Statistical Methods in Medical Research

Guarding against false positive selections is important in many applications. We discuss methods based on subsampling and sample splitting for controlling the expected number of false positives and assigning p-values. They are generic and especially useful for high-dimensional settings. We review encouraging results for regression, and we discuss new adaptations and remaining challenges for selecting relevant variables, based on observational data, having a causal or interventional effect on a response of interest.

In decision-making on optimal treatment strategies, it is of great importance to identify variables that are involved in the decision rule, i.e. those interacting with the treatment. Effective variable selection helps to improve the prediction accuracy and enhance the interpretability of the decision rule. We propose a new penalized regression framework which can simultaneously estimate the optimal treatment strategy and identify important variables. The advantages of the new approach include: (i) it does not require the estimation of the baseline mean function of the response, which greatly improves the robustness of the estimator; (ii) the convenient loss-based framework makes it easier to adopt shrinkage methods for variable selection, which greatly facilitates implementation and statistical inferences for the estimator. The new procedure can be easily implemented by existing state-of-art software packages like LARS. Theoretical properties of the new estimator are studied. Its empirical performance is evaluated using simulation studies and further illustrated with an application to an AIDS clinical trial.

We propose a cross-validated area under the receiving operator characteristic (ROC) curve (CV-AUC) criterion for tuning parameter selection for penalized methods in sparse, high-dimensional logistic regression models. We use this criterion in combination with the minimax concave penalty (MCP) method for variable selection. The CV-AUC criterion is specifically designed for optimizing the classification performance for binary outcome data. To implement the proposed approach, we derive an efficient coordinate descent algorithm to compute the MCP-logistic regression solution surface. Simulation studies are conducted to evaluate the finite sample performance of the proposed method and its comparison with the existing methods including the Akaike information criterion (AIC), Bayesian information criterion (BIC) or Extended BIC (EBIC). The model selected based on the CV-AUC criterion tends to have a larger predictive AUC and smaller classification error than those with tuning parameters selected using the AIC, BIC or EBIC. We illustrate the application of the MCP-logistic regression with the CV-AUC criterion on three microarray datasets from the studies that attempt to identify genes related to cancers. Our simulation studies and data examples demonstrate that the CV-AUC is an attractive method for tuning parameter selection for penalized methods in high-dimensional logistic regression models.

We discuss the identification of features that are associated with an outcome in RNA-Sequencing (RNA-Seq) and other sequencing-based comparative genomic experiments. RNA-Seq data takes the form of counts, so models based on the normal distribution are generally unsuitable. The problem is especially challenging because different sequencing experiments may generate quite different total numbers of reads, or ‘sequencing depths’. Existing methods for this problem are based on Poisson or negative binomial models: they are useful but can be heavily influenced by ‘outliers’ in the data. We introduce a simple, non-parametric method with resampling to account for the different sequencing depths. The new method is more robust than parametric methods. It can be applied to data with quantitative, survival, two-class or multiple-class outcomes. We compare our proposed method to Poisson and negative binomial-based methods in simulated and real data sets, and find that our method discovers more consistent patterns than competing methods.

We review in this article several classification methods, especially for high-dimensional and low-sample size data. We discuss several desirable properties for classifiers in such settings, including predictability, consistency, generality, stability, robustness and sparsity. Specifically, a good classifier should have a small prediction error (predictability); converge to the Bayes-rule classifier asymptotically (consistency); be stable when adding/removing an observation (generality); be stable for different data sets of the same kind (stochastic stability); be stable when there are a small number of contaminated observations (robustness); and have a small number of variables in the classifier (interpretability or sparsity). Several simulation examples and real applications are used to illustrate the usefulness of the existing popular classifiers and compare their performance.