BioMed Research International

Volume 2016, Article ID 8209453, 11 pages

http://dx.doi.org/10.1155/2016/8209453

## High Dimensional Variable Selection with Error Control

Department of Biostatistics and Bioinformatics, Duke University Medical Center, Box 2717, Durham, NC 27710, USA

Received 3 April 2016; Accepted 25 May 2016

Academic Editor: Weiwei Zhai

Copyright © 2016 Sangjin Kim and Susan Halabi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

*Background.* The iterative sure independence screening (ISIS) is a popular method in selecting important variables while maintaining most of the informative variables relevant to the outcome in high throughput data. However, it not only is computationally intensive but also may cause high false discovery rate (FDR). We propose to use the FDR as a screening method to reduce the high dimension to a lower dimension as well as controlling the FDR with three popular variable selection methods: LASSO, SCAD, and MCP.* Method.* The three methods with the proposed screenings were applied to prostate cancer data with presence of metastasis as the outcome.* Results.* Simulations showed that the three variable selection methods with the proposed screenings controlled the predefined FDR and produced high area under the receiver operating characteristic curve (AUROC) scores. In applying these methods to the prostate cancer example, LASSO and MCP selected 12 and 8 genes and produced AUROC scores of 0.746 and 0.764, respectively.* Conclusions*. We demonstrated that the variable selection methods with the sequential use of FDR and ISIS not only controlled the predefined FDR in the final models but also had relatively high AUROC scores.

#### 1. Introduction

Prognosis will continue to play a critical role in patient management and decision making in 21st century medicine. Advanced technologies for genomic profiling are now available and they include millions of sets of molecular data in these assays. A critical element of personalized medicine is utilizing and implementing validated diagnostic signatures (or classifiers) for diagnosing or treating cancer patients. These signatures are built and validated utilizing common statistical methods and machine learning tools. For example, the Decipher signature has been developed as a prognostic model to predict metastasis after radical prostatectomy in patients with prostate cancer [1]. The Decipher score is a 22-feature genomic classifier that has been used to predict metastasis and has been independently validated for prediction of prostate metastasis [2–5]. Another example is oncotypeDx that has been used to stratify randomization and guide treatment in women with breast cancer [6].

A vital step in model building is data reduction. It is assumed that there are several variables that are associated with the clinical outcome in the large dimensional data. The main purpose of the variable selection is to detect only those variables related to the response. Variable selection is composed of two steps: screening and model building. The screening step is to reduce the large number of variables into moderate size while maintaining most of the informative variables relevant to the clinical response. In contrast, in the model building step, investigators develop a single best model utilizing a proper evaluation criterion.

Penalized variable selection methods have played a key role in identifying important prognostic models in several areas in oncology [7–9]. Many articles focused on the development of methodologies related to “small N and large P” with the advent of high throughput technology in cancer. The sure independence screening (SIS) was introduced to reduce the high dimension to below the sample size to efficiently select the best subset of variables to predict clinical responses [10]. Although this approach is popular, it does not perform well under some situations. First, unimportant variables that are heavily correlated with important variables are more highly likely to be selected than important variables that are weakly associated with the response. Second, important variables that are not marginally significantly related to the response are screened out. Finally, there may be collinearity between variables that may impact the calculations of the individual predictors.

The iterative sure independence screening (ISIS) was proposed to overcome the above issues. The procedure is to apply iteratively high dimensional variable screening followed by the proper scale of variable selection until the best subset of variables with high predictive accuracy is obtained. ISIS screening, however, is also computationally intensive and leads to high false discovery rate (FDR) in ultra-high dimensional setting ( mils).

The oncology literature is rich in articles related to the use of validated signatures. Despite their abundance, comparisons and the performance of these various methods have not been studied. We propose to use the false discovery rate (FDR) of the multiple testing correction methods as a screening method to reduce the high dimension to lower dimension as well as controlling the false discovery rate in the final model. We investigate the feasibility of the sequential use of FDR screening method with the ISIS and utilize three popular variable selection methods: LASSO [11], SCAD [12, 13], and MCP [14], through the extensive simulation studies. To the best of our knowledge, this is the first paper that thoroughly analyzes and compares the performance of the variable selection methods with the sequential use of FDR and ISIS screening methods. We use a prostate cancer signature as an example [1] where the number of probes is around 1.4 million and the clinical outcome is binary in nature: presence of metastasis (presence of metastasis = 1, no metastasis = 0) by fitting models based on the simulation results.

