Exploring school factors related to professional learning communities: a machine learning approach using cross-national data

ABSTRACT

In the ongoing endeavour to increase student learning, restructuring schools into professional learning communities (PLCs) remains a popular strategy globally. Multiple studies have investigated positive outcomes associated with PLCs for students and teachers, but limited knowledge exists about factors associated with well-functioning PLCs, such as leadership, organisation, policies, and student and staff composition. We apply machine learning (ML) to explore relationships between PLCs and a wide range of school factors using the Teaching and Learning International Survey (TALIS) 2018. TALIS 2018 provides unique data for this study since it includes substantial information about how schools are managed and the contexts in which they operate across a wide range of countries. We find support for some of the factors mentioned in the literature, as well as identifying other factors not previously explored. Finally, we discuss the potential for further research on how to create optimal conditions for teachers’ engagement in PLCs.

KEYWORDS:

• Professional Learning Communities (PLC)
• International Large Scale Assessment
• Machine Learning (ML)
• exploratory analysis
• school leadership

Introduction

Professional learning communities (PLCs), a concept that emerged in the 1990s, remain a widely used approach to restructuring schools globally. While the early literature is dominated by small-scale qualitative studies seeking to understand the nature of PLCs, there is a growing body of quantitative research investigating their relationship to a variety of different outcomes for teachers and students (Doğan and Adams 2018; Vescio, Ross, and Adams 2008). However, very few studies have investigated factors associated with the successful PLCs (Bellibas, Bulut, and Gedik 2016; Julie, Kruse, and John Tarter 2016). While some factors have been identified as positively associated with PLCs in certain contexts, there remains little empirical knowledge on creating optimal conditions for PLCs globally. This is an important question to raise given the international popularity of the concept and the lack of relevant comparative and cross-national studies.

As the concept of PLCs has been implemented across the globe, mainly drawing on the same body of literature, teachers working in PLCs are expected to engage in similar activities. As such, it can be argued that a “universal” set of factors related to the existence of school-based PLCs may exist. The aim of this paper is to explore which school factors are important in explaining the degree to which teachers perceive their school as a well-functioning PLC. We do this by treating PLCs as the dependent variable and by exploring a wide range of potential predictors. Because the study uses a large data set consisting of 8128 schools with 203 predictors from 42 countries/economies, we apply a learning-based approach to explore all relevant information available in the data.

Machine learning (ML) enables us both to determine which of the factors described in the literature are supported by the data and to identify new factors that may be important but yet unexplored, potentially developing new avenues of research on how to create optimal conditions for PLCs in schools. The use of ML for analysis of large data sets in social science is relatively new, but there is a growing interest (Justin, Roberts, and Stewart 2021; Masci, Johnes, and Agasisti 2018). We test and compare the predictive power of five different ML algorithms, interpret the results of the best-performing algorithm, and discuss our findings in relation to existing knowledge, with a special emphasis on factors that have a greater potential for policy relevance.

Literature review

Professional learning communities

The ultimate aim of implementing PLCs in schools is to improve student learning outcomes. This aim is pursued through close collaboration between teachers, who together continuously strive to improve the quality of their instruction and, thereby, student learning. They do this by clarifying precisely what they expect students to learn, how to assess whether these expectations have been met, and how to intervene when students do not learn as intended. Therefore, instructional practices in individual classrooms and the learning outcomes of every student become teachers’ shared responsibility. Teachers learn with and from each other, and students learn more as their teachers develop greater expertise in addressing student needs (Hargreaves et al. 2012; Hipp and Huffman 2010). While there is no consensus regarding a clear, universal definition of PLCs, and there are considerable differences in measures applied in the research (Lee et al. 2022), Hord and Hirsh highlights several key features of an effective PLC:

… the major goal of the PLC is staff learning together, with the staff’s learning directed to student needs. The staff learning occurs more deeply and richly in interactions and conversations in which staff members pursue intentional learning, share new knowledge, test ideas, ask questions, gain clarification, debate conclusions, and seek consensus on how to transfer new learning to practice.

(Hord and Hirsh 2008, 27)

In the growing literature on PLCs, there is an important distinction between two strains of research: In the first, PLCs are perceived as whole-school phenomena, which entails that factors such as supportive conditions and cultural norms are considered a part of PLCs. In the other strain, PLCs are considered to include only the teachers who meet, and factors like cultural norms in the school are considered as factors influencing PLCs (Zhang, Yin, and Wang 2020). Our approach is aligned with the first, considering PLCs as whole-school phenomena, as it is conceptualised by e.g. Hord (1997).

