Determinants of Adult Education and Training Participation in the United States: A Machine Learning Approach

Research Background

Learning is not confined to the traditional K-12 school system; instead, it permeates all aspects of life. Individuals continually acquire knowledge and skills throughout their entire lifetime, drawing from diverse experiences (Billett, 2010; Ross-Gordon et al., 2017). Learning occurs not only in formal educational settings but also in nonformal environments within organizations, communities, and society at large. This concept of lifelong learning (LLL) has evolved to strengthen human capital across various career stages (Aspin & Chapman, 2007).

As societal challenges and economic complexities, such as job displacement and mobility, continue to grow, the knowledge and skills obtained solely through public school systems may no longer suffice to meet the demands of the adult workforce (Dibra et al., 2014). Consequently, more and more adults are actively seeking additional educational opportunities to foster their professional development and enhance their employability in today's fiercely competitive labor market (Melacarne & Nicolaides, 2019). Particularly for working adults, job-related adult education and training (AET) represent a significant pathway through which individuals acquire new knowledge and upgrade their skills to meet the ever-changing requirements of the world of work.

Over the past few decades, the concept of adult education has undergone extensive review, particularly in terms of its economic and policy implications. From the perspective of LLL, it is widely recognized that work and continuous learning are interconnected (Tikkanen & Nissinen, 2018). Consequently, there has been a growing academic interest in AET participation and its benefits for individuals and organizations. AET, which is often used interchangeably with LLL, encompasses various learning forms that contribute to knowledge and skill accumulation throughout an adult's life (Desjardins, 2015; Rubenson, 2011).

As work-related learning takes center stage in the LLL agenda (Ross-Gordon et al., 2017), timely provisions of occupational AET have garnered increased policy attention (Yamashita et al., 2019). It comes as no surprise that substantial literature provides ample evidence of AET's critical role in improving performance, driving change, and serving as an overarching goal of human development (Watkins & Marsick, 2014).

Empirical research shows that AET participation is positively associated with employment stability and career advancement at an individual level (CEDEFOP, 2016; Rubenson & Desjardins, 2009). Maintaining essential skills and knowledge is a key indicator of employability from a human capital perspective (Yamashita et al., 2019). Moreover, offering AET opportunities has become a paramount concern for employers and governments in their quest to secure a skilled labor force. Investments in AET and the human capital concept reflect a growing awareness of the necessity to foster a learning-oriented society and an effective LLL system to enhance institutional effectiveness (CEDEFOP, 2016; Melacarne & Nicolaides, 2019)

Research Gap

Despite the significant attention given in scholarly works to the potential importance of job-related AET and its associated predictors and outcomes, there has been limited comprehensive examination of the dynamics of individual and situational factors that influence the participation of working adults in AET. Empirical research on the contextual factors most relevant to AET participation among working adults remains inconclusive. While researchers have made efforts to understand the decisive factors in adult education participation (Kalenda et al., 2020), reaching a consensus has proven challenging. The reliance on traditional parametric regression models, which often fail to address the complex multivariate structure of data (Breiman, 2001), limits our understanding of how individual and situational characteristics jointly influence the dependent variable. These limitations highlight the necessity for an integrated approach that considers multiple factors holistically to gain a comprehensive understanding of working adults’ AET participation.

To address this concern, we employ a random forest classifiers (RFCs) technique as the analytical method, one of the machine learning (ML) algorithms. The advantage of using RFCs over traditional regression models lies in its ability to comprehensively examine multiple factors simultaneously. Furthermore, our analytic choice allows us to compare the relative importance of the selected factors. RFCs are a nonparametric modeling strategy that generates variable importance (VI) scores across multiple independent variables. In this context, we seek to reexamine the determinants of job-related AET participation among working adults by applying the proposed RFCs approach. The research questions addressed in this study include:

Research Question 1: What were the main determinants and their association with job-related formal AET participation among working adults?

