Skip to contents

This function performs multimodel inference to evaluate the importance of predictors with a meta-analytical meta-regression model.


multimodel.inference(TE, seTE, data, predictors, method='REML', test='knha',
    eval.criterion='AICc', interaction=FALSE, seed=123)



The precalculated effect size for each study. Must be supplied as the name of the effect size column in the dataset (in quotation marks; e.g. TE = "effectsize").


The precalculated standard error of the effect size for each study. Must be supplied as the name of the standard error column in the dataset (in quotation marks; e.g. seTE = "se").


A data.frame containing columns for the effect size, standard error and meta-regression predictors of each study/effect.


A character vector specifying the predictors to be used for multimodel inference. Names of the predictors must be identical to the names of the columns in the data.frame supplied to data.


Meta-analysis model to use for pooling effect sizes. Use 'FE' for the fixed-effect model. Different random-effect models are available: "DL", "HE", "SJ", "ML", "REML", "EB", "HS" or "GENQ. If 'FE' is used, the test argument is automatically set to 'z', as the Knapp-Hartung method is not meant to be used with fixed-effect models. Default is 'REML', and it is strongly advised to remain with this option to use a standard (mixed-effects) meta-regression model.


Method to use to compute test statistics and confidence intervals. Default is 'knha' which uses the Knapp-Hartung (Knapp & Hartung, 2003) adjustment method. "Conventional" Wald-type tests and CIs are calculated by setting this argument to 'z'. When method='FE', this argument is set to 'z' automatically as the Knapp-Hartung method was not intended to be used with fixed-effect models.


Evaluation criterion to sort the multiple models by. Can be either 'AICc' (default; corrected Akaike's Information Criterion), 'AIC' (Akaike's Information Criterion) or 'BIC' (Bayesian Information Criterion).


If set to FALSE (default), no interactions between predictors are considered. Setting this parameter to TRUE means that all interactions are modeled (interactions will only be modeled if the number of provided predictors is 4 or less).


Set a seed for the function. Default seed is 123.


Returns four tables and a plot:

  • Final Results (Summary Table): Displays the number of fitted models, model formula, method to calculate test statistics and confidence intervals, interactions, and evaluation criterion used.

  • Best 5 Models: Displays the top five models in terms of the evaluation criterion used. Predictors are displayed as columns of the table, and models as rows. A number (weight) or + sign (for categorical predictors) indicates that a predictor/interaction term was used for the model, while empty cells indicate that the predictor was omitted in this model. Other metrics such as the weight, evaluation metric delta compared to the best model, log-Likelihood (logLik) and degrees of freedom are also displayed.

  • Multimodel Inference Coefficients: Displays the estimated coefficients and statistical significance of each regression term in the model.

  • Predictor Importance: Displays the estimated importance of each model term. The table is sorted from highest to lowest. A common rule of thumb is to consider a predictor as important when its importance value is above 0.8.

  • Predictor Importance Plot: A bar plot for the predictor importance data along with a reference line for the 0.8 value often used as a crude threshold to characterize a predictor as important.


Multimodel methods differ from stepwise regression methods as they do not try to successively build the “best” single (meta-regression) model explaining most of the variance. Instead, in this procedure, all possible combinations of a predefined selection of predictors are modeled, and evaluated using a criterion such as Akaike’s Information Criterion, which rewards simpler models. This enables a full examination of all possible models, and how they perform. A common finding using this procedure is that there are many different kinds of predictor combinations within a model which lead to a good fit. In multimodel inference, the estimated coefficients of predictors can then be synthesized across all possible models to infer how important certain predictors are overall.

Multimodel Inference can be a useful way to obtain a comprehensive look on which predictors are more or less important for predicting differences in effect sizes. Despite avoiding some of the problems of stepwise regression methods, it should be noted that this method should still be rather seen as exploratory, and may be used when there is no prior knowledge on how predictors are related to effect sizes in the research field under study.

The multimodel.inference function calls the rma.uni function internally, results of which are then fed forward to an adapted version of the dredge function internally for multimodel inference. Parts of the computations in this function are adapted from a vignette by Wolfgang Viechtbauer, which can be found here.


Harrer, M., Cuijpers, P., Furukawa, T.A, & Ebert, D. D. (2019). Doing Meta-Analysis in R: A Hands-on Guide. DOI: 10.5281/zenodo.2551803. Chapter 9.1

Knapp, G., & Hartung, J. (2003). Improved tests for a random effects meta-regression with a single covariate. Statistics in Medicine, 22, 2693–2710.

Viechtbauer, W. (2019). Model Selection using the glmulti and MuMIn Packages. Link. Last accessed 01-Aug-2019.

See also


Mathias Harrer & David Daniel Ebert


if (FALSE) {
# Example 1: Perform multimodel inference with default settings
mmi = multimodel.inference(TE = 'yi', seTE = 'sei', data = MVRegressionData,
                           predictors = c('pubyear', 'quality',
                                          'reputation', 'continent'))
# Print summary

# Plot predictor importance

# Example 2: Model Interaction terms, set method to 'DL',
# change evaluation criterion to bic
multimodel.inference(TE = 'yi', seTE = 'sei', data = MVRegressionData,
                     predictors = c('pubyear', 'quality',
                                    'reputation', 'continent'),
                     method='DL', eval.criterion = 'BIC', interaction = TRUE)

# Example 3: Use only categorical predictors
multimodel.inference(TE = 'TE', seTE = 'seTE', data = ThirdWave,
                     predictors = colnames(ThirdWave)[4:7], interaction = FALSE)}