School factors related to PLCs

This section provides an overview of the literature, describing various factors that have been associated with the success of PLCs in schools and school departments. Because this is an explorative study and the body of literature is limited, we include sources across different contexts and methodological approaches, and we do not distinguish between the concept of PLCs and the similar concept of professional communities (Stoll et al. 2006).

Leadership

It is widely recognised that PLCs are unlikely to develop and function without active support from the school leadership (Hord, Sommers, and Hargreaves 2008). Literature on PLCs frequently highlights the importance of leadership, but it is unclear which specific leadership models, actions, and related contextual factors increase the likelihood of a successful PLC within a school. In the following, we outline research findings and hypotheses concerning how different leadership models – as well as combinations of different models – and specific behaviours by leaders influence the success of PLCs.

Transformational leadership, which includes the leader’s efforts to motivate, empower, and support teachers in making changes, has been found to have a positive influence on of PLCs (Zhang, Huang, and Xu 2022). Valckx, Vanderlinde, and Devos (2020) also found a positive relationship between transformational leadership and the component “collective responsibility”, as well as an indirect relationship between transformational leadership and “reflective dialogue” through “collective responsibility”.

Successful leadership of PLCs in schools has also been associated with the principal having a strong focus on instructional quality and with specific instructional leadership activities, such as observation of instruction by the principal and establishing and regulating norms for productive professional debate among teachers (Bryk, Camburn, and Louis 1999). Vanblaere and Devos (2016) found that instructional- and transformational leadership had positive associations with distinct components of PLCs, arguing that a combination of these two leadership models can support the successful implementation of PLCs. Keung et al. (2020) found a positive relationship between a leadership scale covering aspects of both transformational- and instructional leadership and kindergarten teachers’ engagement in PLCs.

Distributed- and Shared leadership likewise has a central position in the literature on PLCs (Goos and Martin 2019). For instance, Cherkowski (2012) has argued that sustainable learning communities are dependent on the leader’s ability to create an environment where teachers can work together to challenge existing beliefs and practices, building trust and nurturing shared commitment and responsibility among the community’s members. Similarly, there are arguments that the leader must engage in establishing and maintaining a high-trust environment, and in setting professional norms for civil discourse and productive debate among teachers (Hargreaves and Fink 2006). Other studies have emphasised that it is important for the success of PLCs that leadership is not concentrated on a single person, such as the principal; teachers and other staff must be included in school-wide and instructional decision-making processes (P. Chen et al. 2016; Ho, Ong, and See Tan 2020; Huffman et al. 2001; Vanblaere and Devos 2018).

In addition to specific leadership aspects and practices mentioned above, school leadership also plays a central role in allocating time and resources for teachers’ individual as well as collaborative practices within the school. To develop and sustain PLCs, it has been emphasised that teachers must be provided with the necessary time, space, and knowledge resources to engage in their respective professional communities on a regular basis (Huijboom 2022; McLaughlin and Talbert 2006). Finally, the educational- and professional development beliefs of the principal, as well as the principals’ fear of resistance from teachers, may influence how they perceive the concept of PLCs, which in turn affects their willingness to allocate time and resources for teachers to engage in PLCs (van den Boom-Muilenburg et al. 2021).

School autonomy

Few studies have incorporated the perspective of school autonomy in relation to PLCs. Bolam et al. (2005) found that national and local policies affecting the allocation of resources can inhibit PLCs in schools. Others highlight the importance of schools’ control over teacher recruitment, which allows for employing teachers who are supportive of the idea of a PLC, which can be difficult if recruitment is managed centrally (Bryk, Camburn, and Louis 1999; Valckx, Devos, and Vanderlinde 2018). Additionally, school leadership is recognised for its strategic role in initiating change, for instance through allocation of school resources, including time and personnel, which are necessary for implementing school-wide changes such as PLCs (McLaughlin and Talbert 2006). Therefore, the ability or capacity of schools to make decisions about key policies and practices could also be highly related to PLCs.

Student composition

Specific characteristics of a school’s students have also been highlighted as factors that can potentially inhibit or support the success of PLCs. Bellibas, Bulut, and Gedik (2016) found that schools with lower average student socio-economic status (SES) are less likely to develop PLCs. Similarly, Bolam et al. (2005) identified high percentages of students receiving free school meals, second-language students, and students with special needs are related to ineffective PLCs and argued that the social composition of the student body influences how the school functions, and that students’ ability to respond to changes affects the teacher’s ability to change.

