- Open Access
Effect size measures for multilevel models: definition, interpretation, and TIMSS example
© The Author(s) 2018
- Received: 2 January 2018
- Accepted: 17 July 2018
- Published: 23 July 2018
Effect size reporting is crucial for interpretation of applied research results and for conducting meta-analysis. However, clear guidelines for reporting effect size in multilevel models have not been provided. This report suggests and demonstrates appropriate effect size measures including the ICC for random effects and standardized regression coefficients or f2 for fixed effects. Following this, complexities associated with reporting R2 as an effect size measure are explored, as well as appropriate effect size measures for more complex models including the three-level model and the random slopes model. An example using TIMSS data is provided.
- Multilevel model
- Effect size
The reporting of effect sizes in quantitative research lends interpretability and practical significance to findings and provides comparability across studies (Kelley and Preacher 2012). Further, providing effect sizes in a study can aid future researchers in conducting meta-analysis to synthesize findings from multiple studies (Denson and Seltzer 2011) and power analysis to plan future studies (Kelley and Preacher 2012). Effect size has been defined in various ways, but recent work considers it “a quantitative reflection of the magnitude of some phenomenon that is used for the purpose of addressing a question of interest” (Kelley and Preacher 2012). More specifically, measures of effect size can broadly be categorized as either measures of variance explained (Tabachnick and Fidell 2007) or measures of standardized effect size (Snijders and Bosker 2012).
Because of these important benefits, organizations such as the American Psychological Association (APA) and American Educational Research Association (AERA) are moving toward stronger language requiring the reporting of effect sizes (Kelley and Preacher 2012; Peng and Chen 2013). Many educational and psychological journals are also following suit in requiring effect sizes in tandem with a de-emphasis on null hypothesis significance testing (NHST; Kelley and Preacher 2012). In response to these guidelines, researchers have found that the prevalence of effect size reporting has increased, although there is room for improvement in both prevalence as well as quality of reporting (Peng and Chen 2013). Although the importance and usefulness of effect size reporting is clear, guidance regarding effect size measures for multilevel models is scarce. Further, multilevel models may be particularly relevant in cross-cultural educational research using international datasets due to the nesting of data (i.e. students within schools within countries, etc.).
This paper provides guidance regarding choice and interpretation of effect size measures for multilevel models. A demonstration using large-scale survey data is provided for each topic. Assessing effect size for random effects is demonstrated using the ICC. Following this, assessing effect size for fixed effects is demonstrated using standardized regression coefficients and f2. Lastly, complexities associated with additional topics, including three-level models, R2 as a measure of variance explained, and models with random slopes will be explored.
An example is provided based on a multilevel model estimated from IEA’s Trends in International Mathematics and Science Study (TIMSS) 2011 fourth grade mathematics data. Note that although efforts were made to ensure the data and models estimated are substantively relevant and realistic, this applied analysis is primarily for demonstration purposes. Ten countries were randomly selected (Bahrain, Czech Republic, Denmark, Iran, New Zealand, Norway, Slovak Republic, Spain, Sweden, Tunisia) and used as the level-2 unit. All data preparation and analysis was done in R (R Core Team 2014) and multilevel models were estimated with the lme4 package (Bates et al. 2015) unless specified otherwise. R syntax for each model is provided in the appendix.
The unit of analysis is students (level 1) and the nesting variable is country (level 2). The sample included a total of 46,475 students nested within 10 countries. The largest country sample size was 5760 and smallest was 3121 with an average country size of 4648. In addition, demonstrations involving 3-level models included school membership as a level. The sample included 1817 schools. The largest within-school sample size was 93 and the smallest was 2 with an average school size of 25.6. Missing data was addressed with default listwise deletion resulting in a final analytic sample of 44,800 students.
