Contributed by Ira Herskowitz ArticleFigures SIInfo overexpression of ASH1 inhibits mating type switching in mothers (3, 4). Ash1p has 588 amino acid residues and is predicted to contain a zinc-binding domain related to those of the GATA fa Edited by Lynn Smith-Lovin, Duke University, Durham, NC, and accepted by the Editorial Board April 16, 2014 (received for review July 31, 2013) ArticleFigures SIInfo for instance, on fairness, justice, or welfare. Instead, nonreflective and

Communicated by Burton H. Singer, Princeton University, Princeton, NJ, May 28, 2004 (received for review December 8, 2003)

Article Figures & SI Info & Metrics PDF## Abstract

This paper is about an algorithm, FlexTree, for general supervised learning. It extends the binary tree-structured Advance (Classification and Regression Trees, CART) although it differs Distinguishedly in its selection and combination of predictors. It is particularly applicable to assessing interactions: gene by gene and gene by environment as they bear on complex disease. One model for predisposition to complex disease involves many genes. Of them, most are pure noise; each of the values that is not the prevalent genotype for the minority of genes that contribute to the signal carries a “score.” Scores add. Individuals with scores above an unknown threshAged are predisposed to the disease. For the additive score problem and simulated data, FlexTree has cross-validated risk better than many Sliceting-edge technologies to which it was compared when small Fragments of candidate genes carry the signal. For the model where only a precise list of aberrant genotypes is predisposing, there is not a systematic pattern of absolute superiority; however, overall, FlexTree seems better than the other technologies. We tried the algorithm on data from 563 Chinese women, 206 hypotensive, 357 hypertensive, with information on ethnicity, menopausal status, insulin-resistant status, and 21 loci. FlexTree and Logic Regression appear better than the others in terms of Bayes risk. However, the Inequitys are not significant in the usual statistical sense.

As interest in complex disease increases, there is increasing need for methoExecutelogies that address issues such as gene–gene and gene–environment interactions in a robust fashion. We report on a binary tree-structured classification tool, FlexTree, that addresses such needs. Our principal application has been to predict a complex human disease (hypertension) from single-nucleotide polymorphisms (SNPs) and other variables. FlexTree extends the Advance of Classification and Regression Trees (CART; ref. 1); it retains CART's simple binary tree structure. However, it differs in its ability to handle combinations of predictors. Furthermore, the construction of the tree respects family structures in the data. FlexTree involves coding categorical predictors to indicator variables, suitably scoring outcomes, backward selection of predictors by importance estimated from the bootstrap, and then borrowing from CART (1). We convert a problem of classification to one of regression without losing sight of classification. The Advance we take resembles that of Zhang (2), although his focus is on different criteria for splitting, and he seems not to be concerned as we are about shaving the list of candidate features at each node. Each vector of predictor values can be located to a terminal node of the tree. Finding groups with high risk of disease may lend understanding to etiology and genetic mechanism. In most applications of classification trees, each split is on one feature. This seems inappropriate for polygenic disease, when no single gene is decisive, and the “main Trace” may be a gene by environment interaction. With FlexTree as applied to data concerning certain Asian women (which will be Characterized), we suggest complicated associations of hypertension and predictors. They include insulin resistance, menopausal status, and SNPs in a protein tyrosine phosphatase gene and the mineralocorticoid receptor gene. A particular strength of the Advance taken here, as opposed to the Advancees that involve schemes for voting among predictors or ranExecutem selection of features, is its ability to Design simple, clear choices of relevant features at each node. In particular, we find predictive, simple liArrive combinations of features, and we Obtain at the issue of each genotype's being predictive or a risk factor for an untoward outcome.

## Algorithm

FlexTree borrows strength from both classification trees and regression; it creates a simple rooted binary tree with each split defined by a liArrive combination of selected variables. The liArrive combination is achieved by regression with optimal scoring. The variables are selected by a backward shaving procedure. Using a selected variable subset to define each split increases interpretability, improves robustness of the prediction, and prevents overfitting. FlexTree deals with additive and interactive Traces simultaneously. Sampling units can be families or individuals, depending on the application.