Staff composition

Prerequisites for well-functioning PLCs are regular meetings and collaboration between members. Over time, trust and common expectations develop among the members of such communities; as such, a stable workforce is expected to foster PLCs (Bryk, Camburn, and Louis 1999; Hargreaves 2007). Related to the notion that professional communities take time to flourish, Bellibas, Bulut, and Gedik (2016) recognised a possible connection between low-SES schools and workforce stability, with teachers working in low-SES schools more likely to seek employment elsewhere. This can result in high employee turnover at low-SES schools, making it more difficult to create and sustain PLCs. Additionally, Valckx, Devos, and Vanderlinde (2018) found that male teachers are less likely to engage in specific PLC activities than female teachers are.

School size and location

The size of the school is among the most frequently highlighted contextual factors in research on the effectiveness of PLCs, with findings showing that PLCs are more common and tend to be more successful in smaller schools than in larger schools (Bellibas, Bulut, and Gedik 2016; Vanblaere and Devos 2016). A potential explanation is that smaller schools are generally found to be more engaging work environments, where teachers develop greater trust in colleagues and a stronger sense of belonging to a single professional community (Bolam et al. 2005). In larger schools, with a greater number of students and staff, individual teachers are less likely to see themselves as part of a single school community, and their communication and collaboration with colleagues may be more bureaucratic. Therefore, large school size may be considered an inhibiting factor in relation to PLCs, hindering a common sense of purpose, which is essential to an effective PLC. By contrast, one might expect very small schools to be less likely to develop PLCs due to insufficient critical mass for professional communities centred on different school subjects. For instance, there needs to be more than a couple of mathematics teachers for a PLC on the subject to function as intended.

Regarding the location, Bryk, Camburn, and Louis (1999) have posited that certain features of a school’s surrounding community may pose barriers to adapting teachers’ work and establishing a PLC. Authors also argue that barriers such as high rates of residential mobility and poverty may be more pronounced in urban settings.

Methods

Data

Data source and the variables

This study used data from the Teaching and Learning International Survey (TALIS) 2018. TALIS is an international survey of teachers and school leaders focused on working conditions, learning environments, and professional development. The sampling design of TALIS involves randomly sampling 200 schools (first stage) and then drawing a random sample of 20 teachers within each school (second stage) (OECD 2019). The sampled teachers, as well as the principal at the schools, are selected to answer questionnaires. The data from TALIS are available for secondary analysis, and in this analysis, we use data from the main population (ISCED level 2). While the official data from TALIS 2018 do not include a measure of PLCs, such a measure has been developed at the school level using teacher responses in a secondary analysis (Christensen 2022). We applied this measure of PLC, which is comprised of three subdimensions, with a total of 14 items (Table 1). We used this composite measure as our dependent variable in the analysis presented below.

Table 1. Sub-dimensions and items in the PLC measure.

Collaborative practice with focus on student learning
Main question text: On average, how often do you do the following in this school?	Answer categories Never Once a year or less 2–4 times a year 5–10 times a year 1–3 times a month Once a week or more
Items
Teach jointly as a team in the same class. Observe other teachers’ classes and provide feedback. Exchange teaching materials with colleagues. Engage in discussions about the learning development of specific students. Work with other teachers in this school to ensure common standards in evaluations for assessing student progress. Take part in collaborative professional learning.
Shared vision and responsibilities
Main question text: How strongly do you agree or disagree with these statements, as applied to this school?	Answer categories Strongly disagree Disagree Agree Strongly agree
Items
This school provides staff with opportunities to actively participate in school decisions. This school has a culture of shared responsibility for school issues. There is a collaborative school culture which is characterised by mutual support. The school staff share a common set of beliefs about teaching and learning.
Supportive conditions
Main question text: How strongly do you agree or disagree that the following present barriers to your participation in professional development?	Answer categories Strongly disagree Disagree Agree Strongly agree
Items
There is a lack of employer support.* Professional development conflicts with my work schedule.* There is no relevant professional development offered.* There are no incentives for participating in professional development.*

Answer categories are reverse coded.

As independent variables, we explored the questions included in the TALIS 2018 principal questionnaire, comprising a large set of items concerning external factors and the school’s location, leadership, organisation, and student composition (OECD 2024). The available data including the PLC measure comes from 42 countries/economies (Appendix 1).

Data preparation

Before completing the analysis and comparing the performance of multiple ML algorithms, some processing of the data was necessary. This section describes the steps involved in preparing the data for analysis.

