"... In PAGE 30: ...ncome. Both the elasticity and coefficient estimates show this pattern. The nonparametric elasticity estimate is smaller than the parametric one for the wage rate and larger for nonlabor income. Also, for the nonparametric estimates in the first column of Table6 , the coefficient of w3 is smaller than is the wage coefficient for the parametric estimate in equation (14). As previously noted, the coefficient of w3 gives the wage effect for a linear budget set, because dw is identically zero in that case.... In PAGE 33: ...assuming homoskedasticity leads to a simple Hausman test of the distributional assumption. Comparing the coefficient of w3 in the first column of Table6 with the coefficient of w in the first column of Table 7 gives a Hausman statistic 6.53, that should be a realization of a standard normal distribution.... ..."

"... In PAGE 14: ... We follow the suggested procedure of the last section and use only the CAF P E1 and the CAF P E2a criteria and for reasons of comparison, the linear Schwarz criterion ARSC. Table1 summarizes the results for the lynx data. Except for the CAF P E1 criterion all criteria include lag 1 and 2 in their selection.... In PAGE 14: ... Recalling the results of the previous section, these lags for the CAF P E2a may be due to over tting. To decide whether the more parsimonious model is su cient, we investigated the residuals of all suggested models using the bandwidths of Table1 and conclude that lags 1 and 2 are su cient. A plot of the estimated regression function on a relevant grid is shown in Figure 5.... In PAGE 15: ...and 3 using AF P E1 while Yao and Tong (1994) found lags 1, 3 and 6 using cross-validation. Insert Table1 about here Applying our methods to daily exchange rate data poses a di erent challenge. While there are plenty of data (3212 observations), this bene t is compromised as the data is known to be highly dependent (although only weakly correlated) and therefore asymptotics kick in very slowly.... ..."

"... In PAGE 19: ... (The exceptions were percent relative design biases, in rare cases, of up to 12% for the model-based nonparametric procedures.) Table1 shows the ratios of MSE apos;s for the various estimators to the MSE for the lo- cal polynomial regression estimator with q = 1 (LPR1). Generally, both parametric and nonparametric regression estimators perform better than HT, regardless of whether the un- derlying model is correctly speci ed or not, but that e ect decreases as the model variance increases.... In PAGE 21: ...20 to the classical regression estimator, and the MSE apos;s converge. Clearly, the bandwidth has an e ect on the MSE of LPR1, but Table1 suggests that large gains in e ciency over other estimators can be gained for a variety of bandwidth choices. In particular, for either of the bandwidths considered here LPR1 essentially dominates HT for all populations and essentially dominates REG for all populations except linear, where it is competitive.... ..."

"... In PAGE 3: ... PERFORMANCE TRIALS The interference induced by the NACCH and the resulting data rate on the NACCH was investigated by dynamic dis- crete-event simulations of a complex radio network model [5]. Table4 shows simple assumptions on the packet sizes of some NACCH signalling procedures that were used. Notice that the broadcast channel is modelled to be perma- nently active in downlink with the same constant power as a normal traffic channel.... In PAGE 4: ... Below this threshold, one channel is sufficient in conjunction with a congestion res- olution mechanism. As a reference, simulations were done with signalling phases 10 times as long as specified in Table4 (e.... ..."

"... In PAGE 3: ... PERFORMANCE TRIALS The interference induced by the NACCH and the resulting data rate on the NACCH was investigated by dynamic dis- crete-event simulations of a complex radio network model [5]. Table4 shows simple assumptions on the packet sizes of some NACCH signalling procedures that were used. Notice that the broadcast channel is modelled to be perma- nently active in downlink with the same constant power as a normal traffic channel.... In PAGE 4: ... Below this threshold, one channel is sufficient in conjunction with a congestion res- olution mechanism. As a reference, simulations were done with signalling phases 10 times as long as specified in Table4 (e.... ..."

"... In PAGE 7: ...Table4 depicts the steps that need to be followed in the procedure for modelling survival and migration. It should be noted, however, that the example depicted in Table 4 is simplified in order to illustrate the process.... ..."

"... In PAGE 9: ... For this simple example, many sets of different noisy training data were generated, and the modeling results were consistent and similar to the results shown below, which were typical. It is informative to examine the selection process of the OLS algorithm, listed in Table1 . Notice that the normalized MSE continuously decreased as more terms were added.... In PAGE 10: ... This produced a 15-term model. The model weights had very large value, as can be seen in Table1 . This was a typical sign of over-fitting.... ..."

"... In PAGE 28: ... Overall, it can be concluded that the data provide some preliminary empirical evidence for the validity of hypotheses H2. Finally, Table8 shows two further regressions. In column 8A the likelihood of an opposition is modeled using both procedural indicators and text indicators at the same time.... In PAGE 29: ...Insert Table8 about here Surprisingly, in this joint model of procedural and text indicators, only the number of application claims turns out to have a significant coefficient among all the text indicators. Again, the explanation of this result may only be preliminary and was not necessarily to be expected according to Table 3.... ..."

"... In PAGE 11: ...Table8 includes the translation of a simple TNSDL procedure. 4.... ..."