influence.ME: Tools for Detecting Influential Data in Mixed Effects Models

influence.ME provides tools for detecting influential data in mixed effects models. The application of these models has become common practice, but the development of diagnostic tools has lagged behind. influence.ME calculates standardized measures of influential data for the point estimates of generalized mixed effects models, such as DFBETAS, Cook’s distance, as well as percentile change and a test for changing levels of significance. influence.ME calculates these measures of influence while accounting for the nesting structure of the data. The package and measures of influential data are introduced, a practical example is given, and strategies for dealing with influential data are suggested. The application of mixed effects regression models has become common practice in the field of social sciences. As used in the social sciences, mixed effects regression models take into account that observations on individual respondents are nested within higherlevel groups such as schools, classrooms, states, and countries (Snijders and Bosker, 1999), and are often referred to as multilevel regression models. Despite these models’ increasing popularity, diagnostic tools to evaluate fitted models lag behind. We introduce influence.ME (Nieuwenhuis, Pelzer, and te Grotenhuis, 2012), an R-package that provides tools for detecting influential cases in mixed effects regression models estimated with lme4 (Bates and Maechler, 2010). It is commonly accepted that tests for influential data should be performed on regression models, especially when estimates are based on a relatively small number of cases. However, most existing procedures do not account for the nesting structure of the data. As a result, these existing procedures fail to detect that higher-level cases may be influential on estimates of variables measured at specifically that level. In this paper, we outline the basic rationale on detecting influential data, describe standardized measures of influence, provide a practical example of the analysis of students in 23 schools, and discuss strategies for dealing with influential cases. Testing for influential cases in mixed effects regression models is important, because influential data negatively influence the statistical fit and generalizability of the model. In social science applications of mixed models the testing for influential data is especially important, since these models are frequently based on large numbers of observations at the individual level while the number of higher level groups is relatively small. For instance, Van der Meer, te Grotenhuis, and Pelzer (2010) were unable to find any country-level comparative studies involving more than 54 countries. With such a relatively low number of countries, a single country can easily be overly influential on the parameter estimates of one or more of the country-level variables. Detecting Influential Data All cases used to estimate a regression model exert some level of influence on the regression parameters. However, if a single case has extremely high or low scores on the dependent variable relative to its expected value — given other variables in the model, one or more of the independent variables, or both — this case may overly influence the regression parameters by ’pulling’ the estimated regression line towards itself. The simple inclusion or exclusion of such a single case may then lead to substantially different regression estimates. This runs against distributional assumptions associated with regression models, and as a result limits the validity and generalizability of regression models in which influential cases are present. The analysis of residuals cannot be used for the detection of influential cases (Crawley, 2007). Cases with high residuals (defined as the difference between the observed and the predicted scores on the dependent variable) or with high standardized residuals (defined as the residual divided by the standard deviation of the residuals) are indicated as outliers. However, an influential case is not always an outlier. On the contrary: a strongly influential case dominates the regression model in such a way, that the estimated regression line lies closely to this case. Although influential cases thus have extreme values on one or more of the variables, they can be onliers rather than outliers. To account for this, the (standardized) deleted residual is defined as the difference between the observed score of a case on the dependent variable, and the predicted score from the regression model that is based on data from which that case was removed. Just as influential cases are not necessarily outliers, outliers are not necessarily influential cases. This also holds for deleted residuals. The reason for this is that the amount of influence a case exerts on the regression slope is not only determined by how well its (observed) score is fitted by the specified regression model, but also by its score(s) on the The R Journal Vol. 4/2, December 2012 ISSN 2073-4859 CONTRIBUTED RESEARCH ARTICLES 39 independent variable(s). The degree to which the scores of a case on the independent variable(s) are extreme is indicated by the leverage of this case. A higher leverage means more extreme scores on the independent variable(s), and a greater potential of overly influencing the regression outcomes. However, if a case has very extreme scores on the independent variable(s) but is fitted very well by a regression model, and if this case has a low deleted (standardized) residual, this case is not necessarily overly influencing the outcomes of the regression model. Since neither outliers, nor cases with a high leverage, are necessarily influential, a different procedure is required for detecting influential cases. The basic rationale behind measuring influential cases is based on the principle that when single cases are iteratively omitted from the data, models based on these data should not produce substantially different estimates. If the model parameters change substantially after a single case is excluded, this case may be regarded as too influential. However, how much change in the model parameters is acceptable? To standardize the assessment of how influential a single case is, several measures of influence are commonly used. First, DFBETAS is a standardized measure of the absolute difference between the estimate with a particular case included and the estimate