Variable selection

The first step was to carefully examine the principal questionnaire to identify items that needed to be excluded from the analysis. We found several items that were only administered when specific answers were given to other items so that respondents did not answer questions that did not pertain to them (filtered items). Imputing data on such filtered items would be problematic as irrelevant information would be produced. On this basis, we excluded 14 question batteries, comprising 108 items. Having removed these items, we assessed the amount of missing data for each of the remaining items. One battery consisting of four items had more than 12% missing data for each item, and we excluded these variables as well. Finally, a battery of 11 items was excluded because they were very similar to the teacher-level items used in creating the PLC measure. Including those items would create an overlap between dependent and independent variables. A full list of excluded variables, and the reason for their exclusion, is presented in Appendix 2.

Data splitting and imputation of missing data

To ensure that the model generalises well to new data, we split the data into training and test sets, training our models on 75% of the data while reserving 25% of the data to test model performance on unseen data. The process of training models on one part of the data, the training data, and then evaluating the performance on the test data, is a common approach in ML. This process helps avoid overfitting, where the model becomes too specific to the training data and therefore does not perform well on new data (James et al. 2021). The data splitting was performed with stratification on the PLC variable, which ensures similar proportions of the outcome variable in each of the samples, making the results more reliable. The missing values in the training and test data were imputed using the K-Nearest Neighbors (KNN) algorithm to obtain complete data sets, as most of the algorithms cannot handle missing data. KNN was chosen because of its ability to handle categorical and binary variables and its computational efficiency compared to other imputation methods (Cihan, Kalıpsız, and Gökçe 2019). We used a parameter sweep of odd numbers from 1 to 19 to identify the optimal value of K: The test and training data were imputed using each of the k-values, and then a KNN model was trained using the same value and the performance in terms of MAE, MSE, and RMSE was extracted. The best trade-off between accuracy and complexity was given by a K-value of seven (Appendix 3).

Preprocessing

Finally, some pre-processing steps were necessary to accommodate algorithm-specific data requirements, specifically: Ordinal categorical variables were converted to numerical scores, nominal variables were dummy coded, variables with no variation were removed, and all numeric predictors were standardised to have a mean of zero and a standard deviation of one.

Analysis

Five regression algorithms were compared on predictive accuracy to identify the best-performing model of PLC behaviour within a school. The algorithms were chosen based on their wide use and popularity in predicting numerical outcomes, specifically: Classification and Regression Trees (CART), LASSO regression, Ridge regression, Random forests and XGBoost. CART is generally considered to have poor performance compared to the other algorithms but was included because it is easily interpretable and handles categorical predictors well. LASSO and Ridge regression are widely used extensions of linear regression, while Random forests and XGBoost are extensions of the logic behind CART, which handles non-linearity in the data well. All analysis was completed using the statistical software R, and to a large extent the “tidymodels” package (Kuhn and Wickham 2024; R Core Team 2022).

For each of the five algorithms, we performed a grid search for parameter optimisation with RMSE and RSQ as objective functions. We tested 1000 different tuning parameters for each model using 10-fold cross-validation. The results of this grid search showed that XGBoost was the most accurate on the training data (Appendix 4). The predictive performance of each of the five algorithms was evaluated on the test data. Results are presented in Table 2, which shows that XGBoost outperformed the other four algorithms, having the lowest RMSE value of 0.097 and the highest RSQ value of 0.294.

Table 2. Comparison of performance on test data.

Algorithm	RMSE	RSQ
CART	0.109	0.112
LASSO	0.099	0.275
Random forest	0.100	0.269
Ridge	0.099	0.276
XGBoost	0.097	0.294

XGBoost, short for extreme gradient boosting, is a flexible gradient-boosting algorithm that represents a powerful extension to decision trees. XGBoost segments data based on different variables, splitting the data points into smaller groups that share similarities with reference to the outcome variable. Like other gradient-boosting algorithms, XGBoost relies on ensemble learning – that is, an ensemble of tree-based prediction models that are used together to create predictions. In training an XGBoost model, each new tree model is informed by the performance of previous trees and an ensemble of trees is created in an iterative process (T. Chen and Guestrin 2016; DMLC 2022).

Findings

Interpreting the final model

Variable importance

Variable importance is a frequently used metric in supervised ML, offering a way of understanding which variables have the greatest impact on the predictions made by the model, and thereby how important they are for the predictions. In XGBoost, three different measures of variable importance can be calculated: the gain, the coverage, and the frequency. The gain is a measure of a variable’s relative contribution to the predictions. The coverage is a measure of the number of observations that have been split based on the given variable. Finally the frequency represents the number of times that a variable occurs across the trees in the model (T. Chen et al. 2022). We focused on the measure of gain. Figure 1 illustrates the relative importance of the 20 highest-ranking variables in the final model.

Figure 1. Variable importance of top 20 variables.