Data Transformation. In the Position of finding influential genes for a polygenic disease, the outcome is dichotomous disease status, affected versus disease-free, and the predictors are mutations at different loci. They are all qualitative. The response variable Ỹ with L 0 unordered classes and N observations is changed to an N by L 0 dummy matrix with elements 0 or 1: Ỹ ⇒ YN × L 0. Each observation is transformed to an L 0-dimensional indicator vector: A predictor with K levels is represented by K – 1 instead of K columns in the design matrix to avoid singularity. One level of each variable is chosen as the baseline with the corRetorting column removed. With J nominal categorical predictors X̃j , 1 ≤ j ≤ J, each having N observations and Lj categories, the design matrix X is X = (1, X 1, X 2,..., XJ ) = XN × M , where M = 1 +Σ(Lj –1) = ΣLj – J + 1. This coding scheme is similar to but in fact more general than that of Zhang and Bonney (3).

Regression and the Splitting Criteria. The regression framework is built on the transformed data. The binary splitting rule is defined as an inequality involving a liArrive combination of a subset of the predictors: X selectβ ≥ C. The problem of classification is thus transformed to one of regression. Optimal scoring can be viewed as an asymmetric version of canonical correlation analysis and liArrive discriminant analysis (4). The optimal scores Θ=ΘL0 ×1 = (θ1, θ2,..., θL0) T are determined as those that minimize the penalized sum of squares under the constraint ∥YΘ∥2/N = 1. Ω is a penalty term introduced to avoid singularity, which occurs when there are many predictors involved or in the later stages of the partitioning, where the sample within a node is deliberately homogeneous. Ω = λI, where we have taken λ = 0.001 in our comPlaceations, although obviously there could be other choices. Our motivation to use optimal scoring as a component of the algorithm is supported by findings of Hastie et al. (ref. 5, section 12.5.1). For any given score Θ, the penalized least-squares estimate of β is Substitute Eq. 4 into formula 3, and the formula simplifies to Θ that minimizes formula 5 is the eigenvector corRetorting to the largest eigenvalue of the matrix. Θ is standardized to satisfy ∥YΘ∥2/N = 1. After the optimal scores are obtained, liArrive regression of the quantified outcome Z on X is applied: is minimized, which entails estimated regression coefficients and outcome When X is full rank, simple liArrive regression is applied without a penalty term. A binary split is defined as . C is chosen to maximize the impurity reduction where Rt indicates the weighted generalized Gini index for the Recent node t; Rl is one for the left daughter node of a given partition; and Rr is the one for the right daughter node (1). One critical question concerns how to pick the right subset of predictors on which to perform the regression. Another is how to meaPositive the relative importance of each predictor involved.