Concept and Typology of AET

AET generally refers to “a form of learning that takes place in addition or as a complement to formal education and is distinct from informal learning, which is intentional but less structured” (Widany et al., 2019, p. 8). It encompasses learning activities from personal, social, and work-related perspectives, and this is a crucial component of the concept of LLL (Yamashita et al., 2019). In many countries, AET plays a critical role in improving the alignment between educational qualifications and skills that enhance key competencies, basic skills, and fostering labor market participation (Findsen & Formosa, 2016). Numerous policies and initiatives have been introduced to provide LLL opportunities that align with labor market demands. AET has a positive impact on individuals’ labor market outcomes, and evidence of the quality of AET programs helps organizations make informed decisions regarding investments in adult learning (OECD, 2019). AET typically involves formal and nonformal modes of participation (OECD, 2013).

Formal AET refers to learning that is formally designed and organized, primarily takes place in educational institutions such as colleges or universities (CEC, 2000). The goal of formal AET often centers on obtaining certified educational outputs, such as college degrees and certificates to address workforce education gaps (Yamashita et al., 2019). Therefore, a critical aspect of formal AET involves organizations investing in human resources to improve their employees’ access to training funds and formal education through tuition reimbursement and financial support.

Nonformal learning, on the other hand, serves as an additional and alternative type of learning that complements formal training and is an integral part of the LLL process (Widany et al., 2019). Nonformal AET differs from formal training in its tacit nature of knowledge accumulation, and it generally does not lead to formal credentials (Punksungka et al., 2021). Nonformal AET takes place mostly in institutionalized settings, such as the workplace. Nonformal AET includes short courses, workshops, or seminars that contribute to individuals’ continuous learning and development in their professional journey (Eurostat, 2016; Jarvis, 2010).

Key Drivers of AET Participation

Theoretical Framework

The theoretical model of adult education participation proposed by Boeren et al. (2010) offers an integrated framework for our study. According to their model, the decision to participate in adult education is influenced by various factors at three layers: individual, institution, and socioeconomic contexts. This model emphasizes that individual and organizational factors are embedded within the broader societal context. The interaction between individual demands and organizational offerings plays a central role in shaping adults’ learning participation decisions.

Consequently, AET participation can be viewed as a result of the interplay between individual and work-related contextual factors that are interwoven and complementary. The complexity of factors at both the individual level (e.g., gender, age, education/skills, income) and the institutional level (e.g., job conditions, flexibility, access to learning) can significantly influence the likelihood of adult education participation. The primary objective of this model is to adopt a holistic approach to examine LLL, broadening the scope of inquiry beyond traditional boundaries (Boeren et al., 2010). In our research, we focus on the individual and work-related layers of the model and investigate the associated variables to address our research questions.

Data Source and Sample

The data is sourced from the Program for the International Assessment of Adult Competencies (PIAAC) conducted by the OECD. PIAAC collects data on fundamental skills such as literacy, numeracy, and ICT, alongside extensive background information encompassing demographic, socioeconomic status, education and work history, and behavioral aspects of adult populations. PIAAC data offers nationally representative estimates of adults’ learning and development both within and outside the workplace, assessing their participation in various education and training activities pursued for professional or personal reasons. Hence, PIAAC data proves suitable for use, given the widely accepted recognition of AET's significance as a prerequisite for adults’ social inclusion and effective job performance (CEDEFOP, 2016). For our current study, we utilized the most recent U.S. PIAAC data collected in 2017. We selected respondents aged 25–65 years with recent work experience in the last 12 months, excluding those aged 16–24 years to align with the typical age range of AET participants, which is typically 25 or older (Desjardins, 2015). Our total sample size encompasses 1,334 respondents. ¹

Variables

Analytic Strategy

Scholars in the field of education research have been dedicated to understanding the various factors that influence individual experiences in both formal and nonformal learning settings. To achieve this, they have primarily relied on traditional statistical analysis methods. For example, many studies have used ordinary linear regression analysis to assess how numerous explanatory variables impact dependent variables and determined the statistical significance of these factors. While such studies have significantly advanced our understanding in this area, they do have certain limitations. Classical statistical approaches heavily rely on strong parametric assumptions. For instance, when conducting logistic regression analysis, researchers must satisfy assumptions such as independence of observations, linearity of independent variables and log odds, absence of multicollinearity, and no presence of outliers or influential observations in the data set. Failure to meet these assumptions can lead to biased results in the analysis. To address these challenges, we adopted an alternative analysis technique called RFCs. RFCs are a nonparametric, data-driven modeling strategy ³ that provides a different approach to analyzing data. By utilizing RFCs, we can overcome some of the limitations of classical statistical methods and gain a more comprehensive understanding of the factors influencing formal and nonformal learning experiences.