Variable texts are abbreviated, see Table 3 for full information.

Table 3. Description and interpretation of top 20 variables.

Rank	Variable name	Question text & response type	Interpretation based on plot	General direction
School and context factors
1	TC3G13C	Number of School administrative personnel (including receptionists, secretaries, and administration assistants). (Continuous)	The lowest values have a positive impact on predictions. For slightly higher values, the impact is negative, and this negative impact decreases for the largest values.	Complex/negative
2	TC3G13A	Number of teachers at the school (those whose main professional activity at this school is the provision of instruction to students). (Continuous)	Lowest values add to the predicted values, higher values subtract.	Negative
8	TC3G10	Which best describes this school’s location? (Ordinal)	Positive impact on predictions when schools are in rural areas, negative impact on predictions for schools in big cities.	Negative
9	TC3G45C	How strongly do you agree or disagree with the following statements? C: I am satisfied with the support that I receive from the staff at this school. (Ordinal)	No impact on predictions when principals respond “Strongly disagree”. Negative impact for “Disagree”, and a positive relationship with predictions from “Disagree” through “Strongly agree”.	Complex
10	TC3G13B	Number of personnel for pedagogical support (including all teacher aides or other non-teaching professionals who provide instruction or support teachers in providing instruction, professional curriculum/instructional specialists, educational media specialists, psychologists, and nurses). (Continuous)	Little/no impact among lowest values, positive relationship with predictions as values get higher.	Positive
12	TC3G13E	Number of “Other staff” employed at school. (Continuous)	Generally higher predictions for higher values.	Positive
14	TC3G27A	To what extent do the following statements apply to this school? A: Teachers understand the school’s curricular goals. (Ordinal)	Small negative impact on predictions for the first three values, and a positive impact for the highest values.	Positive
18	TC3G30E	In this school, how often do the following occur amongst students? E: Use/possession of drugs and/or alcohol. (Ordinal)	Negative relationship between the frequency and the predicted value.	Negative
School governance
3	TC3G20B1	Regarding this school, who has a significant responsibility for the following tasks? B1: Dismissing or suspending teachers from employment: Principal. (Binary)	Higher prediction when principal has “significant responsibility”, lower value when it is not the case.	Positive
4	TC3G09C	How strongly do you agree or disagree that the following present barriers to your participation in professional development? C: There is a lack of employer support. (Ordinal)	The more the principal agrees, the lower the prediction. The largest distinction is between categories 2 and 3.	Negative
5	TC3G09G	How strongly do you agree or disagree that the following present barriers to your participation in professional development? G: There are no incentives for participating in professional development. (Ordinal)	The more the principal agrees, the lower the prediction. The largest distinction is between categories 2 and 3.	Negative
Leadership
6	TC3G22C	Please indicate how frequently you engaged in the following activities in this school during the last 12 months. C: I provided feedback to teachers based on my observations. (Ordinal)	The more often the principal provides teacher feedback based on observations, the higher the prediction.	Positive
16	TC3G22K	Please indicate how frequently you engaged in the following activities in this school during the last 12 months. K: I worked on a professional development plan for this school. (Ordinal)	The more often the principal engages in planning professional development, the higher the prediction.	Positive
19	TC3G21A	On average throughout the school year, what percentage of time in your role as a principal do you spend on the following tasks in this school? A: Administrative tasks and meetings (Including regulations, reports, school budget, preparing timetables and class composition, responding to requests from district, regional, state, or national education officials. (Continuous)	Time spent on administrative tasks has a negative impact on predictions.	Negative
Principal characteristics
7	TC3G04D	How many years of work experience do you have regardless of whether you worked full time or part time? D: Year(s) working as a teacher in total (include any years of teaching). (Continuous)	No experience and many years of experience have little impact on predictions. Predictions are higher when principals have some but not many years of teaching experience.	Complex
11	TC3G42	For how many more years do you want to continue to work as a principal? (Continuous)	Positive relationship, which flattens out above a threshold. Only the lowest values have a negative impact on predictions.	Positive
13	TC3G04A	Principal experience A: Years working as a principal at this school. (Continuous)	Higher predictions for higher values. Only the lowest values have a negative impact on predictions.	Positive
15	TC3G04C	Principal experience C: Years working in other school management roles (exclusive of years working as principal). (Continuous)	Varying effect at lower values, and the largest positive impact at the higher values.	Complex
17	TC3G04B	Principal experience B: Year(s) working as a principal in total. (Continuous)	Higher values have a negative impact on predictions.	Negative
20	TC3G03	What is the highest level of formal education you have completed? (Ordinal)	Generally, higher levels of education have a negative impact on predictions.	Negative

