The dimensionality of a reading comprehension assessment with non-stem equivalent multiple-choice (MC) items and open-ended (OE) items was analyzed with German test data of 8523
9th-graders. We found that a two-dimensional IRT model with within-item multidimensionality, where MC and OE items load on a general latent dimension and OE items additionally load on a nested latent dimension, had a superior fit compared to an unidimensional model (p = .05). Correlations between general cognitive abilities, orthography and vocabulary and the general latent dimension were significantly higher than with the nested latent dimension (p = .05). Drawing back on experimental studies on the effect of item format on reading processes, we suppose that the general latent dimension measures abilities necessary to master basic reading processes and the nested latent dimension captures abilities necessary to master higher reading processes. Including gender, language spoken at home, and school track as predictors in latent regression models showed that the well known advantage of girls and mother-tongue students is found only for the nested latent dimension.