Assessing and accounting for between-sample variation of psychometric measurement models

目前患者报告结果的心理测量模型通常假设模型参数不是随机的，而是固定的。在pro分析中，当这些模型用于独立样本（如差异项目功能、因子分析或基于构建的量表的开发）时，经常观察到模型参数通常因样本而异。由于这些差异通常在统计上是显著的，但实际上是不相关的，因此为了改进PRO测量和相关推断，将它们明确地建模为随机效应是有希望的。该项目的目的是通过56项不同的研究调查这一现象，共有17.357名参与者，收集了PHQ-9的个人参与者数据荟萃分析。将评估模型参数的样本间方差，并使用贝叶斯模型将这些信息纳入分析。更具体地说，该项目的目的是(1)调查验证性因子分析和项目反应理论中模型参数估计的样本间变异性；(2)确定研究水平上的变量，如样本量、原产国、疾病群体、年龄和性别，这些变量与模型参数的系统性和相关性变化有关。(3)利用贝叶斯项目反应理论模型估计抑郁症的严重程度，该模型包含了样本间随机参数变化的信息；(4)通过比较该模型估计的抑郁症严重程度的诊断准确性与常用的PHQ-9和得分，研究该模型对预测效度的影响。本项目将深入了解PHQ-9这一常用且高度相关的心理测量模型的系统性和非系统性样本间变异的本质。此外，它将提供在pro的统计分析中解释微小的、实际上不相关的模型差异的方法，从而为未来测量模型的开发提供信息，例如，在开发与工具无关的、基于结构的pro量表时。

Current psychometric models of patient-reported outcomes (PROs) typically assume that model parameters are not random but fixed. In analyses of PROs, where such models are used in independent samples such as differential item functioning, factor analysis or the development of construct-based scales, it has been frequently observed that model parameters usually vary across samples. As these differences are often statistically significant, but practically irrelevant, it is promising to explicitly model them as random effects in order to improve PRO measurement and related inference.The objective of this project is to investigate this phenomena across 56 distinct studies with 17.357 participants, collected within an individual participant data meta-analysis of the PHQ-9. Between-sample variance of model parameters will be assessed and Bayesian models will be used to incorporate such information in analysis. More specifically, the aims of the project are (1) to investigate between-samples variability of model parameter estimates both in confirmatory factor analyis and item-response theory, (2) to identify variables on study level such as sample size, country of origin, disease groups, age and gender, which are associated with systematic and relevant variation of model parameters, (3) to estimate depression severity within a Bayesian item-response theory model which incorporates information about random parameter variation across samples and to (4) to investigate impact of such a model on predictive validity by comparing the diagnostic accuracy of estimates of depression severity derived by these models with PHQ-9 sum scores commonly used.This project will provide insights into the nature of systematic and unsystematic between-sample variation of psychometric models for a frequently used and highly relevant PRO, the PHQ-9. Furthermore, it will provide means to account for minor, practically irrelevant model differences in statistical analysis of PROs and therefore inform development of future measurement models, e.g. in the development of instrument-independent, construct-based scales for PROs.