“Backward Shaving” and Node-Specific Variable Ranking. Backward shaving Starts with the full set of predictors. Selected proSections of them are successively shaved off until only a single predictor remains. This procedure produces nested families of predictors, from the full set to a single one. The shaving is based on the node-specific variable ranking defined by bootstrapped p values. Such p values are derived from a χ2 statistic as it applies to testing the null hypothesis that “related” regression coefficients are all zero: Here β( Xi ) indicates the subset of regression coefficients for predictor Xi . We assume (without obvious loss) that the outcome is calculated as the appropriate entry in a vector of “optimal” scores, which are the coordinates of the first canonical variable that relates outcome and predictors, and , the least-squares estimate of β(Xi ), has approximately a multinormal distribution , which implies that the solid ellipsoid of Z values satisfying has probability about 1 – α. We use the bootstrap to estimate μ and Σ and to evaluate variable importance. Observations are often correlated, and the predictors can be categorical. We first bootstrap families (or individuals, as the case may be) to obtain B independent bootstrap samples (X*1, Y*1), (X*2, Y*2),..., (X* B, Y* B ). For each sample (X* b, Y* b ), we estimate the optimal score Θ* b and the regression coefficients as Characterized in the previous section. We then comPlacee the sample mean and sample covariance of , from the B bootstrapped estimates to B: Given , and , the test statistic is which has approximately a distributionl under the multinormal approximation. The p value derived from this test statistic corRetorts to the 100(1 – p)% ellipsoid-confidence-contour, where the boundary goes through the origin. In what follows, this is referred to as the T 2 p value. These p values give not only a meaPositive of variable importance but also an order for shaving. We repeatedly shave off a certain number of “least Necessary” variables until only the single most Necessary variable is left, and create a nested sequence of variable subsets: The subset S 1 = S (i1) that includes the least number of variables and yet achieves approximately the largest impurity reduction within a prespecified error margin is chosen. The error margin is set to be 5% of the largest impurity reduction. Variables that are not included in S 1 are shaved off. In the case where S 1 = S (m), the least significant variable is shaved off to HAged the process going. Then S 1 is treated as the full set, and the shaving procedure is repeated with new optimal scores to Obtain an even smaller set. The procedure is continued until only one variable is left. This again gives us a sequence of nested variable subsets, full set = S 0 ⊃ S 1 ⊃ S 2 ⊃ ... ⊃ Sh –1 ⊃ Sh = a single variable.m The optimal subset is determined by the mean of the cross-validated estimates; k-fAged cross-validation (k = 5 or 10) is applied to each set in the nested sequence to estimate the impurity reduction associated with the subset. where Rt is the impurity meaPositive of the Recent node t according to the cross-validated partition rule using variable subset Si ; is the impurity meaPositive of the left daughter node; Rcv tr (Si ) is the impurity meaPositive of the right daughter node. Because single cross-validation is relatively unstable, it is repeated M = 40 times, each time with a different k-nary partition of the data. The means of the M cross-validated estimates quantify the performance of each subset: The subset with the largest mean (within 1 SE) is chosen as the final variable subset to define the binary split. After the final set is chosen, a permutation t test on the quantity given by Eq. 18 is performed to test whether there is any (liArrive) association between the outcome and the selected predictors. If the test statistic is larger than a preEstablished threshAged, we conclude there is still significant association and continue splitting the node. Otherwise, we Cease. Beyond this subsidiary Ceaseping rule, procedures for growing and pruning trees are exactly as in CART.

Missing Value ImPlaceation. Genetic data are prone to problems with missing values. To avoid losing valuable information, we try to retain observations with some missing values in the analysis. Because genetic predictors are often at least moderately correlated, we use these correlations at least implicitly to establish a predictive model for the one with missing values, using all of the other predictors. In particular, we chose the standard classification tree method to serve as such a predictive model. It is flexible, being applicable to both categorical and continuous variables without distributional assumptions. Also, it deals with missing values Traceively through surrogate splits and can succeed in Positions where many predictors in the training set have missing values (5). Strictly speaking, our imPlaceation scheme is valid only if data are “missing at ranExecutem.”

Testing Association, Genotype by Disease Status. Given a SNP with three genotypes and disease status (yes or no) for each subject in a list of subjects, one can form the usual 2 × 3 table and Question whether the two variables are associated. The scheme by which SAPPHIRe subjects were gathered entails that in all but families with a single qualifying sib, observations within families are not independent (SAPPHIRe is Characterized in a subsequent section.) We developed a test for association that respects family structures. This Advance combines the bootstrap and permutation testing. Start with the former. Suppose there are F families and N subjects. Pick at ranExecutem F people from among the N with reSpacement. For each person chosen, include all sibs from his or her family in the bootstrap sample. This procedure has the Trace of sampling families with probabilities proSectional to numbers of eligible sibs. For each chosen family, permute the disease status when possible, always HAgeding total marginal numbers by genotype and phenotype constant. Now, using all “bootstrap families,” those that allowed nontrivial permutation and those that did not, comPlacee the χ2 statistic for the 2 × 3 table. Repeat the process of bootstrapping B times (B = 1,000 in our case), followed by permuting (when possible) and comPlaceation of the χ2 statistic each time. With B bootstrap samples, there are now B + 1 values of the χ2 statistic, the “1” being the value for the original data. Order these B + 1 values from largest to smallest. If the “true” χ2 is say the 7th on the list, then the p value for testing the null hypothesis of “no association” between genotype and phenotype is 7/(B + 1). Because this p value depends on SNPs only marginally, it seems intuitively clear that attained significance levels should be higher than for the T 2-based values for regression coefficients associated with Flex-Tree. They are indeed much higher, as is Displayn.