RFCs have gained popularity in the field of ML, and social scientists have recently started using them to calculate the relative importance of explanatory variables (Choi et al., 2020). RFCs are an ensemble ML technique that utilizes multiple decision trees. These decision trees are collected to construct forests that provide information on what factors most efficiently predict and explain the dependent variables without strong parametric assumptions. To understand RFCs, it's essential to comprehend how decision trees work.

Decision trees split the entire data set via a tree-based algorithm. The way the data is split depends on the level of measurement of the explanatory variables in the model. For instance, if an explanatory variable is binary (having values of 0 or 1), the tree will easily split the data based on these two values. On the other hand, if the variable is continuous, researchers often split the data based on its median value.

At each node in the decision tree, the Gini impurity is calculated for each independent variable. The Gini impurity represents the level of disorder or uncertainty in the data. Among the various possible splits at each node (for each variable), the decision tree chooses a variable and its corresponding magnitude to split the data that results in the lowest Gini impurity. This process continues until the tree fully grows or a predefined stopping criterion is met. In this study, the threshold and the independent variable used for the split were determined based on the default Gini impurity criterion provided the Scikit-learn Python package (Pedregosa et al., 2011).

Figure 1 shows an example of one of several decision trees that we created with our data set. This decision tree randomly took samples (in this case, 855 observations) from the entire data set and split them into several subsets to understand relatively important factors associated with individual participation in formal AET. ⁴ This plot has the following information in each node from top to bottom: (1) split decision (inequality), (2) the Gini impurity of the samples at the node (Gini), (3) the total number of unique samples at the node (samples), (4) the number of samples used and their categories (False, True) (through bootstrapping) to calculate the Gini impurity and splits of this particular decision tree (values), and (5) majority class at that node (class). As illustrated, the first variable used to split the data is job satisfaction. This variable is selected because job satisfaction is associated with the lowest Gini impurity. After the initial data set was divided into two subsets, the decision tree continued to split these subsets based on the independent variables associated with the highest variance reduction. The tree kept splitting the data according to the same rule.

OPEN IN VIEWER

Figure 1.

A part of decision tree example.

Random forest classification has two main sources of randomness. Firstly, RFCs build each decision tree using a random subset of the variables in the data set. By using only a subset of the available variables at each split, the algorithm can reduce overfitting and increase the diversity of the trees in the forest. Secondly, RFCs also utilize bootstrap bagging to generate multiple training sets for each decision tree. In the bagging process, samples are randomly selected from the data set with replacements to create a new training set. The randomness of RFCs is important, given that it allows the models to create a diverse set of decision trees. This helps reduce overfitting, improve the stability of the model, and increase the accuracy of its predictions. It also helps decorrelate the individual trees in the forest, making the model more robust to noise and outliers in the data (Prasad et al., 2006).

RFCs utilize multiple decision trees for each forest. In our study, each forest includes 100 decision trees and calculates the mode rankings of each independent variable. Due to the socioeconomic nature of our data set, we also implement a five-fold cross-validation strategy before concluding on the importance of factors on the dependent variable. This means that we create five forests that include a total of 500 decision trees using random sampling with replacement. This strategy helps us improve our model accuracy. It should be noted that RFCs are not parameterized models, therefore, do not provide correlation coefficients or their statistical significance. Therefore, the utilization of RFCs provides information on the relative importance of each independent variable but not on the directionality of these variables. For the purpose of our analyses, we deploy RFCs in Python 3 using the well-known Scikit-learn library (version 0.24.2) with 100 decision trees (Pedregosa et al., 2011).

A key benefit of using RFCs is the ability to compute VIs for our given data set. Here, we leverage permutation importance for assessing which independent variables are most crucial for the predictive accuracy of the trained RFCs. In our work, we introduce the concept of relative decisiveness, which measures the fraction of permutation importances associated with a particular variable. Permutation importances are calculated by randomly permuting one independent variable and then calculating its effect on the baseline accuracy. A larger reduction in accuracy corresponds to greater permutation importance. The relative decisiveness finds the ratio of an individual variable's permutation importance to the sum of all such importances obtained from the model and data. Permutation importance is a more robust indicator of feature importance compared to assessing which variable affects Gini impurity. ⁵

Although RFCs mitigate many potential issues related to parametric assumptions, they may still face challenges when dealing with highly correlated explanatory variables. To test whether our explanatory variables suffer from multicollinearity, we conducted Generalized Variance Inflation Factor (GVIF) tests that measure the multicollinearity in generalized regression models. Typically, a GVIF score of 1 means no multicollinearity, while a GVIF score of 5 indicates multicollinearity in data. Our analysis shows no multicollinearity issues that could either inflate or deflate the relative importance of each variable in our models (Figure 2).

