The many faces of adult learners

Annex A. Methodological description

Latent class analysis

The segmentation approach used in this paper employs a Latent Class Analysis (LCA) methodology to identify sub-groups within the adult population that share similar characteristics. Similar to clustering methods, LCA identifies mutually exclusive and exhaustive latent (or unobserved) classes based on patterns in observed data. LCA estimates class membership probabilities and uses iterative numerical methods to determine the model that best fits the data, based on a statistical criterion.

LCA is widely used applied across disciplines to classify behaviour patterns, attitudes, and preferences. It has been used, for example, to identify consumer preferences (Dolšak, Hrovatin and Zorić, 2020[16]) and to analyse subpopulations based on responses to survey or test items (Bertrand and Hafner, 2013[17]). In the field of education, LCA has been employed to explore educational choices and patterns of learning engagement, particularly within formal education settings. For instance, Denson and Ing (2014[18]) used LCA to classify freshmen according to their ability to work effectively with others of diverse backgrounds, being open to new ideas and different perspectives and being empathetic with other perspectives. This helped higher education institutions target diversity-related interventions. Similarly, Ramesh et al (2014[19]) applied LCA to model and understand student engagement in online courses, using behavioural indicators to predict course completion and identify key survival factors associated with learner persistence.

However, the use of LCA in learning engagement outside formal education settings remains relatively limited. The OECD has applied LCA in multiple policy areas, including identifying different labour market groups based on employment barriers in the “Faces of Joblessness” reports (Fernandez et al., 2016[14]), and defining adult learning profiles in the “OECD Skills Strategy Implementation Guidance for Flanders, Belgium: The Faces of Learners in Flanders” report (OECD, 2022[3]). The methodology developed for Flanders has informed the approach used in this policy paper. The same model has been applied, with two main differences: this paper uses AES 2022 data instead of AES 2016 (see below), and the report for Flanders included additional characteristics for the profiles (e.g. socio-demographic characteristics, learning characteristics, and more).

LCA offers three key advantages over other common segmentation (or clustering) methods: 1) it relies on formal statistical tests guide the selection of the optimal number of profiles and other model parameters; 2) rather than assigning individuals to groups deterministically, LCA estimates the probability of membership in each profile, thereby reducing the risk of classification errors in subsequent analysis; and 3) it is well suited to handling common data challenges, such as missing values and complex survey designs (Collins and Lanza, 2009[20]).

Data

While various surveys collect data on adult education and training participation, motivations to learn and obstacles to participation – including the OECD Survey of Adult Skills, a product of the Programme for the International Assessment of Adult Competencies (PIAAC) – this paper utilises the latest 2022 data from the EU Adult Education Survey (AES) (Eurostat, 2024[8]). The AES was chosen because it is specifically designed to provide detailed insights into adult learning and includes a comprehensive set of indicators on learning activities and learner characteristics.

The AES covers participation in formal, non-formal and informal learning among adults aged 25-64 and is one of the primary data sources for EU lifelong learning statistics. The survey’s reference period for participation is 12 months before the interview.

While PIAAC may not include learning indicators as detailed as those in AES, it offers additional measures – such as skill levels – that could be valuable to explore, and it covers several non-EU countries. For this reason, the paper suggests that conducting a similar analysis using PIAAC data could be a relevant next step.

Methodology

Learner profiles are identified using a LCA methodology, which involves defining a baseline model, selecting relevant indicators and identifying the optimal number of classes (learner profiles), as described in more detail below.

Defining the baseline model and selecting the indicators

LCA estimates were generated for two baseline models representing two distinct groups of adults: i) adults not participating in non-formal or formal learning activities; ii) adults participating in non-formal or formal learning activities.

Unique models were developed for these two groups of learners: one model explored factors that might explain their non-participation in learning, while the other examined factors that might explain their participation (see Table A A.1). The indicators included in the models are generally regarded as the primary factors influencing participation or non-participation and are measured with various surveys on adult education and training, including AES and PIAAC.