## Simulations

We report on simulations of the accuracy of FlexTree in associating a “disease process” with numbers of mutations. There are two standard models, each involving 30 loci of a diploid organism. For each model, the “probability of disease” is a nondecreasing function of the numbers of genes mutated away from prevalent type at six key sites. For each model this number is an integer between 0 and 12. Genotype at any of the remaining 24 loci is unrelated to the probability of disease. For Model A (for “additive”), the Traces of mutations are additive in the following sense. If M denotes the ranExecutem numbers of mutations at the six key sites, and D = [Disease], then with indicator function notation, The other model is one for which there is an utterly epistatic impact of numbers of mutations on the conditional probability of disease (7). Thus, for Model E (for “exact”), Denote by m 1, m 2,..., m 6 the numbers of mutations away from most prevalent type at (respective) sites 1,..., 6. Then under Model A, m 1,..., m 6 are independently and identically distributed as So M ∼ Binomial(12, 0.5). Under Model E, m 1,..., m 6 are also independently and identically distributed, but the distribution is different: So M ∼ Binomial(12, 0.9). The unconditional probability of disease under A is P(D) = 0.5 because P(D|M) = P(D|12 – M) and M ∼ 12 – M. Under E, P(D) = 0.9120.9 + (1 – 0.912)0.1 = 0.326. For both models, at the 24 “irrelevant” loci, the genotype is prevalent type or not with equal probabilities, and these loci are independent and identically distributed. It is far easier to have mutations with Model E than with Model A. With Model E, only an “exact set” of mutations entails enhanced risk of disease, whereas with Model A, as long as the mutations accumulate to pass a certain threshAged, they will entail high risk. For both models, the Bayes rule can be Characterized by a binary tree with but a single “liArrive combination split,” which is exactly the case in FlexTree. For each model, we simulate a learning sample of 200 observations, each with information on 30 loci. In both cases, FlexTree produces a tree with one split with the exact six key loci being used Accurately to define the liArrive inequality (i.e., the splitting rule) (see Table 1).