OPEN IN VIEWER

Figure 2.

GVIF plots for job-related formal and nonformal AET participation models.

Important Factors for Job-Related Formal AET Participation

As mentioned previously, RFCs do not intend to provide quantitative interpretations. However, based on our analysis (see Figures 3 and 4), we assigned variables into five categories: (1) most important, (2) important, (3) somewhat important, (4) less important, and (5) least important factors related to individual participation in job-related formal AET. It should be noted that this categorization is rather subjective. RFCs fit process for this experiment led to a model with an average testing classification accuracy of 88%. ⁶

Firstly, our results indicate that age and literacy skills are the most important factors associated with formal AET participation. Followed by this first group, numeracy skill use at work and ICT skill use at work seem to be more important than other factors. This result shows that except for age, all three skills that individuals utilize at work are more important than any other factors in explaining working adults’ participation in formal AET. The third category (somewhat important factors) includes work flexibility and monthly income. The result shows that education level, managerial status, economic sector, organization size, and employment status are only important at the low level. As we can see from Figure 4, the variance of the mode rankings of these variables is wider, which suggests that the impact of these variables is rather undetermined. Finally, at the lowest level, our result suggests that job satisfaction and gender may be associated with formal AET participation, but they are not very important factors in explaining our dependent variable.

Important Factors for Job-Related Nonformal AET Participation

To understand the relative importance of various factors associated with individual participation in job-related nonformal AET, we also assigned variables into five categories of importance based on our results (see Figures 5 and 6). RFCs fit process for this experiment led to a model with an average testing classification accuracy of 71%. ⁷

Similar to the previous analysis, our results show that literacy skills variable is the most important factor related to nonformal AET participation. Additionally, two other skills that individuals utilize at work (ICT skill use at work and numeracy skill use at work) and organization size seem to be more important than other factors. The importance of organization size seems to be greater for individual participation in nonformal AET compared to formal AET participation. The third set (somewhat important factors) includes the economic sector, monthly income, and work flexibility. These variables are more important than other factors but less important than the first two sets of variables in structuring individual participation in nonformal AET. The fourth category (less important) includes education level and age. Unlike the previous analysis, age does not seem to be an important factor for nonformal AET participation. Finally, at the lowest level, our result suggests that job satisfaction, managerial status, gender, and employment status are the least important factors in explaining nonformal AET participation.

Summary and Discussion

The key findings indicate that age and skills use at work are the most crucial factors influencing formal AET participation, while skills use at work and organization size are the most important factors for nonformal AET participation. Specifically, the findings highlight the strong association between three types of skills use at work and both formal and nonformal AET participation. The findings imply that the extent to which working adults utilize job-related skills and knowledge plays a decisive role in their continuous learning and improvement. The acquisition and utilization of occupational skills emerge as foundational factors influencing AET participation (OECD, 2019), which aligns with previous studies emphasizing the link between skills utilization and AET participation (Tikkanen & Nissinen, 2018).

The empirical findings also support the notion that individuals with higher human capital are more likely to value additional learning experiences (Boeren et al., 2010). Proficient skills reinforce positive attitudes and behaviors toward further learning, thus leading to increased AET participation (Yamashita et al., 2019). Moreover, the proficiency and utilization of job-related skills are interconnected with labor market outcomes such as employment and labor force entry (OECD, 2016), highlighting the importance of continuous skill development throughout individuals’ careers to enhance the competitiveness of the workforce, organizations, and nations. Consequently, many countries recognize the need to expand LLL opportunities for adult workers, aiming to help individuals adapt to the rapidly changing world of work by acquiring and maintaining required skills (OECD, 2019).