The outcome variable is mathematics achievement. TIMSS provides five plausible values, which are multiple imputations of the latent construct (Wu 2005), for this variable. Only the first plausible value was used for analysis (ASMMAT01). Using only one plausible value is not preferred compared to using all five plausible value (Rogers and Stoeckel 2008); however, since using only one plausible value has been shown to typically recover population parameters (Rogers and Stoeckel 2008; Wu 2005) and since the present analysis is primarily for demonstration purposes, use of only one plausible is used simply to avoid overwhelming the reader with many new topics at once. Further, it should be noted that it is inappropriate to use an average of plausible values for analysis, which was therefore not done in the present analysis (Rogers and Stoeckel 2008). The mean for math achievement was 475.18 and the standard deviation was 93.77.
Independent variables include Female (ITSEX; recoded 0 = boy, 1 = girl); whether the student has an internet connection at home (ASBG05E; recoded 0 = no, 1 = yes); and the students’ confidence with math (ASBGSCM, continuous). Overall, the sample was 50% female and 78% of students had an internet connection at home. The mean confidence value was 10.1 and standard deviation was 1.92, although this variable was standardized for most analyses.
Although a detailed description of sufficient sample size for multilevel modeling is beyond the scope of this paper, it should be briefly mentioned that researchers differ in their recommendations. For example, one rule of thumb, the 30/30 rule recommends a minimum of 30 groups, with at least 30 members in each group (Hox 2010) whereas other researchers suggest that with as few as 10 groups, modeling with random rather than fixed effects is appropriate (Snijders and Bosker 2012). Another simulation study specifically recommends at least 50 groups to avoid bias is certain parameter estimates (Maas and Hox 2005). Ten groups were used for the present study, consistent with researcher recommendations (Snijders and Bosker 2012). The primary purpose of analyses in the present study is demonstration; however, applied researchers should be cognizant of differences in the probability of biased parameter estimates and standard errors at various sample sizes.
For the example analysis, the ICC was 0.27 (computed from Eq. 3 based on estimates from Eq. 1) indicating that the proportion of variance in math scores explained by country membership is 0.27. It is unclear what a typical ICC for achievement might be when considering nesting of students within countries, but understanding the extent to which countries differ may be an important first step for further investigating differences in academic achievement between countries.
Fixed effects, such as intercept or slope coefficients, depend on the scale of the independent variable, and so are not comparable across studies or among multiple variables within a single study. Some researchers suggest Cohen’s d, which is a measure of the standardized mean difference between two categories in a binary variable, as a measure of effect size for a binary covariate in a multilevel model (Snijders and Bosker 2012; Spybrook 2008) and by analogy, one could imagine representing the effect size for a continuous covariate as the correlation between that covariate and the outcome variable (for example, between confidence and math achievement in the TIMSS example). However, if Cohen’s d or a bivariate correlation coefficient were to be simply computed (i.e. based on the one-level model with no covariates), it would not be recommended as an effect size measure since it represents the relationship between the two variables without controlling for level-2 unit membership and other associated covariates.
Instead, the standardized coefficients can be used (Ferron et al. 2008; Snijders and Bosker 2012). These can be obtained by standardizing (i.e. M = 0; SD = 1) each variable before analysis (Ferron et al. 2008) or by standardizing each regression coefficient by multiplying it by the standard deviation of X and dividing by the standard deviation of Y (Snijders and Bosker 2012).
For the example analysis, Eq. 2 was estimated, and the standardized coefficient for Female is 0.004 (p > 0.05); for Internet is 0.186 (p < 0.05) and for Confidence is 0.262 (p < 0.05). This indicates that the coefficient for Female is not significantly different from zero; that one standard deviation increase in the Internet variable is related to 0.186 expected standard deviations increase in math achievement; and that one standard deviation increase in the Confidence variable is related to 0.262 expected standard deviations increase in math achievement, controlling for associated covariates.