In addition, we provide a broad review of the existing penalized variable selection methods with screening methods. The remainder of this paper is organized as follows. In Section 2, we provide general details of the screening methods of FDR [15] and ISIS [10] and the variable selection methods with the penalized logistic regression. In Section 3, we describe the simulation studies and in Section 4, we summarize the results of the simulations. We then apply the best screening methods from the simulation studies to the real data in Section 5. Finally in Section 6, we discuss our findings.

#### 2. Methods

We divide this section into several subsections describing the methods used in our paper. The screening section briefly discusses commonly used methods that reduce high dimensionality: false discovery rate (FDR) and iterative sure independence screening (ISIS). We then describe the methods needed to assess variable selection models. The final section considers three existing popular variable selection methods with the logistic regression. All simulations and calculations were carried out using glmnet and ISIS packages in the R library, and the code is available at https://www.duke.edu/halab001/FDR.

##### 2.1. Benjamini and Hochberg False Discovery Rate (FDR)

The false discovery rate is defined as the expected proportion of incorrectly rejected null hypotheses. That is, where is the number of falsely rejected hypotheses and is the total number of rejected hypotheses. We focus on the Benjamini and Hochberg FDR [15] method as a screening method in the simulation studies and application. Briefly, the procedure works as follows. Let denote the FDR, where .(1)Let be the values of the hypothesis tests and sort them from smallest to largest. Denote these ordered values by (2)Let , . If , then reject and if , then there is no rejection of the hypothesis.

##### 2.2. Iterative Sure Independence Screening (ISIS)

The ISIS method was proposed to overcome the difficulties caused by the sure independence screening [16]. Briefly, the algorithm works in the following way:(1)The likelihood of marginal logistic regression (LMLR) is computed for every . Then which is of the top ranked variables of the descending order list of the LMLR is selected to obtain the index set .(2)Apply those variables in to the penalized logistic models to obtain a subset of indices .(3)For every variable , the likelihood of the marginal logistic regression condition on the variables in is solved. Then the likelihood estimators are sorted in descending order and then the top ranked variables are selected to get the index set .(4)Apply those variables in to the penalized logistic models to obtain a new index set .(5)Steps (3) and (4) are repeated until = or .

##### 2.3. Regularizing Methods with Penalized Logistic Regression

The logistic regression is one of the most commonly used methods for assessing the relationship between a binary outcome and a set of covariates and building prognostic models of clinical outcomes. In addition, it is widely used in the classification of two classes such as the development of metastasis in prostate cancer [1]. The purpose of variable selection with the logistic regression model in high dimensional setting is to select the optimal subset of variables that will improve the prediction accuracy [17]. Variable selection in high dimensional setting is composed of two components: a likelihood function and a penalty function in order to obtain better estimates for prediction.

Let the covariates of individual be denoted as for and is the total number of covariates. The penalized logistic regression is as follows:where , a penalty, is function and is 1 for cases and 0 for controls. The probability that individual is a case based on covariates’ information is expressed asThe regression coefficients are obtained by minimizing the objective function (2).

One of the most popular penalty functions is the least absolute shrinkage and selection operator (LASSO) [11]. It forces the coefficients of unimportant variables to be set to 0 and then the LASSO has sparsity property. The LASSO estimates are obtained by minimizing the above penalized logistic regression form (2). It has a satisfactory performance in identifying a small number of representative variables. Though LASSO is widely used in most applications [18–21], its robustness is open to question as it has the tendency to randomly select one of the variables with high correlation and exclude the rest of the predictors [22]. Another disadvantage of LASSO is that it always chooses at most (sample size) number of predictors even though there are more than variables with true nonzero coefficients [23]. The coefficients estimates are obtained by minimizing the following objective function based on the likelihood function of logistic regression:

Another method commonly employed is the smoothly clipped absolute deviation (SCAD) with a concave penalty function that overcomes some of the limitations of the LASSO [12]. The coefficients from SCAD are solved by minimizing the following objective function:The SCAD penalty function, , is defined by with and .

The minimum concave penalty (MCP) is also a recognized method with SCAD, where the coefficients are estimated via minimization of the following objective function:The MCP penalty function, , is defined by for and .

#### 3. Simulation Studies

##### 3.1. Simulation Setup

We performed extensive simulation studies to explore the performance of three popular variable selection methods: LASSO, SCAD, and MCP in high dimensional setting. We employed 10-fold cross validation to tune the regularization parameter for the methods. Figure 1 describes the schema of the simulation procedures.