Talent management consists of the attraction, development, and retention of talent (Luna-Arocas & Danvilla-del-Valle, 2023). It is the process through which an organization attracts, develops, and retains employees who contribute to its competitive advantage (Kossyva et al., 2024). According to Macpherson et al. (2023), talent management is essential in the success of organizations, since it influences many work-related behaviors and outcomes (e.g., job satisfaction, motivation, turnover intention; Kumar, 2022). Given its influence on work outcomes, research has widely examined the role of talent management in turnover intention and concluded that, when effectively implemented, talent management practices help to reduce employee’s turnover intention (Graham et al., 2024; Kumar, 2022).

Turnover intention is defined as the degree to which an employee is willing to leave their current organization (Triningsih & Darma, 2024). One psychological mechanism through which the relationship between talent management and turnover intention works is employee engagement (Sheehan et al., 2019; Vermooten et al., 2019). Traditionally, research on employee engagement has been focused on work engagement (Saks et al., 2022). However, Saks (2006) introduced the concept of organizational engagement, supporting the fact that employee engagement is a multidimensional construct encompassing both work and organizational engagement. Work engagement consists of “a positive, fulfilling, work-related state of mind that is characterized by vigor, dedication, and absorption” (Schaufeli et al., 2002, p. 74). On the other hand, organizational engagement is characterized by a highly positive attitude toward the organization, a firm connection to it, and a willingness to contribute to its success (Saks et al., 2022).

Engaged employees are more likely to feel emotionally connected to their work and organization, which strengthens their commitment and reduces their desire of leaving the organization (Vermooten et al., 2019). Talent management practices, such as career development plans or recognition programs, are functional to increase employees’ engagement, by satisfying employees’ intrinsic and extrinsic needs (Marescaux et al., 2013; Memon et al., 2021; Shahzad et al., 2024). This in turn develops a sense of belonging and satisfaction, which reduces employees’ intentions of leaving the organization (Winarno et al., 2022).

Based on prior research, this study proposed the following hypotheses:

H1: Talent management practices (i.e., selection and recruitment, training and development, performance management, rewards and recognition, and career development) are negatively associated with employees’ turnover intention.

H2: Employee engagement mediates the relationship between talent management practices and turnover intention.

The studies included in this meta-analysis were initially identified through a previously conducted systematic review (Sinisterra et al., 2024), analyzing the relationship between talent management, employee engagement, and turnover intention. The systematic review followed PRISMA guidelines (Moher et al., 2009). The systematic search was performed in February 2023 in four popular databases (Scopus, Web of Science, Business Source Premier, and ProQuest), and was updated in May 2024. The complete search string is available in the Appendix, Table S1. The inclusion criteria used were: (1) empirical quantitative studies; (2) studies available in English or Spanish; (3) studies conducted in an organizational environment; (4) studies addressing one or more of the following talent management practices: selection and recruitment, training and development, rewards and recognition, career development, and performance management (these practices were selected for being the most widely studied talent management practices in relation with engagement and turnover intention, as including all possible ones would be unfeasible and conceptually inconsistent; Pandita & Ray, 2018); and (5) studies assessing the relationship between talent management, employee engagement, and turnover intention. Eventually, 43 were included in the systematic review. For more information about the search strategy and selection process, we forward to Sinisterra et al. (2024). In the current study, the articles identified in the systematic review were re-examined to select only those that provided sufficient statistical data (e.g., correlations, sample sizes) for inclusion in the meta-analysis. Eventually, this study included 29 studies that met the eligibility criteria.

Data was extracted independently by two different authors, and then their findings were cross-checked. Information retrieval during data extraction included: author name, sample size, year of publication, country of origin, type of culture, mean age of participants, percentage of women in the studies, study design, type of industry of participants, and correlation coefficients between variables.

The Appraisal tool for Cross-Sectional Studies (AXIS tool; Downes et al., 2016) was employed to assess the quality of the included studies. This tool comprises 20 questions covering different aspects of study design and quality, and the risk of bias in cross-sectional studies. Questions are answered using “yes”, “no”, or “don’t know”. A score of 1 was assigned to responses marked as “yes,” while “no” and “don’t know” received a score of 0. The general score, expressed as a percentage (0 to 100%) was used to group studies into three categories: high-quality studies (above 80%), moderate-quality studies (60%-80%), and low-quality studies (below 60%).