Table A A.1. Indicators in baseline models for participants and non-participants

Not participating	Participating
Dispositions towards learning	Dispositions towards learning more
Did not want to participate in education and training	Did (not) want to participate in more education and training Did (not) see a need for participating in additional education and training
Obstacles preventing participation	Reasons for participating
Cost of participation Family responsibilities Work schedule Lack of support from the employer or public services Health and/or age obstacles Personal reasons (e.g. negative experiences, other personal reasons) No suitable training available (e.g. no suitable education activities, inaccessible as located too far away, not the right prerequisites, course full or too few registrations)	To do my job better To improve my career prospects To reduce risk of job loss To increase my possibilities of getting a job, or changing a job/profession To start my own business Due to of organisational and/or technological changes at work	To obtain knowledge/skills for own general interest and curiosity To obtain a certificate To meet new people/for fun For health reasons To do voluntary work better Required or obliged to learn by the employer or by law

For the first group of profiles – non-participants – the model includes indicators of both dispositions towards learning and obstacles to participation (see the first column in Table A A.1). In the AES, adults who did not participate in learning activities were asked whether they would have liked to participate, revealing their dispositions towards learning. Those who responded negatively were further asked whether they would this was due to a perceived lack of need. Regardless of their interest in participation, all non-participants were also asked about specific obstacles they faced, providing additional insight into the barriers to participation.

For the second group of profiles – participants – the model draws primarily on indicators related to individual’s reasons for participating in learning, encompassing both job-related and not job-related factors (see the second column in Table A A.1). These indicators offer insights into learners’ attitudes towards learning and may reflect different motivational orientations (e.g. extrinsic and intrinsic motivations to learn). In addition, to gain a more comprehensive understanding of these motivational groups, the model also includes indicators from AES about their interest in engaging in additional learning activities beyond those they already participate in. This approach provides insights into motivations ex post – after engaging in learning – as opposed to solely focusing on the ex ante reasons for participation.

When conducting the analysis, several indicators from the AES for non-participating adults were merged to improve the performance and robustness of the model. This was necessary due to the high prevalence of the response “not wanting to participate in learning” and the structure of the survey, which limited questions on obstacles to only those adults who expressed a desire to learn. As a result, these questions captured only a subset of the non-participating population. Consequently, several categories of obstacles were rarely selected by respondents, reducing their analytical utility. To address this, related categories were combined to ensure sufficient response levels and improve model fit. This limitation did not arise for indicators related to participants, as all individuals who had participated in learning were systematically asked about their motivations, ensuring a more complete dataset for that group.

The indicators used in the LCA estimations are binary, which simplifies the statistical model and enhances the interpretability of results. Moreover, using binary indicators helps minimise information loss when working with categorical variables that have multiple potential response options, thereby improving the robustness and clarity of the analysis.

Assessing the statistical fit of the model

The optimal number of profiles is not automatically determined by LCA. Instead, the analysis requires an assessment of which number of profiles is the best fit. To this end, LCA calculations are conducted with an increasing number of classes (profiles), ranging from two to seven classes. This process yields the following outputs for each model:

1. The share of the population represented by each class, indicating the relative size of each learner profile.

2. The estimated probabilities of each indicator within each class, which reflect the strength of association between individual indicators and specific profiles. For example, a probability of '1' indicates that the indicator is a defining characteristic of a specific profile, whereas a probability of '0' shows that the indicator is not associated with that profile. These probabilities are used to interpret the distinct combinations of characteristics that define each learner profile.

To determine the optimal number of profiles, this paper uses two widely recognised goodness‑of‑fit indicators: The Bayesian Information Criterion (BIC) (Schwarz, 1978[21]) and the Akaike Information Criterion (AIC) (Akaike, 1987[22]). Both criteria support the selection of a model that balances statistical fit with parsimony, helping to identify the best compromise between model complexity and explanatory power While increasing the number of parameters can improve model fit by reducing error, it may also lead to overfitting and reduced generalisability. By evaluating BIC and AIC values across different class specifications, the analysis identifies the optimal number of learner profiles – separately for participants and non-participants – ensuring that the resulting models are both robust and interpretable.