Accumulated local effects

Measuring variable importance provided the relative importance of each variable in the model, but lacked information about the direction of the relationship between the predictors and the outcome. To investigate these relationships, we used accumulated local effect plots (ALE-plots) (Molnar 2019; Molnar, Bischl, and Casalicchio 2018). These ALE-plots illustrate how the prediction was affected, on average, across the range of possible values for the predictor in question. Such relationships may not be linear and can be complex, depending on the possible value that the predictor may take, and the patterns in the data. Figure 2 illustrates how the different values for each variable are associated with the prediction of the outcome variable. As an example, we found that “principal experience working as a teacher”, which is measured by item TC3G04D, did not have a linear relationship with our outcome. The ALE-plot for this predictor revealed that some experience, compared to either no or many years of experience as a teacher, was associated with higher predictions in our model, while many years of experience were associated with lower predictions on average.

Figure 2. Accumulated Local Effects for top 20 variables.

Variable texts are abbreviated, see Table 3 for full information.

Having obtained a list of the most influential variables and an illustration of their average influence across the respective values, we were able to describe the relationship between the predictors and the outcome. Table 3 contains the question text for each of the 20 most important variables and our interpretation of the association with the outcome interest. The variables are arranged by theme and impact.

Comparing the literature to our findings

The following section connects the findings from our analysis to the existing literature, in terms of whether or not they were closely related to previous findings, or if they indicated new themes.

Predictors previously identified in the PLC literature

School size is one of the predictors frequently mentioned in the existing literature, and we found multiple items related to it in our analysis (ranked as 1st, 2nd, 10th, and 12th). The first two predictors, which describe the number of administrative personnel and the number of teachers at the school, are aligned with the literature, showing that PLCs are more likely to be found in smaller schools (Vanblaere and Devos 2016). The predictors ranked 10th and 12th, the number of educational support staff, and the number of “other staff”, respectively, are positively associated with the measure of PLC in the school in the opposite direction. One explanation for the impact of the number of educational or other support staff (e.g. help with technology) could be that the availability of additional staff frees up time for the teachers to engage in PLC-related activities or provides additional support for teachers’ professional collaboration, since time is a factor frequently mentioned in PLC literature (Huijboom 2022; McLaughlin and Talbert 2006).

The 3rd most important predictor in the model is a binary determinant of whether the principal has significant responsibility for “dismissing or suspending teachers from employment”. This is related to a lack of school autonomy, describing limitations on making staff changes, which has been identified as an inhibiting factor in relation to creating effective PLCs in few studies (Bolam et al. 2005; Valckx, Devos, and Vanderlinde 2018).

Schools where the principal agree that there is no incentives to participate in professional development (ranked 5th) are associated with lower predictions of PLC, which is aligned with the finding of van den Boom-Muilenburg et al. (2021) that principals’ beliefs about professional development affect their willingness or motivation to allocate resources towards developing and sustaining PLCs.

The frequency of principal feedback to teachers based on observations is the 6th most important predictor in our model. The impact of the principal’s involvement or interest in instructional quality is supported by the literature linking instructional leadership to PLCs in various contexts (Vanblaere and Devos 2016; Zheng, Yin, and Li 2019). The 16th most important predictor, describing how frequently the principal engages in the planning of professional development activities at the school, further supports this.

Our model ranked school location as the 8th most important predictor, supporting previous findings suggesting that PLCs are more likely in rural areas than in bigger cities (Bryk, Camburn, and Louis 1999). Finally, we found a negative relationship between the frequency of student use/possession of alcohol and PLCs (ranked 18th), which potentially relates to the link between PLCs and student composition and/or neighbourhood context. Previous studies have established links between school climate and students’ use of drugs/alcohol, alongside related factors such as violence and absenteeism (Fisher and Kettl 2003; Henry and Slater 2007; Tomczyk, Isensee, and Hanewinkel 2015). Teachers play a pivotal role in these relationships as they can both influence the prevalence of such issues (Finn and Jeanette Willert 2006; Fletcher, Bonell, and Hargreaves 2008), and as these factors may affect quality of instruction and teacher retention (Fisher and Kettl 2003). Given these circumstances, our results may suggest that positive student behaviour or school climate in general could be a critical factor for effective PLCs to emerge. At the same time, strong PLCs may also be able to reduce such negative conditions (Maag 2009), while this is not the direct focus of the current study.

