A heteroscedastic bivariate distribution arising from a model for rater agreement, and its fitting by simulation (Q1389256)

The paper deals with the problem of finding a suitable model for data in the following situation. Suppose that each person in a population of interviewers exhibits some particular level \(C\) of competency and there are two opinions about it, respectively, \(X_1\) and \(X_2\). The simplest model is \(X_1=C+E_1\), \(X_2=C+E_2\), where \(C\) is a random variable, \(E_1\) and \(E_2\) represent errors independent on \(C\). The author proposed the model \[ X_1=C+E_1\exp(a+bC),\;X_2=C+E_2\exp(a+bC) \] which is more appropriate to real data when the variability of errors is greater for large values of \(C\) than for small ones. Here \(C\), \(E_1\), \(E_2\) can be taken as mutually independent random variables with the standard normal distribution (the case \(b=0\) leads to a bivariate normal model). The author uses simulations to fit the joint distribution of \(X_1\) and \(X_2\). That is why the explicit expression for the joint distribution is not required. A numerical example of data on ability in Russian of 56 tests was analyzed with the help of the statistical package SPIDA. It shows good fitting to the data of the proposed model.