Evaluating Anchoring Methods to Analyze Longitudinal Data with Item Response Models

Ordinal data is ubiquitous in psychological research, but it presents unique challenges in longitudinal analysis. Presently available longitudinal item response models (IRMs) can be computationally prohibitive for large, multiwave datasets, while lower computational alternatives may not produce useful estimates. Longitudinal anchoring is a possible solution to these issues. By anchoring separate IRMs together, person and item estimates can be obtained without limiting the number of timepoints that can be analyzed. A simulation study examining the performance of longitudinal anchoring was conducted. Six anchoring methods were evaluated: Floated, All Times, Time One, Mean, Random, and Cross-sectional. The results suggest that the Mean and the Cross-sectional anchoring methods performed the best. While the Time One, Random, and Floated methods produced similar item and person estimates, model fit was poor. The All Times method should be avoided as it cannot produce reliable change estimates. Longitudinal anchoring is an easily implemented solution when analyzing large, complex longitudinal datasets and shows promise as a low-computation method of producing latent trait estimates.