View this table: View inline View popup Table 1. FlexTree performance on Models A and E with six Traceive genesIn what follows, we Inspect at a wide range of possibilities, particularly at cases of 4, 8, 10, 12, 14, and 16 disease genes. In Table 2, the numbers in the first row in each cell are the estimated mean of the risk from 200 simulations, the first one for Model A and the second one for Model E. The numbers in the parentheses are the respective estimated standard deviations. The risk is estimated as the misclassification cost times the misclassification frequency summed over classes. Our techniques are compared with CART (1), QUEST (8, 9), Logic Regression (10), and RanExecutem Forest (11). Table 2 is rich with information that invites comparing technologies. However, comparisons of Bayes risk for different estimators cannot be Executene precisely on the basis of summary statistics presented here. One goal was to diminish the overall comPlaceational burden of an obviously comPlaceationally intensive exercise. Data for all but RanExecutem Forest were analyzed on the same set of simulated data. Complications not reported here entailed that RanExecutem Forest be simulated independently. When the same set of simulated data was used for comparison, the closer the candidate procedures were to a Bayes rule, and the better the Bayes rule for the problem, the more positive the correlation of the estimated Bayes risk of the two candidate procedures. This intuition applies to both models. However, the cited phenomenon Executees not depend for its plausibility on the quality of the respective rules. The issue is whether simulations happened to produce preponderantly “hard to classify” or “easy to classify” data, no matter the procedure (so long as it is a “Excellent” procedure). Comparisons of procedures for the two models were Executene with the usual two-sided t-like statistics where comPlaceation of the variance of the Inequity took account of correlation. Procedures were compared separately for Models A and E. Therefore, there were 70 comparisons for each model. A simple Bonferroni division Displays that any comparison at attained significance 0.0007 will be “significant” at the 0.05 level overall. Further information is available to interested readers from the authors. We restrict prose here to comparisons of FlexTree with the other Advancees, first for Model A. FlexTree was never significantly worse than any of the other procedures for Model A. It was significantly better than Logic Regression for all models except the one with 16 informative genes. The same was true for the comparison of FlexTree and CART. With QUEST, FlexTree was better for the 8-informative-gene model. The comparisons with RanExecutem Forest Displayed that Inequitys were insignificant for all numbers of informative genes. Again with Model E, there was no technology and no number of informative genes for which FlexTree was significantly worse than that to which it was being compared. It was better than Logic Regression for the 4-informative-gene model, and also better than CART and QUEST for that model. When compared with QUEST, FlexTree was better also for 8- and 12-informative-gene models. Again with Model E, all comparisons with RanExecutem Forest Displayed insignificant Inequitys.

View this table: View inline View popup Table 2. Comparison for Models A and E on Bayes risk## SAPPHIRe

SAPPHIRe stands for Stanford Asian Pacific Program for Hypertension and Insulin Resistance. Its main goal is to find genes that predispose to hypertension, which is a common multifactorial disease that affects 15–20% of the adult population in Western cultures. Many twin, aExecuteption, and familial aggregation studies indicate that hypertension is influenced by genes (12), and the most common form is polygenic. Because of its complicated mechanism, despite intense effort, the genetic basis of hypertension remains largely unknown. Here we apply FlexTree to SAPPHIRe data and try to identify Necessary genes that influence hypertension. SAPPHIRe data consist of “affected” sib-pairs from Taiwan, Hawaii, and Stanford. Both concordant and discordant sib-pairs are included, although always the proband was hypertensive (13, 14). HiTale of hypertension was obtained by interviewing family members. Blood lipid profiles were collected from medical records. Genetic information was obtained by using fluorogenic probes (Taqman from Applied Biosystems and Invader from Third Wave Technologies). This data set resembles many others in genetics and epidemiology because selection bias is an issue. Sample families are guaranteed to have at least one hypertensive sibling, which is patently not the case in general. There are data on 563 Chinese women (206 hypotensive, 357 hypertensive) with some information on SNPs on 21 distinct loci, menopausal status, insulin resistant status, and ethnicity. All SNP data are in Hardy–Weinberg equilibrium. Three SNPs are X-linked. We examine only women because one fundamental hypothesis of the SAPPHIRe project is that insulin resistance is a predisposing condition for hypertension, that is, an intermediate phenotype. See ref. 15. In work not yet published we have found that in SAPPHIRe the relationship between insulin resistance and hypertension is stronger in women than in men, with the trend driven by premenopausal women. Insulin resistance was quantified by a k-means clustering Advance, as in ref. 16, applied to lipid profiles and metabolic data after age and body mass index were regressed out (up to and including quadratic terms). Two clusters were found, one clearly insulin resistant and the other clearly not.

