Computing Integrated Ratings from Heterogeneous Phenotypic Assessments: A Case Study of Lettuce Postharvest Quality and Downy Mildew Resistance

Comparing performance of a large number of accessions simultaneously is not always possible. Frequently, only subsets of all accessions are tested in separate trials with only some (or none) of the accessions overlapping between subsets. Using standard statistical approaches to combine data from such a sparsely populated accession × trial matrix is precluded if different rating scales are used to evaluate accessions in those trials. Here we compare two approaches that can compute an overall linear rating for the performance of all accessions across a set of trials, even for accessions that were never tested together and were rated on dissimilar scales. We use data from lettuce ( Lactuca sativa L.) postharvest quality collected on 178 accessions in 18 trials and assessment of lettuce resistance to downy mildew ( Bremia lactucae Regel) performed on 583 accessions in 53 trials. The projected values (PV) approach uses a combination of principal component analysis and resampling to merge trial results and calculates an overall rating from real values. In contrast, the rank‐aggregation (RA) approach uses an extension of the Rasch model to combine rank‐ordered data from individual trials. We found high correlation between ratings produced by the two approaches for the postharvest quality ( r = 0.803) and the resistance to downy mildew ( r = 0.748). Combining data from multiple experiments identified lettuce accessions with a high level of resistance to the disease and a slow rate of deterioration when processed for salad. The PV and RA approaches also allow combining data from different laboratories or databases.