without that particular case (Belsley, Kuh, and Welsch, 1980). Second, Cook’s distance provides an overall measurement of the change in all parameter estimates, or a selection thereof (Cook, 1977). In addition, we introduce the measure of percentile change and a test for changing levels of significance of the fixed parameters. Up to this point, this discussion on influential data was limited to how single cases can overly influence the point estimates (or BETAS) of a regression model. Single cases, however, can also bias the confidence intervals of these estimates. As indicated above, cases with high leverage can be influential because of their extreme values on the independent variables, but not necessarily are. Cases with high leverage but a low deleted residual compress standard errors, while cases with low leverage and a high deleted residual inflate standard errors. Inferences made to the population from models in which such cases are present may be incorrect. Detecting Influential Data in Mixed Effects Models Other options are available in R that help evaluating the fit of regression models, including the detection of influential data. The base R installation provides various plots for regression models, including but not limited to plots showing residuals versus the fitted scores, Cook’s distances, and the leverage versus the deleted residuals. The latter plot can be used to detect cases that affect the inferential properties of the model, as discussed above. These plots, however, are not available for mixed effects models. The LMERConvenienceFunctions package provides model criticism plots, including the density of the model residuals and the fitted values versus the standardized residuals (Tremblay, 2012). However, while this package works with the lme4 package, it only is applicable for linear mixed effects models. The influence.ME package introduced here contributes to these existing options, by providing several measures of influential data for generalized mixed effects models. The limitation is that, unfortunately, as far as we are aware, the measure of leverage was not developed for generalized mixed effects models. Consequently, the current installment of influence.ME emphasizes detecting the influence of cases on the point estimates of generalized mixed effect models. It does, however, provide a basic test for detecting whether single cases change the level of significance of an estimate, and therefore the ability to make inferences from the estimated model. To apply the logic of detecting influential data to generalized mixed effects models, one has to measure the influence of a particular higher level group on the estimates of a predictor measured at that level. The straightforward way is to delete all observations from the data that are nested within a single higher level group, then re-estimate the regression model, and finally evaluate the change in the estimated regression parameters. This procedure is then repeated for each higher-level group separately. The influence function in the influence.ME package performs this procedure automatically, and returns an object containing information on the parameter estimates excluding the influence of each higher level group separately. The returned object of class "estex" (ESTimates EXcluding the influence of a group) can then be passed on to one of the functions calculating standardized measures of influence (such as DFBETAS and Cook’s Distance, discussed in more detail in the next section). Since the procedure of the influence function entails re-estimating mixed effects models several times, this can be computationally intensive. Unlike the standard approach in R, we separated the estimation procedure from calculating the measures of influence themselves. This allows the user to process a single model once using the influence function, and then to evaluate it using various measures and plots of influence. In detecting influential data in mixed effects models, the key focus is on changes in the estimates of variables measured at the group-level. However, most mixed effects regression models estimate the effects of both lower-level and higher-level variables simultaneously. Langford and Lewis (1998) developed a procedure in which the mixed effects model is modified to neutralize the group’s influence on the higher-level estimate, while at the same time allowing the lower-level observations nested within The R Journal Vol. 4/2, December 2012 ISSN 2073-4859 40 CONTRIBUTED RESEARCH ARTICLES that group to help estimate the effects of the lowerlevel predictors in the model. For each higher-level unit evaluated based on this method, the interceptvector of the model is set to 0, and an (additional) dummy variable is added to the model, with score 1 for the respective higher level case. This way, the case under investigation does not contribute to the variance of the random intercept, nor to the effects of variables measured at the group-level. influence.ME provides this functionality, which is accessed by specifying delete=FALSE as an option to the influence function. As a result of the specific modification of the model-specification, this specific procedure suggested by Langford and Lewis (1998) does not work when factor-variables are used in the regression model. Finally, influence.ME also allows for detecting the influence of lower-level cases in the mixed effects model. In social science applications of mixed effects models, with a great number of lower-level observations nested in a limited number of groups, this will not always be feasible. Detecting influence of lower-level observations is supported for applications in various disciplines where mixed effects models are typically applied to only a limited number of observations per group. This procedure is accessed by specifying obs=TRUE as an option to the influence function. The influence function can either determine the influence of higher-level cases, or of single observations, but not both at the same time. The Outcome Measures The influence function described above returns an object with information on how much the parameter estimates in a mixed effects model change, after the (influence of) observations of higher-level groups and their individual-level observations were removed from it iteratively. This returned object can then be passed on to functions that calculate standardized measures of influence. influence.ME offers four such measures, which are detailed in this section.