Page 290 - Contributed Paper Session (CPS) - Volume 7

CPS2102 Iris Reinhard
3. Result
Results I – Type I Error
Figure 1 illustrates effects of varying sample sizes and scenarios on the
type I error rates (in accordance with more results not shown due to limited
number of pages). The performance improves with increasing sample size,
N=100 seems to be sufficient, while N=50 cannot be recommended. For the
impact of the covariate correlation one can find a better performance for
uncorrelated covariates.
For the correlation of random effects no obvious effect was found, the level
properties seem to be slightly better when uncorrelated. The simulations
indicate that the type I error level is kept to a great extent. For the linear mixed
model, moderate to strong differences are found, depending on the
parameter configuration, partly worse in the normal component (because of
“modelling”), see Figure 2. Especially when the effect is located only in the zero
component, the type I error for the corresponding effect in the normal
component which is the only one in the linear mixed model, is inflated,
because it might comprise the fixed effect on the probability for the
occurrence of a non-zero outcome. This also holds for large sample sizes.
Results II – Mean Squared Error (MSE)
The simulation results which are summarized in parts in Table 1
demonstrate that the empirical biases for the two-part model are negligible.
With increasing sample size the performance regarding the mean squared
error improves. The correlation between the covariates does not show an
obvious effect. The correlation of the random effects of the two components
does not indicate an impact on the MSE (not shown). The evaluation of the
two-part model showed a much better performance than the classical linear
model, the estimates of which are heavily biased. This also applies for large
sample sizes.

277 | I S I W S C 2 0 1 9