Model selection

Description

Model selection using the stepwise procedure and the chosen criterion.

Usage

selectModel(X, y, fitFun = fitLinear, crit = mbic, Xm = NULL,
  stay = NULL, minpv = 0.15, multif = TRUE, crit.multif = bic,
  maxf = min(ncol(X), 70), minb = 0, fastST = FALSE, maxp = 1e+06,
  verbose = TRUE, file.out = NULL, ...)

Arguments

Details

To find the best model (linear or generalized), the following algorithm (a modification of the stepwise selection) is used [3]. In the first step the likelihood ratio tests between two regression models are performed: 1) with only the intercept, 2) with the intercept and every single variable from the matrix X. P-values are calculated and variables with p > minpv are excluded from the model selection procedure. In the second step (multi-forward) we start with the null model and add variables which decrease crit.multif (in order from the smallest p-value). The step is finished after we add maxf variables or none of remaining variables improve crit.multif. Then the classical backward selection is performed (with crit). When there is no variables to remove, the last step, the classical stepwise procedure, is performed (with crit).

Results from this four-step procedure should be very similar to the classical stepwise procedure (when we start with the null model and do not omit variables with high p-values) but the first one is much quicker. The most time-consuming part is the forward step in the stepwise selection (in the multi-forward step we do not add the best variable but any which decrease crit.multif) and it is performed less often when we start with a reasonable model (sometimes you can find the best model without using the stepwise selection). But you can omit the first three steps if you set multif=FALSE and minpv=1. Resignation from the multi-forward step can be reasonable when you expect that the final model should be very small (a few variables).

If your data are too big to store in RAM, you should read them with the read.big.matrix function from the bigmemory packages. The selectModel function will recognize that X is not an ordinary matrix and split your data to smaller parts. It will not change results but is necessary to work with big data.

The default criterion in the model selection procedure is a modification of the Bayesian Information Criterion, mBIC [1]. It was constructed to control the so-called Family-wise Error Rate (FWER) at the level near 0.05 when you have a lot of explanatory variables and only a few of them should stay in the final model. If you are interested in controlling the so-called False Discovery Rate (FDR) is such type of data, you can change crit to mBIC2 [2], which controls FDR at the level near 0.05. There are more criteria to choose from or you can easily define your own (see 'Examples')

If you do not have the desing matrix in one file, you have to combine them. It can be problematic if your data are big, so you can use the combineBigMatrices function from this package. You can use it also when you have to transpose X (because you have variables in rows) or you want to change names of columns.

Value

The names of variables in the final model.

Author(s)

Piotr Szulc

References

[1] M. Bogdan, J.K. Ghosh, R.W. Doerge (2004), "Modifying the Schwarz Bayesian Information Criterion to locate multiple interacting quantitative trait loci", Genetics 167: 989-999.

[2] F. Frommlet, A. Chakrabarti, M. Murawska, M. Bogdan (2011), "Asymptotic Bayes optimality under sparsity for generally distributed effect sizes under the alternative". Technical report at arXiv:1005.4753.

[3] F. Frommlet, F. Ruhaltinger, P. Twarog, M. Bogdan (2012), "A model selection approach to genome wide association studies", Computational Statistics and Data Analysis 56: 1038-1051.

Examples

set.seed(1)
n <- 100
M <- 10
X <- matrix(rnorm(M*n), ncol=M)
colnames(X) <- 1:M
y <- X[, 2] - X[, 3] + X[, 6] - X[, 10] + rnorm(n)
selectModel(X, y, p=M)

# more examples: type ?bigstep