relative improvement of fit decreases rather quickly and seems to become rather small after four classes. While the BIC favored the solution with 4 classes over 3 classes, the LRT test suggested that the 3-class solution did not improve fit as compared with the 2-class models for the 25% and 8% groups. However, as BIC kept improving, the entropy value increased, and the class classification probabilities for the four classes were well over 0.90 (the “resilient” group being close to 1; see Table 2), we were confident in selecting the solution with 4 classes. The entropy value was well over the recommended value of 0.8, signifying an acceptable level of classification precision.

Table 2 Average latent class probabilities for most likely latent class membership (row) by latent class (column)
Full size table

The probabilities presented in Table 2 indicated distinctive group belongingness of the individuals categorized in the 4-class solution. The latent profile model results for Sweden are presented in Table 3. The first columns presented the mean values of each latent profile indicator in the four classes, along with SE for the means.

Table 3 Latent profiles for three disadvantaged groups in Sweden
Full size table

The number of students characterized as resilient is 343 in the 33% sample. As may be noted, these are the students forming the resilient groups in the 25% and 8% samples. As indicated by the mean values, the resilient groups differ significantly in their characteristics compared with the other groups. We tested the differences between the groups using a one-way ANOVA with post hoc tests. We applied the Bonferroni corrections to minimize the type 1 error rates, and we applied the Games-Howell test for variables that had unequal variances across the groups (Field, 2017). In the 33% and 25% groups, the “resilient” group outperforms the non-resilient group by 0.2–0.3 SD in reading achievement, which must be considered substantial given the accumulative nature of educational disadvantage. It should be noted that the resilient group still performs some points below the national mean (554), but is nevertheless higher than might be expected for students at risk. Interestingly, the 8% at-risk children perform somewhat above the national mean. Notably, the two groups perform in a relatively similar way, and the resilient group actually performs somewhat lower than group 3 (Table 4).

Table 4 Differences between resilient group (4) and the three other low-SES groups
Full size table

Though we found substantial between-group differences in reading achievement, the most striking differences relate to the indicators of academic resilience. The resilient class has substantially higher means for all groups (33%, 25%, and 8%). Substantial differences are found for school belonging, where the effect size amounts to over 2 SD in the 33% sample.

Sensitivity analysis

To validate our results, we replicated the analyses with a set of countries participating in the PIRLS 2016. We selected Finland as it has a similar culture as in Sweden, but another development in terms of school segregation, which in Finland is very low. Sweden has a more extensive voucher system than Finland, and the other Nordic countries. We also selected a few countries based on their varying performances in PIRLS, their different geographical location, and differences in language and culture, as well as their different proportions of resilient students (OECD, 2019). The included countries are Canada, Chile, Finland, and Singapore. Canada has a decentralized school-system. However there is a weaker association between students’ social background and achievement than in Sweden. Canada, in comparison, has a higher share of resilient students (OECD, 2019). Singapore, on the other hand, has a highly centralized school system, and has a lower share of resilient students. Chile’s school system is similar to Sweden in many ways, despite the countries’ different geographical location. Chile too has a voucher system which compares to that of Sweden. However, the performance in Chile is about half a standard deviation lower than in Sweden. Naturally, other countries would be relevant for a sensitivity analysis. However, the share of missing in the selection variable HRL was quite substantial in many countries, and since some countries did not participate in the home questionnaire (e.g., USA), the selection of countries was also restricted to those where missingness was less than 20%.

The model results for the selected countries are found in Appendix 1. The number of students categorized as disadvantaged varied somewhat across countries since the overall sample size varied quite substantially across countries. Although classification probabilities for the different profiles were slightly lower than for Sweden, the results indicate that a resilient group could be identified in all countries. There were, however, some differences across countries that are worth noting. First, the reading achievement levels for disadvantaged groups were closer to the overall national mean in some countries. In Singapore, for example, the average reading achievement for the total sample was 572, thus significantly higher than in Sweden (554). However, for the 25% disadvantaged group, the average score was the same in Singapore as that in Sweden (521). Interesting differences were also revealed when a resilient group was in focus. As in Singapore, the resilient class in Finland had a higher average on all indicators of academic resilience. Furthermore, reading achievement differed less than in the other countries. In Sweden, quite substantial differences were observed between the achievement of the resilient group and the other groups. In Finland, the differences were not more than a couple of points.

