An R package for the analysis of (genetic) risk prediction studies.

Description

An R package for the analysis of (genetic) risk prediction studies.

Details

Fueled by the substantial gene discoveries from genome-wide association studies, there is increasing interest in investigating the predictive ability of genetic risk models. To assess the performance of genetic risk models, PredictABEL includes functions for the various measures and plots that have been used in empirical studies, including univariate and multivariate odds ratios (ORs) of the predictors, the c-statistic (or AUC), Hosmer-Lemeshow goodness of fit test, reclassification table, net reclassification improvement (NRI) and integrated discrimination improvement (IDI). The plots included are the ROC plot, calibration plot, discrimination box plot, predictiveness curve, and several risk distributions.

These functions can be applied to predicted risks that are obtained using logistic regression analysis, to weighted or unweighted risk scores, for which the functions are included in this package. The functions can also be used to assess risks or risk scores that are constructed using other methods, e.g., Cox Proportional Hazards regression analysis, which are not included in the current version. Risks obtained from other methods can be imported into R for assessment of the predictive performance.

The functions to construct the risk models using logistic regression analyses are specifically written for models that include genetic variables, eventually in addition to non-genetic factors, but they can also be applied to construct models that are based on non-genetic risk factors only.

Before using the functions fitLogRegModel for constructing a risk model or riskScore for computing risk scores, the following checks on the dataset are advisable to be done:

(1) Missing values: The logistic regression analyses and computation of the risk score are done only for subjects that have no missing data. In case of missing values, individuals with missing data can be removed from the dataset or imputation strategies can be used to fill in missing data. Subjects with missing data can be removed with the R function na.omit (available in stats package). Example: DataFileNew <- na.omit(DataFile) will make a new dataset (DataFileNew) with no missing values;

(2) Multicollinearity: When there is strong correlation between the predictor variables, regression coefficients may be estimated imprecisely and risks scores may be biased because the assumption of independent effects is violated. In genetic risk prediction studies, problems with multicollinearity should be expected when single nucleotide polymorphisms (SNPs) located in the same gene are in strong linkage disequilibrium (LD). For SNPs in LD it is common to select the variant with the lowest p-value in the model;

(3) Outliers: When the data contain significant outliers, either clinical variables with extreme values of the outcomes or extreme values resulting from errors in the data entry, these may impact the construction of the risk models and computation of the risks scores. Data should be carefully checked and outliers need to be removed or replaced, if justified;

(4) Recoding of data: In the computation of unweighted risk scores, it is assumed that the genetic variants are coded 0,1,2 representing the number of alleles carried. When variants are coded 0,1 representing a dominant or recessive effect of the alleles, the variables need to be recoded before unweighted risk scores can be computed.

To import data into R several alternative strategies can be used. Use the Hmisc package for importing SPSS and SAS data into R. Use "ExampleData <- read.table("DataName.txt", header=T, sep="\t")" for text files where variable names are included as column headers and data are separated by tabs. Use "ExampleData <- read.table("Name.csv", sep=",", header=T)" for comma-separated files with variable names as column headers. Use "setwd(dir)" to set the working directory to "dir". The datafile needs to be present in the working directory.

To export datafiles from R tables to a tab-delimited textfile with the first row as the name of the variables, use "write.table(R_Table, file="Name.txt", row.names=FALSE, sep="\t")" and when a comma-separated textfile is requested and variable names are provided in the first row, use "write.table(R_Table, file="Name.csv", row.names=FALSE, sep=",")". When the directory is not specified, the file will be saved in the working directory. For exporting R data into SPSS, SAS and Stata data, use functions in the the foreign package.

Several functions in this package depend on other R packages:

(1) Hmisc, is used to compute NRI and IDI;

(2) ROCR, is used to produce ROC plots;

(3) epitools, is used to compute univariate odds ratios;

(4) PBSmodelling, is used to produce predictiveness curve.

Acknowledgements

The authors would like to acknowledge Lennart Karssen, Maksim Struchalin and Linda Broer from the Department of Epidemiology, Erasmus Medical Center, Rotterdam for their valuable comments and suggestions to make this package.

Note

The current version of the package includes the basic measures and plots that are used in the assessment of (genetic) risk prediction models and the function to construct a simulated dataset that contains individual genotype data, estimated genetic risk and disease status, used for the evaluation of genetic risk models (see Janssens et al, Genet Med 2006). Planned extensions of the package include functions to construct risk models using Cox Proportional Hazards analysis for prospective data and assess the performance of risk models for time-to-event data.

Author(s)

Suman Kundu

Yurii S. Aulchenko

A. Cecile J.W. Janssens

References

S Kundu, YS Aulchenko, CM van Duijn, ACJW Janssens. PredictABEL: an R package for the assessment of risk prediction models. Eur J Epidemiol. 2011;26:261-4.

ACJW Janssens, JPA Ioannidis, CM van Duijn, J Little, MJ Khoury. Strengthening the Reporting of Genetic Risk Prediction Studies: The GRIPS Statement Proposal. Eur J Epidemiol. 2011;26:255-9.

ACJW Janssens, JPA Ioannidis, S Bedrosian, P Boffetta, SM Dolan, N Dowling, I Fortier, AN. Freedman, JM Grimshaw, J Gulcher, M Gwinn, MA Hlatky, H Janes, P Kraft, S Melillo, CJ O'Donnell, MJ Pencina, D Ransohoff, SD Schully, D Seminara, DM Winn, CF Wright, CM van Duijn, J Little, MJ Khoury. Strengthening the reporting of genetic risk prediction studies (GRIPS)-Elaboration and explanation. Eur J Epidemiol. 2011;26:313-37.

Aulchenko YS, Ripke S, Isaacs A, van Duijn CM. GenABEL: an R package for genome-wide association analysis. Bioinformatics 2007;23(10):1294-6.