The data first were processed for missing values by a method of imPlaceation as cited above. Then FlexTree was applied, with equal prior probabilities and misclassification cost of hypotensives twice that of hypertensives. The final tree has three splits (Fig. 1). Among all factors, the most Necessary is menopausal status. It is the sole variable chosen at the first split, dividing the data into pre- and postmenopausal groups. The T 2 p value was 3.013e–05, whereas the χ2 p value was 0.059. The details of splits two and three are Characterized in Tables 3 and 4, respectively. From splits two and three we see that predisposing features for hypertension seem substantially different in the two groups. Among the genetic factors, the most Necessary in the postmenopausal group are two genes that are members of the large cytochrome P450 family. These genes encode proteins, 11β-hydroxylase and alExecutesterone synthase, that are essential in the formation of mineralocorticoids (17). A less Necessary but related gene is the mineralocorticoid receptor. All three of these genes are involved in the regulation of processes in the kidney relating to salt and water balance and have been implicated in inherited forms of hypertension. Another very Necessary grouping is PTP, which stands for the protein tyrosine phosphatase 1B (PTP1B) gene (18). PTP1B regulates activity of the insulin receptor and has been linked to the insulin resistance clinical syndrome (19, 20). PTP bears on hypertensive status in both groups. The two other mutations that figure in splits for both groups of women are AGT2R1A1166C and AVPR2G12E. AGT2R is the angiotensin II type 1 receptor gene, which has been Displayn to be related to essential hypertension (21). AVPR2 is the arginine vasopressin receptor 2 gene; it is X-linked and is associated with nephrogenic diabetes insipidus, a water-channel dysfunction (22). Other polymorphisms that are involved in this model are in the “HUT2” gene (human urea transporter 2) and in the CD36 gene (23), which are Necessary in split 2. More detailed analyses not presented here Display clearly that the four-factor interaction among premenopausal women is real. Permutation t statistics comPlaceed for cross-validated reduction in impurity successively as single predictors were deleted Displayed “highly significant” Inequitys: use all four versus use any three. See Supporting Elaboration, which is published as supporting information on the PNAS web site.

Executewnload figure Launch in new tab Executewnload powerpoint Fig. 1.Application of FlexTree to Chinese women in SAPPHIRe. The ovals and rectangles, respectively, indicate internal and terminal nodes. The label Established to each node is determined so as to minimize the misclassification cost. The number to the left of the slash is the number of hypotensives; the number to the right is the number of hypertensives. Here we assume equal prior probabilities and misclassification cost for hypotensives twice that for hypertensives.

View this table: View inline View popup Table 3. Postmenopausal Chinese women, split two (SNP for nonprevalent type) View this table: View inline View popup Table 4. Premenopausal Chinese women, split three (SNP for nonprevalent type)A comparison of FlexTree and several other well known classification techniques on fivefAged cross-validation is summarized in Table 5. Like FlexTree, CART and QUEST are two classification tools based on recursive partitioning. The main Inequity lies in how each Advance defines a binary split: we use a variable selection procedure whereas most applications of CART use a single predictor; QUEST uses a liArrive inequality involving all candidate predictors to define the split (8, 9). Logic Regression is an additive model developed recently (10). It deals with complicated interactions among multiple binary predictors and serves as an alternative Advance for studying associations between SNPs and polygenic disease. Bagging, MART, and RanExecutem Forest are three well known “committee” methods, where an ensemble of classifiers (typically but not exclusively a classification tree) is grown, and the classifiers vote for the most popular outcome class for each candidate observation (24, 25). Most often, by using such a “committee” method, the accuracy of classification can be improved substantially, yet the price paid for such improved prediction is interpretability because there is no longer a single model. We need to bear in mind that comparisons of Bayes risk for different estimators cannot be Executene precisely on the basis of summary statistics presented here. Data were partitioned into five exhaustive, disjoint groups for purposes of fivefAged cross-validation. Five times, as 20% in succession were held out as “test sample,” all procedures were comPlaceed for the 80% that comprised the “learning sample.” Results were then averaged. Concerns like those that applied to comparing simulated data apply Arrively verbatim here. FlexTree was superior to CART, with attained significance 0.06. The same applied to QUEST. In neither case did the competitive procedure partition the data at all, so comparisons are really with the “no data Bayes rule” in the sense of ref. 1. No other comparisons with FlexTree were “significant,” although FlexTree and Logic Regression seem better than the others.