In most countries, students between 25 and 33% performed relatively well, thus mirroring results found in Sweden. In Chile, however, both the lowest 33% and 25% resilient groups outperformed the total sample. There were no clear-cut resilient groups in the between 25 and 33% sample. Relatively more competent readers are thus located in the most disadvantaged quartile in Chile, which is in stark contrast to the other countries tested.

Discussion and conclusions

The present study used latent profile analysis to identify homogeneous clusters of socio-economically disadvantaged students. We used several different definitions of disadvantage and applied our model to several countries participating in the PIRLS 2016. Based on a set of predictors of academic resilience and reading achievement, we identified a group of students that could be regarded as academically resilient in most countries studied.

The substantial differences regarding the predictors of academic resilience for students with different degrees of disadvantage in their family background indicate that the modeling approach to study academic resilience is a promising alternative to the definition-driven approach, since academically resilient students are not only low in SES and high in achievement. Rather, being resilient is a result of an interaction process between risk factors and protective factors (e.g., Bronfenbrenner, 1999). Given that such a process occurs both within the individual and in different contexts around the individuals, i.e., teachers, classrooms, and schools (e.g., Masten., 2018), determining resilience requires the simultaneous modeling of both risk and protective factors in these contexts and their interaction with individuals. Oftentimes, the academic resilient group was clear-cut in our modeling, while reading achievement was at about the same level for the two groups in some countries. Typically, one of the groups with higher reading achievement had clearly higher values on the predictors used in the analysis, even though their reading achievement was not much different from another group (e.g., Finland). In contrast, a third group had lower achievement and lower values on the predictors. These results clearly illustrate that a disadvantaged group is not homogenous. Achievement can be a protective factor, and several studies have shown a high correlation between achievement and future well-being (e.g., Crede et al., 2015). However, individuals who experience a sense of belonging, high confidence, strong engagement, and interest in reading are likely to be more persistent in their learning, particularly younger students (Thorsen et al., 2021).

While we have presented a person-centered modeling approach based on selected indicators to identify academically resilient students, cutoff values were still involved in classifying individuals as disadvantaged. It should be noted that indicators such as parental educational level and SES composites are rather crude measures of underprivileged students and may mis-specify resilience. However, it was necessary to define groups with disadvantages to model resilient groups — inclusion of SES measures in the modeling rendered typically groups based on social strata rather than clear-cut resilient groups.

In the present study, we tested a person-centered modeling approach to study academic resilience and demonstrated the appropriateness of such an approach that is guided by theory. However, it is reasonable to believe that other studies may propose slightly different and/or more protective factors in the operationalization of resilient students. Carrying out our study on the large-scale international data, we were restricted to factors that were available in PIRLS. The factors were measured using self-reported data, and items were similarly phrased regardless of the national context. However, while we found the selected factors suitable to the Swedish setting, in future research, in-depth studies should be carried out, especially for academic resilience, and designs should include relevant factors that are geared towards the specific context. We also considered the national context, rather than adopting a fixed criterion for all countries. We did not use cut-off values (above the 75th percentile performance, below 25th percentile SES) as previous research has suggested a large misclassification for such approach (Ye et al., 2021). Applying a strict cutoff in Sweden suggests that only 90 students, corresponding to 2.4%, performed at or above the 75th percentile while also characterized as socio-economically disadvantaged (below 25th percentile SES). Notably, the group that we classified as academically resilient is around 10% in most of the countries tested. This indicates that the number of resilient students could be underestimated in previous research.

We argue that the modeling approach resonates well with the theoretical definition of academic resilience in capturing the interplay between risk and protective factors on different integrated levels and in varying contexts. In research on the measurement of resilience, Rudd et al. (2021) conclude that in identifying academic resilience, it is necessary to capture this complex interplay between risk and protective factors. They argue that the process-driven approach is the most appropriate in capturing this interplay. However, unlike the process-driven approach, the modeling approach also includes adversity as a factor. Additionally, the modeling approach integrates achievement in understanding the dynamic relationship between risk and protective factors in the identification of academic resilience. Merely capturing the absence of academic failure, as in definition driven approaches, does not capture the positive adaptation integral to academic resilience and provides a narrow view of education (Tudor & Spray, 2018). Being engaged and experiencing self-confidence are also important aspects of positive adaptation in education.