Although these measures are now comparable, the interpretation for binary covariates, such as Female, may still be somewhat difficult; instead, the researcher can dummy code the binary covariate and standardize the outcome variable resulting in partially standardized coefficients. For the example analysis, this results in a coefficient of 0.008 (p > 0.05) for Female and 0.45 (p < 0.05) for Internet indicating that females are not significantly different from males on math achievement, when controlling for associated measures, and that students with internet connection are expected on average to score 0.45 standard deviations higher on math achievement, controlling for associated measures.
Both standardized and partially standardized coefficients provide information about the magnitude of the effect (after controlling for other covariates and nesting) and these measures are generally comparable amongst themselves and across studies with similar populations (and analogously similar analytic samples). One limitation is that these measures are dependent on the sample standard deviation of each variable, which will be particular to the given sample, and may vary from sample to sample. When working with large samples however, this sampling variability, or sample to sample variation, should be quite small meaning that comparability should typically not be a problem.
In the present example with a random effect for school membership added, there are 46,475 students nested within 1817 schools nested within 10 countries. Based on random effect estimates from the empty model (Eq. 7: φ2 = 2376; τ2 = 1906; σ2 = 4778), all three ICC values were computed as measures of effect size for random effects. Results indicate that ICC at level 3 (Eq. 8) is 0.26, the first version of ICC at level 2 (Eq. 9) is 0.21 and the second version of ICC at level 2 (Eq. 10) is 0.47. Each of these three measures provides different information and is interpreted in a different way. The present results indicate that 26% of variance in math scores is accounted at the country level, 21% is accounted at the school level; and 47% at the level of schools nested within countries.
Four random effects will be estimated: Var(eij) = σ2; Var(U0j) = τ 0 2 ; Var(U1j) = τ 1 2 ; Cov(U0j, U1j) = τ01. Interpretation can proceed in multiple ways. First, since the model assumes that the distribution of U1j terms (the difference between the overall slope and group-specific slope for each group) is normally distributed with variance τ 1 2 , the range within which 95% of level 2 units would fall can be computed (Snijders and Bosker 2012). In the present example, the average slope for internet (β2) was estimated as 0.17 (t = 8.77, p < 0.05) and τ 1 2 = 0.003 (SD = 0.06). So a hypothetical country with a “high” slope could be computed by starting with the average slope and adding two times the square root of τ 1 2 . Analogously, a hypothetical country with a “low” slope could be computed similarly, but by subtracting two times the square root of τ 1 2 . In the present example, this would imply that 95% of countries would have a slope for Internet between 0.05 and 0.29. Since these represent standardized coefficients, these slopes can be interpreted in the same way as fixed effects based on standardized covariates. Since there is no clear effect size measure to aid in interpretation of the random slope, this interpretation may offer a helpful alternative in the spirit of presenting the scope of the effect.
Researchers may also want to interpret the covariance term, τ01. As with any covariance, this term can be standardized to report the correlation and interpreted according to effect size criteria for r: 0.10 is small; 0.30 is medium; and 0.50 is large (Cohen 1992). In the present example, the correlation is 0.07 indicating that countries with higher slopes tend to have slightly higher intercepts, but that this relationship is fairly weak, according to Cohen’s (1992) criteria. However, caution regarding the intercept interpretation should be applied. The intercept will be specific only to the case where all predictors (X) are equal to zero (Snijders and Bosker 2012). In the present example, this implies that the relationship between Internet and Math is stronger for countries with higher average math scores for average Internet connectivity, although the effect is relatively small. If Internet had not been grand mean centered, the intercept-slope covariance term would have taken a different value based on the fact that the intercept would take a different meaning. Thus, centering must be carefully considered within the context of a random slopes model due to the fact that interpretation of the intercept-slope covariance parameter depends on how covariates are centered. Centering does not, however, necessarily directly contribute to issues of comparability across studies.
