I read Achim Zeileis's paper: Model-Based Recursive Partitioning for a long time. But I still confused with the splitting criterion in this paper.

My understanding is that it would evaluate the instability for every partitioning variable respectively in the beginning. If there is one partitioning variable's instability is largest (it means that splitting this variable would let model-fitting become better) and p-value is lower (it means that our estimation of instability is better) than significant level.

However, I can't interpret anything from formula in this paper. Could anyone who understands those formulas give me some tips or idea to interpret those formulas (3)~(7) in this paper.


1 Answer 1


Equations (3) and (4) in the preprint version (3.1 and 3.2 in the published JCGS journal version) pertain to the first order conditions to the maximization problem formulated in Equation (1) (or 2.1). If you obtain the parameter estimates by maximum likelihood (ML) for example, then Equation (4) is the score function corresponding to the log-likelihood function. This is sometimes also called the estimating function. The gradient of the summed objective function sums these score contributions and this needs to be zero at the optimal parameter estimates (Equation 3).

A simple example is the estimation of a constant mean (either via OLS or ML). The score function then simply corresponds to the residuals. And the sum of the residuals is zero at the estimated parameter (= empirical mean).

The empirical score contributions can thus be seen as measures of model deviation and the idea of the subsequently discussed parameter instability tests in Section 3.2 is the following: If the model fits well to the current subsample, then the scores should fluctuate randomly around zero. If there is a systematic change in parameters, then there should be a shift in the scores as well. This can be captured by looking at cumulative sums of the scores (Equation 5 or 3.3) and then brought out by suitable test statistics (Equations 6/7 or 3.4/3.5).

More details are given in the references cited in the paper. Also some of our follow-up work has more details, e.g.:

  • Carolin Strobl, Florian Wickelmaier, Achim Zeileis (2011). Accounting for Individual Differences in Bradley-Terry Models by Means of Recursive Partitioning. Journal of Educational and Behavioral Statistics, 36(2), 135–153. http://dx.doi.org/10.3102/1076998609359791

  • Edgar C. Merkle, Achim Zeileis (2013). Tests of Measurement Invariance without Subgroups: A Generalization of Classical Methods. Psychometrika, 78(1), 59–82. http://dx.doi.org/10.1007/s11336-012-9302-4

  • Ting Wang, Edgar C. Merkle, Achim Zeileis (2014). Score-Based Tests of Measurement Invariance: Use in Practice. Frontiers in Psychology, 5(438). http://dx.doi.org/10.3389/fpsyg.2014.00438


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.