### Table 1. Performances of the Bayesian DTs with restarting strategy

"... In PAGE 5: ... However, practically this effect disappears for domain problems including more than 300 data points. Table1 lists the characteristics of the 7 domain problems used in our experiments; here C, m, train, and test are the numbers of classes, input variables, training and testing examples, respectively. This table also provides the performances of the Bayesian DT technique on these data.... In PAGE 6: ...3% confidently incorrect classifications. In fact, as Table1 shows, the Bayesian DTs seldom make confident, but incorrect classifications, though they make more uncertain classifications. Although it may be unrealistic to expect the confidently incorrect rate to approach the Bayes error rate with small datasets, these results suggest that the randomised ensemble tends to pro- duce over-confident ensembles, while the Bayesian ensembles make few confident but incorrect classifications.... ..."

Cited by 1

### Table 2: Results on Senseval-2 English lexical sam- ple using different Bayesian network approaches.

"... In PAGE 5: ... As the dataset is used extensively for this purpose, only the Senseval-3 lex- ical sample task is used for evaluation. Selecting Bayesian Network The best achievable result, using the three different Bayesian network approaches, when validating on Senseval-2 test data is shown in Table2 . The parameters that are used... ..."

### Table 2: Results on Senseval-2 English lexical sam- ple using different Bayesian network approaches.

"... In PAGE 5: ... As the dataset is used extensively for this purpose, only the Senseval-3 lex- ical sample task is used for evaluation. Selecting Bayesian Network The best achievable result, using the three different Bayesian network approaches, when validating on Senseval-2 test data is shown in Table2 . The parameters that are used... ..."

### Table 2: A comparison of rst-order MQL, second-order PQL and Bayesian tting (with a di use prior) in model (2) applied to the Rodr guez-Goldman simulated Guatemalan child health data set number 1. Figures in square brackets in the upper table are true parameter values; gures in parentheses in the upper table are SEs (for the ML methods) or posterior SDs (for the Bayesian method). Bayesian point estimates are posterior means, and 95% central posterior intervals are reported.

in for

"... In PAGE 4: ...he main (approximate) likelihood alternatives (e.g., Goldstein 1995) currently employed with greatest frequency by multilevel modelers in substantive elds of inquiry (based upon empirical usage in the recent literature) are marginal quasi-likelihood (MQL) and penalized (or predictive) quasi-likelihood (PQL), in both of which the investigator has to specify the order of the Taylor-series approximation, and a variety of prior distributions may be considered in the Bayesian approach. Table2 summarizes a comparison between rst-order MQL, second-order PQL, and Bayesian tting|again with a particular di use prior to be discussed in Section 2.3.... In PAGE 27: ... We are not aware of large- scale simulation results on the calibration of these approaches in small samples; the literature seems particularly silent on the quality of interval estimates produced by these methods. One important likelihood-Bayesian comparison we have not addressed is computational speed, where ML/REML and MQL/PQL approaches have a distinct advantage (for example, PQL2 tting of model (2) to the Rodr guez-Goldman data set in Table2 takes less than 3 seconds on a 3GHz PC versus 1.8 minutes using MCMC with 25,000 monitoring iterations).... ..."

### Table 2: Results from the application of two Bayesian approaches and a MYCIN-based

1997

"... In PAGE 25: ... Postprocessing drastically reduces the number of incorrectly recognized motifs and consequently increases the quality of the recognition procedure #28Rost, Casadia, amp; Farisellis, 1996#29. Table2 presents a comparison of results of applying the three methods for identifying secondary structure motifs to the experimental electron density map of penicillopepsin.... In PAGE 28: ...According to Table2 , the MYCIN-based approach appears to have greater success in recognizing helical motifs in experimental maps. Example #286#29 in Figure 12 depicts one of the three helix motifs that was correctly recognized using the MYCIN-based approach.... In PAGE 28: ... Jumps are problematic and may seriously hinder the recognition rate, especially in experimental maps blurred by noise and errors. Table2 illustrates that the consideration of noisier data in the training set #28module Bayes 2 #29 leads to an improvement in the number of identi#0Ced #0B-helices with respect to the #0Crst Bayesian module #28Example #286#29 in Figure 12#29. However, this also leads to a number of incorrectly identi#0Ced segments.... ..."

Cited by 6

### TABLE III SUMMARY OF OBTAINED RESULTS BY MEANS OF ALTERNATIVE APPROACH

### Table 18: Study Two Results Comparison Study Two Alternative Approach

2000

"... In PAGE 5: ...able 17: Study Two Results ............................................................................................................................ 14 Table18 : Study Two Results Comparison.... ..."

### Table 19: Study Three Results Comparison Study Three Alternative Approach

2000

"... In PAGE 5: ...able 18: Study Two Results Comparison........................................................................................................ 14 Table19 : Study Three Results Comparison.... ..."

### Table 4: Effective Income Tax Rates Under Alternative Approaches

"... In PAGE 18: ... Overall and Sectoral Income Tax Burden Using both income and expenditure measures, households are first assigned to deciles of income or spending distribution. The effective rate of income tax -the ratio of income (expenditure) taken by income tax- is then calculated for each income (expenditure) decile (see Table4 ). Whether this ratio rises or falls (or is constant) as income (expenditure) rises determines whether the tax system is progressive or regressive (or proportionate).... In PAGE 18: ... For instance, the rich (top income decile) pay more than four times higher effective income tax rate than the poor. Table4 also shows the fraction of household per capita exenditure devoted to income tax for households grouped by total expenditures quot;. Similar pattern of tax burden emerges.... In PAGE 19: ... Furthermore, the exclusion of in-kind income/expenditure eliminates much of the share income (expenditure) of the poor from taxation, increasing the progressivity. Second, although the effective income tax rates based on expenditure are higher than those based on income, the figure shows that the variation in expenditure shares across deciles is the same as the variation in income tax outlays as a share of income 1 The urban and rural effective income tax rates for each income (expenditure) decile are also given in Table4 . Urban households pay 5.... ..."

### Table 3: Two random intercepts: Maximum Simulated Likelihood

2006

"... In PAGE 12: ... This indicates that with a single term there is no advantage of using MSL over the Bayesian approach. [ Table3 and 4 about here] In the following the complexity of the estimation increases by allowing the unobserved heterogeneity to difier between the alternatives. Here the advantage of computational time of Halton based simulation over Gauss Hermite quadrature becomes evident.... ..."