Pearson correlation coefficients between the variables of interest (i.e., talent management, engagement, and turnover intention) were extracted from primary studies. Afterwards, two types of analyses were carried out. First, each bivariate correlation separately was combined to obtain an overall effect and run moderator, publication bias, and sensitivity analysis. Second, OSMASEM (Jak & Cheung, 2020) was implemented to test a meta-analytic mediation model. Details on each of these procedures are provided in the next paragraphs.

To obtain an overall effect of each bivariate relationship, Pearson correlation coefficients were initially converted to Fisher’s Z-scores (Cooper et al., 2019). These Z-scores were then combined, and the resulting overall effect sizes were converted back to Pearson correlation coefficients, which are the values presented in the manuscript. In several studies, several Pearson correlations were available for a given bivariate relationship, often extracted from the same sample. To avoid biased estimates, statistical methods that account for the dependency among effect sizes extracted from the same study were implemented. Specifically, we applied three-level models (Van den Noortgate et al., 2013, 2015) together with a posteriori robust variance correction (Fernández-Castilla et al., 2021; Pustejovsky & Tipton, 2022). Besides confidence intervals, 95% prediction intervals were also calculated (using the standard errors provided by the robust variance correction). These intervals estimate the range within which the true effects are expected to fall for 95% of similar future studies (IntHout et al., 2016). Forest plots were created for each bivariate relationship, taking into account the presence of multiple effect sizes within studies (Fernández-Castilla et al., 2020).

When applying a three-level model, two sources of variance are estimated: the variability of effect sizes within studies (Level 2) and the variability of effect sizes across studies (Level 3). The magnitude of these variances, both between studies and within a study, was assessed using likelihood ratio tests. Subsequently, moderator analyses were conducted by introducing study characteristics as predictor variables in the three-level model. The results of these meta-regressions were corrected using robust variance correction.

For qualitative moderator variables, meta-regressions were performed only for categories that contained at least five effect sizes. When comparing more than two categories, multiple comparisons of their overall effects were conducted using Tukey’s correction to prevent inflated Type I error. When the moderator variables were quantitative, they were centered before analysis. Finally, analyses to detect the potential presence of publication bias were implemented: funnel plots were created and visually analyzed, three-level Egger regression tests were performed, and the statistic of the Trim and Fill test was obtained, as explained in Fernández-Castilla et al. (2021). If the distribution of effect sizes in the funnel plot was asymmetrically distributed at the bottom of the figure, and/or the three-level Egger regression test was statistically significant, and/or the value of the statistic was above two, we concluded that publication bias might be present. If that was the case, selection methods proposed by Vevea and Woods (2005) were applied to obtain an overall estimate adjusted for publication bias. Finally, an outlier analysis was conducted to identify extreme effect sizes. We calculated the studentized deleted residuals for each effect size (Viechtbauer & Cheung, 2010), looking for values that exceeded ± 1.96. Upon identifying potential outliers, we excluded them and reran all analyses to ensure robustness.

To fit the meta-analytic mediation analysis, we used OSMASEM approach. The input was the 3 x 3 correlation matrices of each study, containing the three correlations of interest. Through this approach, a structural equation model can be fitted on the observed correlation coefficients extracted from the studies (for more technical información, we refer to Cheung and Jak & Cheung, 2020 and Jak et al., 2021). This method uses full information maximum likelihood as estimation method, which makes it robust to the missingness of some correlations in some studies. In our study, we began by fitting a simple regression model to examine the relationship between talent management and turnover intention. If this relationship was statistically significant, we proceeded with a second model, incorporating engagement as a mediator variable. In this second model, we focused on the significance of both the direct effect of talent management on turnover intention and the indirect effect through engagement. If the second model showed that the direct effect of talent management on turnover intention was not significant, while the indirect effect was significant, we concluded that full mediation was present. If both the direct and indirect effects were statistically significant, we concluded that partial mediation was present. The implementation of this approach was carried out in R (RStudio Team, 2023) using the metaSEM package (Cheung, 2015).

