Comparando formas de análise para dados censurados por razões práticas em programas de melhoramento vegetal

Abstract

Experiments in several areas are subject to occurrences of censored data for practical reasons, such as detection thresholds given by devices beyond which no values recorded. In these cases one can assume that a continuous latent random variable contains these thresholds. Thus allows for investigating the consequences informed truncation of data to fit distributions associated with experimental populations. Present work, inspired by a sweet potato genetic improvement experiment, aims to develop an analysis method for data with left censoring by implementing an algorithm for its conditional prediction using Gibbs sampling and verifying its properties. We compared the analysis on a simulated example with similar properties. Simulated experiment was an incomplete block design partially balanced (PBIB) (square lattice with v = k 2 ,r = 3, k = 11,b = 33,λ1 = 1 e λ2 = 0). The methods carried out were: the uncensored analysis of the complete "DC"data, the usual Zero Censorship ("C0" considering zero for the censorships) and Left Censorship ("CE" considering missing observations in the censorship) and the proposed analysis of Conditional Prediction ("PC" with conditional imputation of censored data). Competing analysis with censored data were compared to references in two scenarios: 1) moderate or high censoring (∼ 30% ∼ 50%). We evaluated selective accuracy, precision and bias in the estimates of genetic parameters (variance components and heritabilities). We also obtained correlations of simulated and predicted the censored observations . Finally, in each analysis, the Pearson and Spearman correlations between predicted genetic values and respective parametic values were calculad. "PC" analysis was sensible and accurate for selection purposes, showing correlations between treatment effects and parametric values close to the uncensored case. The proposed method has the simulated values very likely in the respective marginal a posteriori distributions. The usual forms of analysis ("C0" and "CE") have zero correlation between the values taken as zero and the parametric values. The "CE" analysis was bad in both scenarios regarding the estimation of genetic parameters (especially variances and heritabilities) and for presenting low correlation, but for the selection of the elite genotypes it was better than the "C0" analysis in the scenario two). For scenario 1) the "C0" analysis seems to be an promising alternative, but has shown a considerable worsening with increasing censorship. Although the "PC" analysis produced statements that were more difficult to interpret because it overestimated heritabilities, it was the most indicated to make selection decisions.