Predictors not mentioned in the PLC literature

We found that a lack of employer support for the principals’ participation in professional development was the 4th most important predictor in our analysis; principal agreement that there is a lack of employer support is associated with lower predictions. Although not related directly to PLCs, there is research on how the professional development of principals may improve their leadership practices and capacities (Goff et al. 2014; Grissom and Harrington 2010; Simkins et al. 2009). Therefore, not being able to participate in professional development might negatively affect principals’ capacity to develop PLCs in their schools.

The predictors related to principals’ experience gave multifaceted results: Experience working as a teacher (7th) and working in “other school management roles” (15th) have rather complex associations with PLCs. Experience working as a teacher has a positive influence at lower levels of experience, indicating that principals having some experience with the teaching profession positively predicts the presence of a PLC at the school, while the impact for many years of experience is negligible. Experience in other management roles is not important at the lowest values, but positively predicts PLC level at larger values. While this finding aligns with the result of an earlier study focusing on student achievement (Gümüş et al. 2021), it is difficult to uncover what the mechanism may be, due to the sparse research on this specific qualification. However, years of experience as a principal has an overall negative relationship with PLCs (17th). A potential explanation for this could be that experienced principals have a more traditional mindset, placing little value on close collaboration between teachers. Meanwhile, principals’ experience in their current position had a positive relationship with PLCs (13th), which could reflect the time it takes to foster the trusting relationships between teachers and leaders that are necessary when creating PLCs (Cherkowski 2012). We found a negative relationship between principals’ formal education and PLCs (20th). Although this result seems surprising, there are similar findings on the relationship between principals’ formal education and other school outcomes, such as student achievement (see Gümüş et al. (2021) for an overview). One possible explanation could be the increasing hierarchy between principals and teachers or the lack of collegiality when principals have a higher level of formal education (e.g. a Master’s degree or Ph.D.), with a potentially negative impact on professional collaboration.

There was a complex but mainly positive relationship between PLCs and the level of perceived support from staff (9th) in our model. This result could relate to the importance of teachers’ enthusiasm and support for the PLC process or to the literature on school autonomy, which has described how PLCs are more likely when the principal is able to hire teachers based on their willingness to participate in PLCs (Bolam et al. 2005; Valckx, Devos, and Vanderlinde 2018). We found that the number of years that principals would like to remain in their current position (11th) and whether they feel the teachers understand curricular goals (14th) are positively related to PLCs, both of which are factors that are not mentioned in the literature. Potential explanations might be related to the principals’ overall satisfaction with their schools and with the professional capacity of staff.

Finally, we found a negative relationship between PLCs and the amount of time the principal spends on administrative tasks and meetings (19th). While this is not explicitly mentioned in the literature on PLCs, greater attention to administrative tasks has been linked to less of a focus on instructional issues and professional development (Meyer and Macmillan 2001; Pollock, Fei, and David 2017).

Discussion

The important variables presented in this analysis cover a wide range of school factors, some of which have a greater potential for change than others. The size of the school, for example, which is a factor highlighted by much previous research as well as in our results, is not something that can be easily changed to support PLCs. In the following, we will highlight factors identified in our analysis, that have a greater potential for change, which makes them more relevant for informing further research into enhancing conditions for PLCs.

Leadership

Among the different leadership styles mentioned in the literature on PLCs, we found support for a positive relationship with instructional leadership through principals’ observation of teaching practice and feedback to teachers, as well as time spent on planning professional development. These specific tasks have been identified as some of the core responsibilities of instructional leaders (Hallinger 2010; Jared and Bowers 2018). This suggests that future research on leadership and PLCs should include the perspective of instructional leadership. Future studies might also focus more on the role of principals’ administrative obligations, given that we found a negative association with the amount of time principals spent on administrative tasks. While the existing studies highlight that principals’ time spent on administrative tasks and bureaucracy may lead to less of a focus on instructional issues (Gümüş et al. 2024; Hallinger 2005; Supovitz and Poglinco 2001), how the engagement in such tasks influence PLCs has not received much attention.

Principal’s qualifications

Principals’ experience as a teacher was ranked highly in terms of importance in our model. This experience may provide principals with some foundation to support PLCs in their schools to a certain degree, while a substantial amount of experience may be related to being more traditional mindset, similar to our explanation for the effect of the total experience of being principal. Arguably, this link has potential for further investigation, as teaching experience could become a more widely recognised prerequisite for principal appointment if additional studies support this relationship. The same applies to the negative relationship between principals’ formal education and PLCs. Principal experience and education have been frequently studied in relation to, for example, student achievement, but not with PLCs specifically (Bush 2018; Gümüş et al. 2021; Hitt and Player 2019).

