Five independent observers assessed more than 1600 stems of Pinus radiata
D.Don for crookedness on a 0-9 scale. Inconsistencies in the scoring of individual observers resulted in erratic changes in the mean and the variance as the work progressed; this was probably the origin of four distinct interactions, which contributed a small but statistically significant part of the total variance. The error variance of the individual observer was about 0.5 and constituted about 32% of the total variance.
The frequency distributions of the errors generally showed significant departures from the normal, but for no consistent reason. In some there was skewness, in some positive kurtosis, and most showed apparently anomalous frequencies in some classes. As expected, the mean score of five observers per tree was greatly superior in its statistical properties to the single scores, the departure of its frequency distribution from the normal falling well short of the 5% significance level.
The variance of the errors showed significant heterogeneity and were to some extent correlated with the means. Attempts to eliminate these undesirable features by transformations were only partly successful; but, despite rather severe changes brought about by the transformation, the analyses of variance made before and after transformation gave essentially the same results.
The data were analysed as five separate sets (one from each observer) and as a single set combining the five scores for each tree. The results consistently indicated the presence of a substantial families component in the total variance (P < 0.001). The best estimate of heritability was 0.44 (90% confidence limits about 0.28 and 0.87).
In ranking the means of 26 open-pollinated families, the five observers showed good agreement (Kendall and Babington Smith's coefficient of concordance W = 0.81; P<0.001). There were also differences in the amount of tree-to-tree variation within families. When the families were ranked according to their variances, the observers were again found to have displayed a highly significant concordance (W = 0.55; P < 0.001).
It is concluded that a single observer, under conditions like those experienced in this study, could score crookedness accurately enough if the purpose were solely the ranking of group means; but in general the assessment of crookedness should be based on the sum of scores by two or more independent observers per tree. This is particularly true when the data are to be used for statistical work involving measures of dispersal, as in the analysis of variance and covariance.