Evaluation and extensions of a structural equation modeling approach to the analysis of survival data.

摘要

Recently a new method for the analysis of survival data using a structural equation modeling approach has been suggested by Pickles and colleagues using twin data they demonstrated the application of this model to study the correlation in age of onset. The purpose of the current research is twofold: 1) to evaluate the statistical performance of the model as presented by Pickles and colleagues, and 2) to expand and evaluate the model in more applications, including both genetically informative data and other multivariate examples. Results evaluated from this study involve three areas of method performance: Type-I error rates, power, and parameter estimates under four different distributions (normal, Gamma-2, Gamma-6 and g-and-h) and four different sample sizes (n = 125, 250, 500 and 750). Results based on the original Pickles model indicated that in all sample size and distribution conditions the Type-I error rate was adequate, in fact below the nominal level of .05. Additionally, power was greater than .80 for sample sizes of 500 or more for all distribution conditions. Parameter estimates were upwardly biased when the population value was rho = .20. This bias varied across distributions; the g-and-h distribution showed the largest bias. Results from the expanded model indicated that Type-I error rates were adequate. Power results were not affected by distribution type; sample sizes of 500 were above the .80 level. Parameter estimates continued to be upwardly biased in this more general model, although the degree of bias was smaller.