An R package for Automated Generation of Item Response Theory Model NONMEM Control File
Leticia Arrington(1,2), Rikard Nordgren(1), Malidi Ahamadi(2), Sebastian Ueckert(1), Sreeraj Macha(2), Mats O. Karlsson(1)
(1) Uppsala University, Uppsala Sweden (2) Merck& Co., Inc., Kenilworth, NJ, USA,
Introduction
Pharmacometrics Item response theory (IRT) models have been recognized as a valuable modeling approach for analyzing healthcare related composite assessments; it provides a natural framework to combine different outcomes from the same disease into a joint disease model. However, the implementation of Pharmacometrics IRT presents several technical challenges, e.g. it has many components increasing the risk of coding errors, IRT models contain large number of parameters that require initial estimates, and with many contributing data types to the elaborate model diagnostic. These challenges are especially hampering in a drug development setting where analysts need to be skilled in a wide area of techniques. There is a need to develop a tool that facilitates the implementation of Pharmacometrics IRT. Existing ready to use IRT-R packages are mainly focused on psychometrics with limited flexibility for modeling longitudinal data.
Objectives: The objective of this work was to develop a modeling tool in the form of an R package, called nmIRT, that streamline the implementation of Pharmacometrics IRT. The functionality of nmIRT is showcased using Parkinson’s disease composite assessment data (i.e. MDS-UPDRS).
Methods:The nmIRT R package aims to streamline the implementation of Pharmacometrics IRT models. The overall package can be decomposed in two main components: the assembler, responsible for creating NONMEM-IRT model and the Inspector, responsible for the generation of diagnostic plots. The assembler allows the users to specify the desired structure of the IRT model and then generate the corresponding NONMEM code that can be ran in any NONMEM modeling and simulation platform. The calculation of initial values for the NONMEM code as well as the identifiability analysis will rely on existing IRT R-packages (e.g. mirt). Currently the types of data that are supported are ordered categorical and binary. Once the NONMEM-IRT model is fitted, the “Inspector” part of nmIRT will perform model diagnostic that include mirror plots, Item characteristic curves and residual correlation plots. Additional capabilities such as data checkout and the creation of a IRT model based directly from a provided data set are included in nmIRT package. The ability of the newly designed workflow to handle IRT model development was assessed using a Parkinson’s disease model and data from Gottipati et al.
Results:The model assembler was able to generate the IRT model structure including all components required for estimation and simulation for ordered categorical and binary items and provided a default structure for the latent variable after execution of a few lines of code. The item parameter estimates were comparable between the manual IRT model and the autogenerated model from the nmIRT package. The model diagnostics provided information on item parameter fit and provided supportive evidence of the functionalities within the nmIRT package.
Conclusions:
The nmIRT model assembler first generation is able to expedite the IRT model development process by generating an editable IRT NONMEM control file for MDS-UPDRS scale for use in NONMEM with input from the user. While the current package internal predefined scale is MDS-UPDRS, the current package is able to handle datasets from other scales with categorical and binary data, which allows general use for other similar assessments.
References:
[1] Gottipati G, Karlsson MO, Plan EL. Modeling a composite score in Parkinson's disease using item response theory. AAPS J. 2017;19(3):837–845