From the initial set of studies identified in the systematic review (Sinisterra et al., 2024), which included 43 studies, a total of 29 articles met the inclusion criteria of this meta-analysis and were included.

The quality of studies was assessed using the AXIS tool, since almost all included studies used a cross-sectional design (Downes et al., 2016). Rating indicated high quality for 1 study, moderate quality for 26 studies, and low quality for 2 studies. The average score was 67.07%. Demographic information of the 29 studies included in the meta-analysis can be found in the Appendix, Table S2.

Among the existing qualitative variables, we were only able to analyze the type of talent management, culture, and industry, as the remaining variables had fewer than five effect sizes across all categories. None of the moderator variables significantly explained the variability observed across the effect sizes. Results from the moderator analysis can be found in Table 2.

Funnel plots for the three correlations of interest can be found in Figures 2a, 2b, and 2c. A visual inspection of the figures indicated that there was no substantial asymmetry in the distribution of effect sizes. The results of this visual analysis was confirmed by the results of the three-level Egger regression test, where it was found that the standard errors of the effect sizes did not predict their magnitude (B = 0.412, p = .864 for the correlation between talent management and engagement, B = 2.171, p = .585 for the correlation between talent management and turnover intention, and B = 3.447, p = .385 for the correlation engagement and turnover intention). Furthermore, the statistic of the Trim and Fill was below two for the three correlations ( = 0.169 for the correlation between talent management and engagement; and = 0 for the correlation between talent management and turnover intention and for the correlation between engagement and turnover intention).

Finally, regarding the sensitivity analyses, several potential outliers were detected. For the correlation between talent management and engagement, two correlations were reported in Babakus et al. (2017) and Zhong et al. (2016) that were smaller than the others and therefore showed studentized deleted residuals out of bounds. These Fisher’s Z were 0.11 and 0.17, respectively. Without these outliers, the overall effect became slightly larger (r = .49). As for the correlation between talent management and turnover intention, four Fisher’s Z were detected as outliers (Z = 0.81 from Dalal & Akdere, 2021; Z = 0.60 from Kossyva et al., 2024; Z = 0.70 from Rezwan & Takahashi, 2021; and Z = 0.55 from Wen et al., 2022). These effects had in common that they were all positive, whereas the Pearson correlation reported by most of the other studies was negative. Without these effects, the overall effect became -.41. Finally, for the correlation between engagement and turnover intention, the Fisher Z reported in the study of Kossyva et al. (2024), which was .60, was detected as an outlier. After removing this effect, the overall effect remained almost the same (r = -.42).

Moving on to the next phase of the analysis, the results of the meta-analytic mediation model can be found in Figures 3a and 3b. In Figure 3a, it can be seen that Talent Management significantly predicts turnover intention. When Engagement is entered as a mediator variable (Figure 3b), this effect drops from -.331 to -.189 but remains statistically significant. Furthermore, the indirect effect of Talent Management on turnover intention through Engagement is statistically significant (indirect effect = -.142, 95% CI [-0.215, -0.086]). All these results suggest that Engagement “partially mediates” the relationship between Talent Management and turnover intention.

Despite the relevant contributions of this meta-analysis, it is also important to consider its limitations. The first limitation (1) of this study is the lack of complete information in some studies, specifically regarding the mean age. Sixteen studies did not report the mean age of participants. To address this, emails requesting this information were sent to the respective authors, and only four of them responded with the required information. Additionally, some studies reported the distribution of participants across generational intervals. However, this data cannot be used in meta-analytic analyses, which require the mean age of participants. Therefore, the moderator analysis for age was conducted using only 17 out of the 29 included studies. Although the quality of included studies, as measured by the AXIS tool, was moderate and acceptable, the available data was not enough for robust meta-analytic analyses. Previous studies showed differences in perception of talent management, employee engagement and turnover intention depending on age (Chen et al., 2020; Douglas & Roberts, 2020; Ramaprasad et al., 2021). However, the influence of age on these variables in the current study was not significant. This could be explained by the fact that missing data could have influenced the results of the moderator analysis related to age, potentially limiting the robustness of the findings in this regard. In addition, variables such as organization tenure could not be examined as moderators due to the lack of this information in the included studies. Without complete information, it will not be able to achieve robust and comprehensive conclusions. Future research should address this limitation by reporting all necessary information about demographic information, specifically the mean age of participants and its standard deviation.