It is clear from the plot that if Internet were recoded with a different substantive meaning for the value zero, the intercept variance could change, as well as the intercept-slope covariance. For example, if 2 was subtracted from each value for Internet, the y-intercept would be further right on the plot (Fig. 1) at the location where Internet = 2. In that case, the y-intercept for each country would be different and the variability of these intercepts (τ 0 2 ) as well as the correlation between these intercepts and the slopes (τ01) would therefore differ as well. As a researcher interprets and attempts to explore the scope of these effects, their conditional nature should be considered and emphasized accordingly.
Model comparison for random slopes
No random slope
With random slope
Lower BIC indicates a better fit, and a difference of greater than 10 indicates “very strong” evidence for the more complex model (Raftery 1995) which is provided in the present example. Although this model comparison can aid in selecting a model, and even possibly assess the strength of that evidence (i.e. BIC), the evidence cannot be used as a measure of effect size. Each of the model fit indices (AIC, BIC, and log-likelihood) are a function of deviance which is, among other things, a function of sample size (McCoach and Black 2008) and so therefore would not be appropriate as measures of effect size.
Measures of variance explained are also generally inappropriate for random slopes, because allowing the slope to differ for each level 2 unit does not necessarily explain additional variance. It can be noted that the scope of the effect of random slopes is represented by the variance of those slopes. Although there is not strictly a measure of standardized effect size or variance explained for variance terms, the square root of this variance (square root of τ 0 2 ) is a standard deviation which is considered an interpretable measure of a distribution’s spread (Darlington and Hayes 2017).
One other possibility would be to approximate the scope of the effect by modeling the level 2 unit as a fixed effect (if this is substantively plausible) and including the interaction between the covariate and each dummy indicator for group membership. Based on this conceptualization, any method appropriate for effect size of a set of fixed effects (i.e. the interaction terms) would be feasible.
Many large-scale datasets involve complex sampling plans and other additional complexities which may need to be considered during analysis, including use of plausible values for achievement measures, inclusion of sampling weights due to non-equal probability of selection, and inclusion of replicate weights to account for multi-stage sampling. These aspects of analysis are beyond the scope of the present examination of effect size, but they should not be ignored. Several resources are available to learn more about these topics (Martin and Mullis 2012; Meinck and Vandenplas 2012; Rogers and Stoeckel 2008; Snijders and Bosker 2012; Wu 2005) and specific software is available to aid in analysis, such as the BIFIE package (BIFIE 2017) which is implemented in R.
Reporting measures of effect size is a crucial part of interpretation for applied multilevel modeling studies. Researchers can use the ICC to represent the magnitude of random effects which could represent country and/or school effects and standardized regression coefficients or f2 to represent the magnitude of fixed effects which may represent relationships of interest when examining substantive questions using international datasets. Complexities associated with three-level models, reporting R2, and random slopes have been explored. The topics in the present study have been demonstrated using TIMSS data, but the suggestions provided could be applied to any multilevel analysis of primary or secondary data.
The author read and approved the final manuscript.
The author would like thank Chao-Ying Joanne Peng and Patricia Martinkova for their comments on prior versions of this manuscript.
This research is based on work presented at the IEA International Research Conference (IRC) in June 2017.
The author declares that she has no competing interests.
Availability of data and materials
TIMSS data is publicly available from the IEA. R syntax used for the analyses in this paper is provided in Appendix.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
- Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing and interpreting interactions. Newbury Park: Sage.Google Scholar
- Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed effects models using lme4. Journal of Statistical Software, 67(1), 1–48. https://doi.org/10.18637/jss.v067.i01.View ArticleGoogle Scholar
- BIFIE. (2017). BIFIEsurvey: Tools for survey statistics in educational assessment. R package version 1.13-24. https://CRAN.R-project.org/package=BIFIEsurvey. Accessed 29 Jan 2018.
