nullModel-methods: Create Null Model for Association Test

In UBod/podkat: Position-Dependent Kernel Association Test

Create Null Model for Association Test

Description

Method for creating a null model that can be used for association testing using assocTest

Usage

## S4 method for signature 'formula,data.frame'
nullModel(X, y, data,
          type=c("automatic", "logistic", "linear", "bernoulli"),
          n.resampling=0,
          type.resampling=c("bootstrap", "permutation"),
          adj=c("automatic", "none", "force"), adjExact=FALSE,
          n.resampling.adj=10000, checkData=TRUE)
## S4 method for signature 'formula,missing'
nullModel(X, y, data,
          type=c("automatic", "logistic", "linear", "bernoulli"),
          n.resampling=0,
          type.resampling=c("bootstrap", "permutation"),
          adj=c("automatic", "none", "force"), adjExact=FALSE,
          n.resampling.adj=10000, checkData=TRUE)
## S4 method for signature 'matrix,numeric'
nullModel(X, y,
          type=c("automatic", "logistic", "linear"), ...)
## S4 method for signature 'matrix,factor'
nullModel(X, y,
          type=c("automatic", "logistic", "linear"), ...)
## S4 method for signature 'missing,numeric'
nullModel(X, y,
          type=c("automatic", "logistic", "linear", "bernoulli"),
          ...)
## S4 method for signature 'missing,factor'
nullModel(X, y,
          type=c("automatic", "logistic", "linear", "bernoulli"),
          ...)

Arguments

X	a formula or matrix
y	if the formula interface is used, y can be used to pass a data frame with the table in which both covariates and traits are contained (alternatively, the data argument can be used for that purpose). The other methods (if X is not a formula) expect y to be the trait vector. Trait vectors can either be numeric vectors or a factor with two levels (see details below).
data	for consistency with standard R methods from the stats package, the data frame can also be passed to nullModel via the data argument. In this case, the y must be empty. If y is specified, data is ignored.
type	type of model to train (see details below)
n.resampling	number of null model residuals to sample; set to zero (default) to turn resampling off; resampling is not supported for plain trait vectors without covariates
type.resampling	method how to sample null model residuals; the choice “permutation” refers to simple random permutations of the model's residuals. If “bootstrap” is chosen (default), the following strategy is applied for linear models (continuous trait): residuals are sampled as normally distributed values with mean 0 and the same standard deviation as the model's residuals. For logistic models (binary trait), the choice “bootstrap” selects the same bootstrapping method that is implemented in the SKAT package.
adj	whether or not to use small sample correction for logistic models (binary trait with covariates). The choice “none” turns off small sample correction. If “force” is chosen, small sample correction is turned on unconditionally. If “automatic” is chosen (default), small sample correction is turned on if the number of samples does not exceed 2,000. This argument is ignored for any type of model except “logistic” and small sample correction is switched off.
adjExact	in case small sample correction is switched on (see above), this argument indicates whether or not the exact square root of the matrix P_0 should be pre-computed (see Subsection 9.5 of the package vignette). The default is FALSE. This argument is ignored if small sample correction is not switched on.
n.resampling.adj	number of null model residuals to sample for the adjustment of higher moments; ignored if small sample correction is switched off.
checkData	if FALSE, only a very limited set of input checks is performed. The purpose of this option is to save computational effort for repeated input checks if the function is called from a function that has already performed input checks. The default is TRUE. Only change to FALSE if you know what you are doing!
...	all other parameters are passed on to the nullModel method with signature formula,data.frame.

Details

The podkat package assumes a mixed model in which the trait under investigation depends both on covariates (if any) and the genotype. The nullModel method models the relationship between the trait and the covariates (if any) without taking the genotype into account, which corresponds to the null assumption that the trait and the genotype are independent. Therefore, we speak of null models. The following types of models are presently available:

Linear model (type “linear”):

a linear model is trained for a continuous trait and a given set of covariates (if any); this is done by standard linear regression using the lm function.

Logistic linear model (type “logistic”):

a generalized linear model is trained for a binary trait and a given set of covariates (if any); this is done by logistic regression using the glm function.

Bernoulli-distributed trait (type “bernoulli”):

a binary trait without covariates is interpreted as instances of a simple Bernoulli process with p being the relative frequencies 1's/cases.

The type argument can be used to select the type of model, where the following restrictions apply:

• For linear models, the trait vector must be numerical. Factors/factor columns are not accepted.

• For logistic models and Bernoulli-distributed traits, both numerical vectors and factors are acceptable. In any case, only 0's (controls) and 1's (cases) are accepted. Furthermore, nullModel quits with an error if the trait shows no variation. In other words, trait vectors that only contain 0's or only contain 1's are not accepted (as association testings makes little sense for such traits anyway).