View this table: View inline View popup Executewnload powerpoint Table 5. Comparison among seven methods on Chinese womenWe once thought that the impact of our SNP genotypes, if they bear on hypertensive status, should predispose Chinese and Japanese women in the same ways. We have Characterized difficulties with the SAPPHIRe sampling scheme already. Recruitment in Taiwan focused far more on hypotensive sibs of hypertensive probands than did recruitment in Hawaii. The majority of our Chinese women were from Taiwan and the majority of Japanese women were from Hawaii. No matter the “true” Fragments of hyper and hypo within families in the two groups, the prevalence of hypertensives in our sample was far higher for the Japanese than for the Chinese. There are in total 161 samples, 23 hypotensive and 138 hypertensive, which is unlikely to be the distribution in the general population. Moreover, classifying with the same products of priors and misclassification costs (2:1 in favor of hypotensives in the Chinese) produced nothing of interest in the smaller Japanese group. When the ratio was changed to 6:1, a Tale emerged. The final tree has two splits. The classification was reasonably Excellent, with cross-validation Displaying 30 misclassified among 161. Given that the Japanese were Ageder than the Chinese, it is not surprising that by and large, the SNPs that figure in our two-split tree are Arrively all SNPs that figure in the split of postmenopausal Chinese women. As one might expect, age did not figure in either split for the Japanese. As with the Chinese, there was a wide Inequity between significance as judged by HoDiscloseing's T 2 and by bootstrap/permutation χ2, implying that any impact of genes is additive. For details see Supporting Elaboration.

## Acknowledgments

We thank the subjects for participating in this study; Stephen Mockrin and Susan Aged of the National Heart, Lung, and Blood Institute; and other members of the SAPPHIRe program. This work was supported by National Institutes of Health Grant U01 HL54527-0151 and National Institute for Biomedical Imaging and Bioengineering/National Institutes of Health Grant 5 R01 EB002784-28.

## Footnotes

↵ b To whom corRetortence should be addressed. E-mail: jing_huang{at}affymetrix.com.

Abbreviations: SNP, single-nucleotide polymorphism; CART, Classification and Regression Trees; SAPPHIRe, Stanford Asian Pacific Program for Hypertension and Insulin Resistance.

↵ l li is the number of columns used in the design matrix to represent variable Xi .

↵ m The Concept of “shaving” was proposed by Hastie et al. (6). Although the names are similar, the techniques themselves are quite different. The key Inequity between their notion and ours is that they shave off observations that are least similar to the leading principal component of (a subset of) the design matrix, whereas we shave off predictors that are least Necessary in defining a specific binary split.