To illustrate this process, Figure A A.1 shows the percentage changes in BIC and AIC values as the number of classes (profiles) increases from three to seven for participating adults in Ireland. The model corresponding to the lowest BIC or AIC value – indicated by a negative change that is closest to zero – is typically considered the most appropriate, as it reflects the best balance between model fit and simplicity. Based on this criterion, the six-class model is selected for the case of participating adults in Ireland. Using this methodology, the optimal number of profiles may vary across countries and between the non-participating and participating adult populations. Table A A.2 presents the selected number of profiles for each group in Bulgaria, Finland, Ireland and Portugal.

Figure A A.1. Change in AIC and BIC for increasing number of classes for participating adults in Ireland

Note: The X‑axis indicates the number of classes (profiles) estimated in the model, ranging from 3 classes to 7 classes. The Y‑axis represents the percentage difference in BIC (the Bayesian Information Criterion) and AIC (the Akaike Information Criterion) between different classes. For instance, the percentage for 3 classes indicates the difference in AIC and BIC between models with 2 classes and 3 classes.

Source: Adapted from Eurostat (2022[15]), Adult Education Survey 2022, https://ec.europa.eu/eurostat/cache/metadata/en/trng_aes_12m0_esms.htm.

Table A A.2. Optimal number of classes (profiles) in Bulgaria, Finland, Ireland and Portugal

	Not participating	Participating
Bulgaria	2 classes	4 classes
Finland	4 classes	5 classes
Ireland	3 classes	6 classes
Portugal	4 classes	5 classes

Source: Adapted from Eurostat (2022[15]), Adult Education Survey 2022, https://ec.europa.eu/eurostat/cache/metadata/en/trng_aes_12m0_esms.htm.

While changes in BIC and AIC were the primary criteria used to determine the optimal number of profiles, face validity was also an important consideration. This involved assessing whether each profile represented a coherent and distinct set of characteristics, and whether the profiles meaningfully reflected different types of learners – for example, avoiding overlap or profiles lacking clear defining traits. This consideration was particularly important in the case of Bulgaria. Among non-participating adults, low motivation levels were widespread, resulting in limited variation in the data and making it difficult to identify a meaningful profile structure based on statistical criteria alone. For participating adults, the sample size was small, and although the lowest AIC and BIC values were observed for seven profiles, several of these profiles displayed identical or near-identical characteristics, limiting their interpretability. As a result, qualitative judgement was applied to ensure that the final selection reflected distinct and policy-relevant learner profiles. Ultimately, two profiles were identified for non-participating adults and four for participating adults in Bulgaria.

Interpreting LCA findings: Data and methodological limitations

When interpreting the findings of this report, it is important to consider the specific characteristics and limitations of the survey data used – namely, the AES. The interpretation of key concepts such as “motivations to learn”, “obstacles to participation”, and “reasons to learn” is shaped by the way in which survey questions are phrased. For instance, when respondents are asked about their reasons for participating in learning, they choose from a set of predefined options (e.g. “to do my job better”, “to improve my career prospects”, or “to be less likely to lose my job”). While these categories capture common motivations, other relevant factors may not be reflected, potentially constraining the interpretative scope of the data.

In addition, although LCA is a powerful tool for identifying unobserved subgroups within a population, it is subject to several methodological limitations. LCA assigns individuals to classes based on the probability of membership, given their responses to a set of observed indicators (Muthén and Muthén, 2000[23]). As these classifications are probabilistic rather than deterministic, interpreting the results can be complex. One potential challenge is the “naming fallacy”, where the labels assigned to classes may not fully reflect the diversity of experiences and characteristics within each profile (Weller, Bowen and Faubert, 2020[24]).

Moreover, LCA relies on the Local Independence Assumption (LIA), which posits that all observed indicators are conditionally independent given the latent variable. While this assumption facilitates estimation and supports causal interpretation, it may oversimplify relationships among the variables. In practice, some indicators may be interrelated in ways that are not fully accounted for by the latent classes, particularly when multiple unobserved factors influence learning behaviour. As a result, some complexity may be lost, and the profiles generated may not capture the full range of interdependencies present in the data (Visser and Depaoli, 2022[25]; Subtil, de Oliveira and Gonçalves, 2012[26]).