The following interfaces are available to specify the traits and the covariates (if any):

Formula interface:

the first argument X can be a formula that specifies the trait vector/column, the covariate matrix/columns (if any), and the intercept (if any). If neither the y argument nor the data argument is specified, nullModel searches the environment from which the function has been called. This interface is largely analogous to the functions lm and glm.

Trait vector without covariates:

if the X argument is omitted and y is a numeric vector or factor, y is interpreted as trait vector, and a null model is created from y without covariates. Linear and logistic models are trained with an intercept. For type “bernoulli”, the trait vector is written to the output object as is.

Trait vector plus covariate matrix:

if the X argument is a matrix and y is a numeric vector or factor, y is interpreted as trait vector and X is interpreted as covariate matrix. In this case, linear and logistic models are trained as (generalized) linear regressors that predict the trait from the covariates plus an intercept. The type “bernoulli” is not available for this variant, since this type of model cannot consider covariates.

All nullModel methods also support the choice type="automatic". In this case, nullModel guesses the most reasonable type of model in the following way: If the trait vector/column is a factor or a numeric vector containing only 0's and 1's (where both values must be present, as noted above already), the trait is supposed to be binary and the type “logistic” is assumed, unless the following conditions are satisfied:

1. The number of samples does not exceed 100.

2. No intercept and no covariates have been specified. This condition can be met by supplying an empty model to the formula interface (e.g. y ~ 0) or by supplying the trait vector as argument y while omitting X.

If these two conditions are fulfilled for a binary trait, nullModel chooses the type “bernoulli”. If the trait is not binary and the trait vector/column is numeric, nullModel assumes type “linear”.

For consistency with the SKAT package, the podkat package also offers resampling, i.e. a certain number of vectors of residuals are sampled according to the null model. This can be done when training the null model by setting the n.resampling parameter (number of residual vectors that are sampled) to a value larger than 0. Then, when association testing is performed, p-values are computed also for all these sampled residuals, and an additional estimated p-value is computed as the relative frequency of p-values of sampled residuals that are at least as significant as the test's p-value. The procedure to sample residuals is controlled with the type.resampling argument (see above).

For logistic models (type “logistic”), assocTest offers the small sample correction as introduced by Lee et al. (2012). If the adjustment of higher moments should be applied, some preparations need to be made already when training the null model. Which preparations are carried out, can be controlled by the arguments adj, adjExact, n.resampling.adj, and type.resampling (see descriptions of arguments above and Subsection 9.5 of the package vignette).

If any missing values are found in the trait vector/column or the covariate matrix/columns, the respective samples are omitted from the resulting model (which is the standard behavior of lm and glm anyway). The indices of the omitted samples are stored in the na.omit slot of the returned NullModel object.

Value

returns a NullModel object

Author(s)

Ulrich Bodenhofer [e-mail]

References

http://www.bioinf.jku.at/software/podkat

Lee, S., Emond, M. J., Bamshad, M. J., Barnes, K. C., Rieder, M. J., Nickerson, D. A., NHLBI Exome Sequencing Project - ESP Lung Project Team, Christiani, D. C., Wurfel, M. M., and Lin, X. (2012) Optimal unified approach for rare-variant association testing with application to small-sample case-control whole-exome sequencing studies. Am. J. Hum. Genet. 91, 224-237. DOI: 10.1016/j.ajhg.2012.06.007.

See Also

NullModel, lm, glm

Examples

## read phenotype data from CSV file (continuous trait + covariates)
phenoFile <- system.file("examples/example1lin.csv", package="podkat")
pheno <-read.table(phenoFile, header=TRUE, sep=",")

## train null model with all covariates in data frame 'pheno'
model <- nullModel(y ~ ., pheno)
model
length(model)
residuals(model)

## read phenotype data from CSV file (binary trait + covariates)
phenoFile <- system.file("examples/example1log.csv", package="podkat")
pheno <-read.table(phenoFile, header=TRUE, sep=",")

## train null model with all covariates in data frame 'pheno'
model <- nullModel(y ~ ., pheno)
model
length(model)
residuals(model)

## "train" simple Bernoulli model on a subset of 100 samples
model <- nullModel(y ~ 0, pheno[1:100, ])
model
length(model)
residuals(model)

## alternatively, use the interface that only supplies the
## trait vector
model <- nullModel(y=pheno[1:100, ]$y)
model

UBod/podkat documentation built on Feb. 18, 2024, 2:32 a.m.