Given these findings, it is crucial for policymakers to establish LLL policies that encourage and support working adults’ skills utilization and development. Such policies should incentivize employers to invest in the skills development of their workforce (OECD, 2016). By promoting continued investment in the skills of working adults, organizations can foster a competitive workforce that is adaptable to the evolving demands of the labor market. The study's results provide valuable evidence and rationale for the implementation of LLL policies that facilitate skills utilization among working adults and contribute to the overall growth and development of individuals and the economy (OECD, 2016).

Second, the study found that work flexibility and monthly income were also significant factors associated with working adults’ AET participation. The results support the idea that job discretion and autonomy influence adults’ engagement in further education (Tikkanen & Nissinen, 2018). When employers grant workers the flexibility to pursue learning opportunities, it can lead to improved skills development and increased productivity for both the individual and the organization (OECD, 2013). Creating flexible working conditions that facilitate learning-rich environments can enhance access to AET opportunities for employees (Rubenson & Desjardins, 2009). Additionally, the study reconfirms the well-established finding that income level is a major driver of AET participation (Boeren et al., 2010). Working adults’ financial status remains an influential factor that can either facilitate or hinder their participation in AET.

Third, factors explaining AET participation varied somewhat across different types of AET. When we compare working adults’ participation in formal and nonformal AET, the respondent's age was the most salient and decisive factor related to formal AET participation, whereas the relative importance of age was low in nonformal AET. Previous studies have shown a significant “age-participation effect” (Fouarge & Schils, 2009); all else being equal, the willingness to invest in LLL tends to decrease as people age. Similarly, in a UK study, the younger the age, the higher participation rates were found for formal AET (Egglestone et al., 2018). As such, researchers have supported the negative relationship between age and AET participation (Boeren et al., 2010; Desjardins, 2015; Roosmaa & Saar, 2017). Despite the generally accepted belief about age and AET participation, our findings should be interpreted with caution; RFCs do not tell us the direction of the relationship among the study variables. A reasonable conclusion is that age is more associated with formally designed learning that leads to college diplomas and certifications compared to nonformal learning. This is probably because there are still existing perceptions that higher education is more suitable for certain age groups’ career development.

Regarding work-related contextual factors, organization size ranked higher in nonformal AET participation compared to formal AET. Organization size is related to formalization, and larger firms with more centralized decision-making processes may influence social capital in the workplace (Grinyer & Yasai-Ardekani, 1981). Social support from colleagues and significant others plays a positive role in AET participation (Brown & Bimrose, 2018; Clochard & Westerman, 2020). Nonformal learning often occurs through interpersonal relationships and knowledge sharing (Jarvis, 2010), and therefore, a learning-rich work environment can foster nonformal AET participation by facilitating employee relations and support.

Significance of the Study

First, this study holds theoretical significance as it enhances our understanding of job-related AET participation. By utilizing the theoretical model of adult education participation (Boeren et al., 2010) and thoroughly reviewing individual and situational factors associated with AET participation, we contribute valuable insights to the field. The well-documented effects of selected independent variables on AET participation justify their inclusion in the RFCs model configurations for calculating VIs. The second major contribution lies in the methodology employed. We adopted nonparametric models using the ML technique—RFCs algorithm. ML techniques have gained popularity for their accuracy and reliability in prediction and have been increasingly applied in various research domains, including this study. The RFCs method effectively handles nonlinear and complex relationships within the data set, providing high predictive accuracy in determining the relative importance of target variables. As a result, the study's methodological significance lies in its novel approach to providing a comprehensive understanding of AET participation.

Limitations and Future Research

This study has some limitations that should be acknowledged. First, the study's findings are limited to Western cultural settings due to the study population being from the United States. To enhance the generalizability of the findings, future research could compare similar study constructs and methods with non-Western populations. Second, while RFCs, an ML-driven and nonparametric modeling strategy, are a robust and powerful analytic approach, they do not provide information about the direction of relationships between the study variables. To complement the VI measures, future studies could consider employing additional analytic methods that provide estimates of effect sizes for independent variables. Lastly, considering human capital theory, it is essential to analyze not only adult education but also various individual and work-related predictors to identify a comprehensive range of potential interventions (Cummins et al., 2018). For instance, focusing on workforce skilling could be a viable strategy to improve AET participation, but further research is needed to explore causal relationships among the studied variables.