Second (2), the limited number of effect sizes in some categories restricted the scope of analyses. It has not been possible to analyze the separate effect of each talent management practice on engagement and intention to quit, in order to examine if some practices have greater impact than others, because we did not have at least five effect sizes of each practice. It happened the same with “industry” as moderator variable. Additionally, outlier analyses revealed the presence of extreme effect sizes in certain studies. However, sensitivity analyses confirmed the overall stability of the results. In future studies, talent management practices could be examined separately in relation to each of the dependent variables (i.e., employee engagement and turnover intention), rather than as a total measure of talent management. Therefore, this will allow for the identification of differences in the effect sizes of each talent management practice on employee engagement and turnover intention, providing evidence of which practice has a greater impact on these variables. This information could guide organizations in the design and development of their talent management strategy, by providing information about which practice should be prioritized to increase the desire of employees to stay in their organization.

Third (3), the heterogeneity in how talent management practices and turnover intention were measured across studies has represented a limitation in achieving more robust and wider results. As highlighted by Sinisterra et al. (2024), the extant literature on talent management fails to capture an appropriate conceptualization and definition of the construct. Despite the many approaches to the concept, there is still no consensus on the definition of talent management, nor is there a validated universal scale for its measurement. Consequently, it has not been possible to analyze whether variations in the way talent management is measured influence the results, as each study used a different scale. A similar challenge occurs with the measurement of turnover intention, complicating the ability to compare findings across studies. In addition, as there is no consensus on the practices that conform the concept of talent management and the list of proposed practices can be endless, this meta-analysis included the five most studies ones (selection and recruitment, career development, performance management, training and development, and rewards and recognition; Pandita & Ray, 2018). However, this selection may limit the robustness of the findings, since some interesting studies considering other talent management practices may have not been identified or excluded. Future research should make efforts to arrive at a consensus in the definition and conceptualization of talent management and to develop a universal talent management scale. In addition, when conceptualizing talent management, researchers should arrive at an agreement of which practices are included within this concept, so a broader systematic review and meta-analysis including all talent management practices and their relation with employee engagement and turnover intention could be conducted.

Fourth (4), the limited literature addressing organizational engagement – only 2 out of the 29 studies included in this meta-analysis considered this dimension of employee engagement – represents another limitation of this study. This lack of research has made it impossible to examine whether the mediating effect of employee engagement on the relationship between talent management and turnover intention differs when considering work engagement versus organizational engagement. Future studies should consider organizational engagement as an important dimension of employee engagement, and study its mediating role between talent management practices and turnover intentions. The limited research regarding organizational engagement has identified it as an important antecedent of employee wellbeing (Farndale et al., 2014; Saks et al., 2022) and has confirmed its mediating role in the relationship between talent management and turnover intention (Juhdi et al., 2013; Saks, 2006). However, studies including organizational engagement in these relationships should be increased to validate these results.

The fifth limitation (5) of this meta-analysis is that the majority of studies are cross-sectional. Cross-sectional studies gather data at a single point in time, difficulting the prediction of causal relationships between variables (Wang & Chen, 2020). Future studies should consider examining these relationships using a longitudinal design, to infer causal relationships over time.

A further limitation (6) of this metanalysis concerns the search strategy. Although electronic database searches are a standard and well-established method for identifying relevant studies, best practices in meta-analysis recommend complementing these with additional procedures, such as searching dissertation databases or conducting manual searches of key journals (Harari et al., 2020). The absence of these strategies may have limited the comprehensiveness of the study selection process and should be considered when interpreting the findings.

Finally, the last limitation (7) of this study is the analysis of demographic variables (i.e., gender, age, and culture) independently, rather than considering its combined effects, especially in increasingly diverse workplaces (Lahey & Oxley, 2021). Future studies should address this limitation by examining how demographic variables interact to influence workplace outcomes. Understanding these joint effects is crucial for promoting inclusive organizational strategies that satisfy the unique needs of diverse employee groups (Croitoru et al., 2022; Drydakis et al., 2023).