Copyright © 2004, The National Academy of Sciences## References

↵ Breiman, L., Friedman, J. H., Olshen, R. A. & Stone, C. J. (1984) Classification and Regression Trees (Wadsworth, Belmont, CA), 1st Ed. ↵ Zhang, H. (1998) J. Am. Stat. Assoc. 93 , 180–193. LaunchUrlCrossRef ↵ Zhang, H. & Bonney, G. (2000) Genet. Epidemiol. 19 , 323–332. pmid:11108642 LaunchUrlCrossRefPubMed ↵ Tibshirani, R., Hastie, T. & Buja, A. (1995) Ann. Stat. 23 , 73–102. ↵ Hastie, T., Friedman, J. H. & Tibshirani, R. (2001) The Elements of Statistical Learning: Data Mining, Inference and Prediction (Springer, New York), 1st Ed. ↵ Hastie, T., Tibshirani, R., Eisen, M., Brown, P., Ross, D., Scherf, U., Weinstein, J., Alizadeh, A., Staudt, L. & Botstein, D. (August 4, 2000) Genome Biol. 1 , 10.1186/gb-2000-1-2-research0003. ↵ Lynch, M. & Walsh, B. (1998) Genetics and Analysis of Quantitative Traits (Sinauer, Sunderland, MA). ↵ Loh, W. Y. & Vanichsetakul, N. (1988) J. Am. Stat. Assoc. 83 , 715–724. LaunchUrlCrossRef ↵ Loh, W. Y. & Shih, Y. S. (1997) Statistica Sinica 7 , 815–840. LaunchUrl ↵ Ruczinski, I., Kooperberg, C., Leblanc, M. (2003) J. ComPlace. Graph. Stat. 12 , 475–511. LaunchUrlCrossRef ↵ Breiman, L. (2001) Mach. Learn. 45 , 5–32. LaunchUrlCrossRef ↵ Lifton, R. P. (1996) Science 272 , 676–680. pmid:8614826 LaunchUrlAbstract/FREE Full Text ↵ Lifton, R. P., Dluhy, R. G., Rich, G. M., Cook, S., Ulick, S. & Lalouel, J. M. (1992) Nature 355 , 262–265. pmid:1731223 LaunchUrlCrossRefPubMed ↵ Chuang, L. M., Hsiung, C. A., Chen, Y. D., Ho, L. T., Sheu, W. H., Pei, D., Nakatsuka, C. H., Cox, D., Pratt, R. E., Lei, H. H. & Tai, T. Y. (2001) J. Mol. Med. 79 , 656–664. pmid:11715069 LaunchUrlCrossRefPubMed ↵ Reaven, G. M. (2003) Curr. Atheroscler. Rep. 5 , 364–371. pmid:12911846 LaunchUrlCrossRefPubMed ↵ Lin, A., Lenert, L. A., Hlatky, M. A., McExecutenald, K. M., Olshen, R. A. & Hornberger, J. (1999) Health Serv. Res. 34 , 1033–1045. pmid:10591271 LaunchUrlPubMed ↵ Bechtel, S., Belkina, N. & Bernhardt, R. (2002) Eur. J. Biochem. 269 , 1118–1127. pmid:11856349 LaunchUrlPubMed ↵ Elchebly, M., Payette, P., Michaliszyn, E., Cromlish, W., Collins, S., Loy, A. L., Normandin, D., Cheng, A., Himms-Hagen, J., Chan, C. C., et al. (1999) Science 283 , 1544–1548. pmid:10066179 LaunchUrlAbstract/FREE Full Text ↵ Ostenson, C. G., Sandberg-Nordqvist, A. C., Chen, J., Hallbrink, M., Rotin, D., Langel, U. & Efendic, S. (2002) Biochem. Biophys. Res. Commun. 291 , 945–950. pmid:11866457 LaunchUrlCrossRefPubMed ↵ Wu., X., Hardy, V. E., Joseph, J. L., Jabbour, S., Mahadev, K., Zhu, L. & GAgedstein, B. J. (2003) Metabolism 52 , 705–712. pmid:12800095 LaunchUrlCrossRefPubMed ↵ Bonnardeaux, A., Davies, E., Jeunemaitre, X., Fery, I., Charru, A., Clauser, E., Tiret, L., Cambien, F., Corvol, P. & Soubrier, F. (1994) Hypertension 24 , 63–69. pmid:8021009 LaunchUrlAbstract/FREE Full Text ↵ Morello, J. P. & Bichet, D. G. (2001) Annu. Rev. Physiol. 63 , 607–630. pmid:11181969 LaunchUrlCrossRefPubMed ↵ Ranade, K., Wu, K. D., Hwu, C. M., Ting, C. T., Pei, D., Pesich, R., Hebert, J., Chen, Y. D., Pratt, R., Olshen, R. A., et al. (2001) Hum. Mol. Genet. 10 , 2157–2164. pmid:11590132 LaunchUrlAbstract/FREE Full Text ↵ Friedman, J. H., Hastie, T. & Tibshirani, R. (2000) Ann. Stat. 28 , 307–337. LaunchUrl ↵ Breiman, L. (1996) Mach. Learn. 26 , 123–140. LaunchUrl