- Cohen, J. (1992). A power primer. Psychological Bulletin, 112, 155–159.View ArticleGoogle Scholar
- Darlington, R. B., & Hayes, A. F. (2017). Regression analysis and linear models: Concepts, applications, and implementation. New York: Guilford Press.Google Scholar
- Denson, N., & Seltzer, M. H. (2011). Meta-analysis in higher education: An illustrative example using hierarchical linear modeling. Research in Higher Education, 52, 215–244.View ArticleGoogle Scholar
- Ferron, J. M., Hogarty, K. Y., Dedrick, R. F., Hess, M. R., Niles, J. D., & Kromrey, J. D. (2008). Reporting results from multilevel analyses. In A. A. O’Connell & D. B. McCoach (Eds.), Multilevel modeling of educational data (pp. 391–426). Charlotte: Information Age Publishing Inc.Google Scholar
- Hedges, L. V., & Hedberg, E. C. (2007). Intraclass correlation values for planning group-randomized trials in education. Educational Evaluation and Policy Analysis, 29(1), 60–87.View ArticleGoogle Scholar
- Hox, J. J. (2010). Multilevel analysis: Techniques and applications. Great Britain: Routledge.Google Scholar
- Kelley, K., & Preacher, K. J. (2012). On effect size. Psychological Methods, 17, 137–152. https://doi.org/10.1037/a0028086.View ArticleGoogle Scholar
- Kirk, R. E. (2013). Experimental design: Procedures for the behavioral sciences. Los Angeles: Sage.View ArticleGoogle Scholar
- Maas, C. J. M., & Hox, J. J. (2005). Sufficient sample sizes for multilevel modelling. Methodology: European Journal of Research Methods for the Behavioral and Social Sciences, 1(3), 86–92. https://doi.org/10.1027/1614-18126.96.36.199.View ArticleGoogle Scholar
- Martin, M. O., & Mullis, I. V. S. (Eds.). (2012). Methods and procedures in TIMSS and PIRLS 2011. Chestnut Hill: TIMSS & PIRLS International Study Center, Boston College.Google Scholar
- McCoach, D. B., & Black, A. C. (2008). Evaluation of model fit and adequacy. In A. A. O’Connell & D. B. McCoach (Eds.), Multilevel modeling of educational data (pp. 245–272). Charlotte: Information Age Publishing Inc.Google Scholar
- Meinck, S. & Vandenplas, C. (2012). Sample size requirements in HLM: An empirical study. IERI monograph series issues and methodologies in large-scale assessments. IER institute, special issue 1, educational testing service and international association for the evaluation of educational achievement.Google Scholar
- Peng, C. Y. J., & Chen, L. T. (2013). Beyond Cohen’s d: Alternative effect size measures for between-subject designs. Journal of Experimental Education, 82, 22–50. https://doi.org/10.1080/00220973.2012.745471.View ArticleGoogle Scholar
- R Core Team. (2014). R: A language and environment for statistical computing (Computer software), R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.org/. Accessed 29 Jan 2018.
- Raftery, A. E. (1995). Bayesian model selection in social research. Sociological Methodology, 25, 111–163.View ArticleGoogle Scholar
- Rogers, A. M., & Stoeckel, J. J. (2008). NAEP 2008 arts: Music and visual arts restricted-use data files data companion (NCES 2011–470). Washington, D.C: US Department of Education Institute of Education Sciences, National Center for Education Statistics.Google Scholar
- Snijders, T. A. B., & Bosker, R. J. (2012). Multilevel analysis: An introduction to basic and advanced multilevel modeling. Thousand Oaks: Sage Publishing.Google Scholar
- Spybrook, J. (2008). Power, sample size, and design. In A. A. O’Connell & D. B. McCoach (Eds.), Multilevel modeling of educational data (pp. 273–311). Charlotte: Information Age Publishing Inc.Google Scholar
- Tabachnick, B. G., & Fidell, L. S. (2007). Using multivariate statistics. Boston: Pearson.Google Scholar
- Wu, M. (2005). The role of plausible values in large-scale surveys. Studies in Educational Evaluation, 31, 114–128.View ArticleGoogle Scholar