Math versus meaning in MAIHDA: A commentary on multilevel statistical models for quantitative intersectionality.

RATIONALE: Intersectionality has been increasingly adopted as a theoretical framework within quantitative research, raising questions about the congruence between theory and statistical methodology. Which methods best map onto intersectionality theory, with regard to their assumptions and the results they produce? Which methods are best positioned to provide information on health inequalities and direction for their remediation? One method, multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA), has been argued to provide statistical efficiency for high-dimensional intersectional analysis along with valid intersection-specific predictions and tests of interactions. However, the method has not been thoroughly tested in scenarios where ground truth is known. METHOD: We perform a simulation analysis using plausible data generating scenarios where intersectional effects are present. We apply variants of MAIHDA and ordinary least squares regression to each, and we observe how the effects are reflected in the estimates that the methods produce. RESULTS: The first-order fixed effects estimated by MAIHDA can be interpreted neither as effects on mean outcome when interacting variables are set to zero (as in a correctly-specified linear regression model), nor as effects on mean outcome averaged over the individuals in the population (as in a misspecified linear regression model), but rather as effects on mean outcome averaged over an artificial population where all intersections are of equal size. Furthermore, the values of the random effects do not reflect advantage or disadvantage of different intersectional groups. CONCLUSIONS: Because first-order fixed effects estimates are the reference point for interpreting random effects as intersectional effects in MAIHDA analyses, the random effects alone do not provide meaningful estimates of intersectional advantage or disadvantage. Rather, the fixed and random parts of the model must be combined for their estimates to be meaningful. We therefore advise caution when interpreting the results of MAIHDA in quantitative intersectional analyses.

Estimating New Quantities from Longitudinal Test Scores to Improve Forecasts of Future Performance.

Psychometric models for longitudinal test scores typically estimate quantities associated with single-administration tests, like ability at each time-point. However, models for longitudinal tests have not considered opportunities to estimate new quantities that are unavailable from single-administration tests. Specifically, we discuss dynamic measurement models - which combine aspects of longitudinal IRT, nonlinear growth models, and dynamic assessment - to directly estimate capacity, defined as the expected future score once the construct has fully developed. After discussing the history and connecting these areas into a single framework, we apply the model to verbal test scores from the Intergenerational Studies, which follow 494 people from 3 to 72 years old. The goal is to predict adult verbal scores (Age ≥ 34) from adolescent scores (Age ≤ 20). We held-out the adult data for prediction and compared predictions from traditional longitudinal IRT ability scores and proposed dynamic measurement capacity scores from models fit to the adolescent data. Results showed that the R2 from capacity scores were 2.5 times larger than the R2 from longitudinal IRT ability scores (43% vs. 16%), providing some evidence that exploring new quantities available from longitudinal testing could be worthwhile when an interest in testing is forecasting future performance.

Variability in Palliative Care Use after Intracerebral Hemorrhage at US Hospitals: A Multilevel Analysis.

BACKGROUND: Palliative care (PC) is an essential component of comprehensive care of patients with intracerebral hemorrhage (ICH). In the present study, we sought to characterize the variability of PC use after ICH among US hospitals. METHODS: ICH admissions from hospitals with at least 12 annual ICH cases were identified in the Nationwide Inpatient Sample between 2008 and 2011. We used multilevel logistic regression modeling to estimate between-hospital variance in PC use. We calculated the intraclass correlation coefficient (ICC), proportional variance change, and median OR after accounting for individual-level and hospital-level covariates. RESULTS: Among 26,791 ICH admissions, 12.5% received PC (95% CI 11.5-13.5). Among the 629 included hospitals, the median rate of PC use was 9.1 (interquartile range 1.5-19.3) per 100 ICH admissions, and 150 (23.9%) hospitals had no recorded PC use. The ICC of the random intercept (null) model was 0.274, suggesting that 27.4% of the overall variability in PC use was due to between-hospital variability. Adding hospital-level covariates to the model accounted for 25.8% of the between-hospital variance observed in the null model, with 74.2% of between-hospital variance remaining unexplained. The median OR of the fully adjusted model was 2.62 (95% CI 2.41-2.89), indicating that a patient moving from 1 hospital to another with a higher intrinsic propensity of PC use had a 2.63-fold median increase in the odds of receiving PC, independent of patient and hospital factors. CONCLUSIONS: Substantial variation in PC use after ICH exists among US hospitals. A substantial proportion of this between-hospital variability remains unexplained even after accounting for patient and hospital characteristics.