Principal professional development

Two of our highest-ranked predictors were related to principal professional development; one described a lack of employer support for participating in professional development, and the other that the principal believed that there were no incentives to participate. Since this theme has received limited attention in PLC research (van den Boom-Muilenburg et al. 2021), further investigation is warranted to understand whether these relationships with PLCs may indicate the importance of the beliefs of the principal regarding professional development as well as the knowledge and skills that principals may acquire from it, or if they indicate that some contexts may be less favourable for engagement in professional development for both teachers and principals. Depending on the underlying causes, these factors have the potential for change through national or regional policies.

Specialized staff

Most of the items related to school size (measured in terms of the number of employees in different positions) are aligned with the finding that PLCs are more likely in smaller schools. However, the positive relationship between the number of educational support staff and “other staff” could be investigated further: It could be the case that teachers can invest more time in the core task of providing high-quality instruction through PLCs, when the school has sufficient specialised staff to assist with other tasks (Huijboom 2022). This could also relate to the specialised roles such as coaches, coordinators or head teachers which may be assigned specifically to support and develop PLCs in schools (Bolam et al. 2005; Talbert 2009; Vangrieken et al. 2015).

School autonomy

On the policy level, giving schools autonomy on staffing matters (here denoted as the principal having significant responsibility for “dismissing or suspending teachers”) is something that has only received limited attention in PLC research (Bolam et al. 2005; Valckx, Devos, and Vanderlinde 2018). Given that not all teachers may wish to engage in PLCs, schools aiming to implement PLCs may be hindered in this ambition if teachers are assigned to their school at the district or municipality level, rather than hired directly at the school. This again is a factor that has potential for policy changes, by giving schools more autonomy to handle staffing internally, should further studies support this finding.

Limitations

The findings of this study should be considered merely exploratory, with the potential to inspire further research on the factors that facilitate and inhibit PLCs in schools. In the approach taken here, we did not address statistical significance, unlike the inferential approach common in quantitative educational research. The data were pooled international data from 42 different countries/economies that participated in TALIS 2018, and we did not take country differences and probability sampling into account, as well as possibly important predictors not covered by TALIS. Instead, we used this large data set to investigate potentially “universal” predictors for the widespread use of the concept of PLCs. We make no causal claims about the relationships found in this explorative analysis. Nonetheless, in the discussion above, we tried to connect our findings to previous literature on PLCs and, when relevant, provide possible explanations that could inform further research and policy on how to create optimal conditions for PLCs in schools.

Conclusions

In this study, we utilised supervised Machine Learning (ML) to examine the relationships between various predictors and outcome of interest, Professional Learning Communities (PLCs), using a large dataset. ML serves as a valuable tool for exploring extensive data sets, especially when empirical studies are sparse, allowing us to “let the data speak”. This approach allowed us to assess the degree to which there was statistical support for the factors mentioned in the existing literature and to compare their individual impacts using a large cross-national data set. Many of the factors mentioned in the literature were among the most important predictors in our study, giving some support to the implicit assumption that, to a certain degree, PLCs have universal characteristics, transcending national contexts. We also identified several strong statistical relationships between factors not previously mentioned in the literature and provided possible explanations. Therefore, this study supports most existing findings and identifies potential directions for future studies investigating new factors or adding nuance to existing knowledge, thereby advancing our understanding of how to create optimal conditions for PLCs.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplemental data

Supplemental data for this article can be accessed online at https://doi.org/10.1080/03055698.2024.2369855.

Additional information

Funding

The work was supported by the Carlsbergfondet [CF19_0751].

Notes on contributors

Anders Astrup Christensen

Anders Astrup Christensen is currently a researcher at the department of Educational Sociology, Danish School of Education at Aarhus University, Denmark. His research primarily focuses on school organization and teacher practices, and how these factors relate to student learning and wellbeing, as well as inequality in education and international comparative research on these subjects.

Kristoffer Laigaard Nielbo

Prof. Kristoffer Laigaard Nielbo has conducted basic research in natural language and developed research infrastructures. His expertise lies in developing AI tools designed to analyze extensive text databases. Notably, his contributions to applied natural language processing have been recognized with awards and have created multiple tools for large-scale text data analysis, particularly in Danish. Furthermore, he has developed computational and data infrastructures that are used widely in Denmark and Scandinavia. Prof. Nielbo leads a large research center and infrastructure service at Aarhus University.

Sedat Gümüş

Sedat Gümüş, PhD, is an associate professor in the Department of Educational Policy and Leadership at The Education University of Hong Kong. His current research focuses on the relationship between school leadership and various teacher and student outcomes as well as the contextualization of leadership models and practices.