Evaluating Statistical Methods Using Plasmode Data Sets in the Age of Massive Public Databases: An Illustration Using False Discovery Rates

Department of Statistics, Kansas State University, Manhattan, Kansas, United States of America.
PLoS Genetics (Impact Factor: 8.17). 07/2008; 4(6):e1000098. DOI: 10.1371/journal.pgen.1000098
Source: PubMed

ABSTRACT Plasmode is a term coined several years ago to describe data sets that are derived from real data but for which some truth is known. Omic techniques, most especially microarray and genomewide association studies, have catalyzed a new zeitgeist of data sharing that is making data and data sets publicly available on an unprecedented scale. Coupling such data resources with a science of plasmode use would allow statistical methodologists to vet proposed techniques empirically (as opposed to only theoretically) and with data that are by definition realistic and representative. We illustrate the technique of empirical statistics by consideration of a common task when analyzing high dimensional data: the simultaneous testing of hundreds or thousands of hypotheses to determine which, if any, show statistical significance warranting follow-on research. The now-common practice of multiple testing in high dimensional experiment (HDE) settings has generated new methods for detecting statistically significant results. Although such methods have heretofore been subject to comparative performance analysis using simulated data, simulating data that realistically reflect data from an actual HDE remains a challenge. We describe a simulation procedure using actual data from an HDE where some truth regarding parameters of interest is known. We use the procedure to compare estimates for the proportion of true null hypotheses, the false discovery rate (FDR), and a local version of FDR obtained from 15 different statistical methods.

Download full-text


Available from: Steve Xiang, Jun 18, 2015
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Validating statistical analysis methods for RNA sequencing (RNA-seq) experiments is a complex task. Researchers often find themselves having to decide between competing models or assessing the reliability of results obtained with a designated analysis program. Computer simulation has been the most frequently used procedure to verify the adequacy of a model. However, datasets generated by simulations depend on the parameterization and the assumptions of the selected model. Moreover, such datasets may constitute a partial representation of reality as the complexity or RNA-seq data is hard to mimic. We present the use of plasmode datasets to complement the evaluation of statistical models for RNA-seq data. A plasmode is a dataset obtained from experimental data but for which come truth is known. Using a set of simulated scenarios of technical and biological replicates, and public available datasets, we illustrate how to design algorithms to construct plasmodes under different experimental conditions. We contrast results from two types of methods for RNA-seq: (1) models based on negative binomial distribution (edgeR and DESeq), and (2) Gaussian models applied after transformation of data (MAANOVA). Results emphasize the fact that deciding what method to use may be experiment-specific due to the unknown distributions of expression levels. Plasmodes may contribute to choose which method to apply by using a similar pre-existing dataset. The promising results obtained from this approach, emphasize the need of promoting and improving systematic data sharing across the research community to facilitate plasmode building. Although we illustrate the use of plasmode for comparing differential expression analysis models, the flexibility of plasmode construction allows comparing upstream analysis, as normalization procedures or alignment pipelines, as well.
    Frontiers in Genetics 09/2013; 4:178. DOI:10.3389/fgene.2013.00178
  • [Show abstract] [Hide abstract]
    ABSTRACT: Longitudinal healthcare claim databases are frequently used for studying the comparative safety and effectiveness of medications, but results from these studies may be biased due to residual confounding. It is unclear whether methods for confounding adjustment that have been shown to perform well in small, simple nonrandomized studies are applicable to the large, complex pharmacoepidemiologic studies created from secondary healthcare data. Ordinary simulation approaches for evaluating the performance of statistical methods do not capture important features of healthcare claims. A statistical framework for creating replicated simulation datasets from an empirical cohort study in electronic healthcare claims data is developed and validated. The approach relies on resampling from the observed covariate and exposure data without modification in all simulated datasets to preserve the associations among these variables. Repeated outcomes are simulated using a true treatment effect of the investigator’s choice and the baseline hazard function estimated from the empirical data. As an example, this framework is applied to a study of high versus low-intensity statin use and cardiovascular outcomes. Simulated data is based on real data drawn from Medicare Parts A and B linked with a prescription drug insurance claims database maintained by Caremark. Properties of the data simulated using this framework are compared with the empirical data on which the simulations were based. In addition, the simulated datasets are used to compare variable selection strategies for confounder adjustment via the propensity score, including high-dimensional approaches that could not be evaluated with ordinary simulation methods. The simulated datasets are found to closely resemble the observed complex data structure but have the advantage of an investigator-specified exposure effect.
    Computational Statistics & Data Analysis 04/2014; 72:219–226. DOI:10.1016/j.csda.2013.10.018 · 1.15 Impact Factor
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Normalization of gene expression data has been studied for many years and various strategies have been formulated to deal with various types of data. Most normalization algorithms rely on the assumption that the number of up-regulated genes and the number of down-regulated genes are roughly the same. However, the well-known Golden Spike experiment presents a unique situation in which differentially regulated genes are biased toward one direction, thereby challenging the conclusions of previous bench mark studies. This study proposes two novel approaches, KDL and KDQ, based on kernel density estimation to improve upon the basic idea of invariant set selection. The key concept is to provide various importance scores to data points on the MA plot according to their proximity to the cluster of the null genes under the assumption that null genes are more densely distributed than those that are differentially regulated. The comparison is demonstrated in the Golden Spike experiment as well as with simulation data using the ROC curves and compression rates. KDL and KDQ in combination with GCRMA provided the best performance among all approaches. This study determined that methods based on invariant sets are better able to resolve the problem of asymmetry. Normalization, either before or after expression summary for probesets, improves performance to a similar degree.
    BMC Bioinformatics 06/2011; 12:222. DOI:10.1186/1471-2105-12-222 · 2.67 Impact Factor