How much missing data is too much? A strawberry case study

Abstract

In the Australian Strawberry Breeding Program (ASBP), missing data can arise due to no or low numbers of fruit at season tails, or too few undamaged fruit available to accurately measure or rate a plot for a particular trait. The ASBP wanted to understand just how much missing phenotypic data is too much for the accurate estimation of genetic variances, heritability and reliability.
A simulation study was performed to investigate how the level of missingness at both the plot and genotype level affected the estimation of total, additive and non-additive genetic variances, heritability and reliability. These parameters were explored for a range of traits, including yield, fruit quality and plant architecture characteristics. The level of missingness was investigated in increments of 5 from 0% to 50%, at both the plot (6 plant plots) and genotype (replicated 1-3 times across plots) levels. A non-pedigree model and a pedigree model were both applied to each simulated dataset iteration. In this study, 250 iterations were performed for each combination of level (either plot or genotype) and missing data amount, resulting in 5,500 simulations in total for each trait. From these models total, additive, and non-additive genetic variances were estimated for each iteration to provide a measure of these variances’ stability.
Initial results suggest that, at the plot level, the estimates of genetic parameters decrease, and fairly rapidly, with higher levels of missing data compared to at the genotype-level, particularly for heritability, i.e. missing plot level data has more of an impact than missing genotype level data. For some traits, genetic variance and heritability remained somewhat stable for 0-25% missing data but then declined from 25-50%, particularly at the plot level. At the genotype level for some traits, total genetic variance remained stable until about 30% missing data and then increased slightly. These results suggest that traits with over 25% missing data at the plot level and 30% at the genotype level do not provide accurate estimation of genetic variances and should not be included in downstream analysis, including genomic prediction, as it could affect selection decisions. This knowledge will also help the ASBP conserve resources whereby traits are not measured unless an appropriate number of plots and genotypes are available to collect data.