Introduction: The What Works Clearinghouse (WWC) reviews rigorous research on educational interventions with a goal of identifying "what works" and making that information accessible to educators and policymakers. The WWC has historically prioritized internal validity over external validity in rating the quality of research. One critique of the WWC is that there has been too little attention to external validity (Hedges, 2018; Joyce, 2019; Ming & Goldenberg, 2021). The larger concern is that paying too little attention to external validity could lead to inaccurate conclusions about the potential effectiveness of interventions for different student populations (Briggs, 2008; William, 2019). Researchers have also found that educators are more likely to use evidence as a factor in decision-making when the studies are implemented in similar contexts to their own (Mills, Lawlor, Neal, Neal, & McAlindon, 2020; Nakajima, 2021). This white paper uses publicly available WWC study data to explore both the main sample and subgroup findings included in the WWC evidence base. This paper addresses the following research questions: (1) Are there systematic differences in subgroup findings relative to the full sample findings (within studies)?; (2) Are there compositional effects (across studies) related to student characteristics?; and (3) How much treatment heterogeneity is there in the findings? Data: The analytic sample includes 3,933 full sample findings and 3,427 subgroup findings from 1,045 studies. The advantage of using WWC study data is that studies must demonstrate internal validity and outcome measures must demonstrate minimum reliability requirements to be included (WWC, 2022). Therefore, the WWC data allow for the examination of effect size differences without poor study quality as a potential confounder. Method: Multivariate meta-analysis with robust variance estimation was used to account for the dependency of multiple findings within the same study and to identify systematic differences in effect sizes for each subgroup (Pustejovsky , 2019; Viechtbauer, 2010). Models were estimated separately for each subgroup. For the first research question, the models include a dummy indicator to distinguish the subgroup from full sample findings, fixed study effects to hold constant differences across studies, and dummy indicators for outcome domains, which can vary within studies. For the second research question, the models include a dummy indicator to distinguish the subgroup from full sample findings, the percent of students in the full sample with the subgroup characteristic, and dummy indicators for outcome domains--note that sample characteristics are only available at the study level, not at the findings level. The idea is to see to disentangle the subgroup effects within studies and compositional effects across studies (Pustejovsky et al., 2023). Findings: Findings show small positive subgroup effects within studies for Black, economically disadvantaged, and female students, and small negative subgroup effects for high-performing students, on average. This means that some program evaluations found differentially positive and negative effects for some subgroups, although the differential effects were small and largely underpowered. However, it is possible that there is more publication bias in the subgroup findings than the full sample findings, which would yield larger observed effect sizes for subgroups in general. The next step for this paper is to explore the degree of publication bias in the full sample versus subgroup findings. Next, findings show that there are also compositional effects, meaning that effect sizes systematically differ according to the characteristics of the student sample for reasons other than differential treatment effectiveness for particular subgroups. In other words, there may be contextual factors correlated with sample characteristics that influence the magnitude of effect sizes. The proportions of economically disadvantaged and special education students are each positively associated with effect size magnitudes, and the proportion of female students is negatively associated with effect size magnitudes, although the relationships are small. These compositional effects may have nothing to do with the effectiveness of interventions, and may instead relate to measurement issues across studies, such as outcome measure type and degree of variation in the outcome measures (Wolf & Harbatkin, 2022). On the other hand, compositional effects may be a (albeit weak) signal for where interventions may be more or less effective, due to things like available resources in the setting or the lack of an equally effective business-as-usual condition. Finally, there is substantial treatment heterogeneity across and within studies, and the subgroup and compositional effects have limited utility in explaining treatment heterogeneity. The degree of treatment heterogeneity is so large that consumers of research may need more than quantitative research to determine which interventions are likely to work in their settings. Consumers of research need to consider the theory of action of the intervention, implementation fidelity and quality, other contextual factors in studies, the full range of findings for the intervention or similar interventions, and the outcomes most relevant and important to them.