empiricalBayesLM: Empirical Bayes-moderated adjustment for unwanted covariates

In WGCNA: Weighted Correlation Network Analysis

Empirical Bayes-moderated adjustment for unwanted covariates

Description

This functions removes variation in high-dimensional data due to unwanted covariates while preserving variation due to retained covariates. To prevent numerical instability, it uses Empirical bayes-moderated linear regression, optionally in a robust (outlier-resistant) form.

Usage

empiricalBayesLM(
  data,
  removedCovariates,
  retainedCovariates = NULL,

  initialFitFunction = NULL,
  initialFitOptions = NULL,
  initialFitRequiresFormula = NULL,
  initialFit.returnWeightName = NULL,

  fitToSamples = NULL,

  weights = NULL,
  automaticWeights = c("none", "bicov"),
  aw.maxPOutliers = 0.1,
  weightType = c("apriori", "empirical"),
  stopOnSmallWeights = TRUE,

  minDesignDeviation = 1e-10,
  robustPriors = FALSE,
  tol = 1e-4, maxIterations = 1000,
  garbageCollectInterval = 50000,

  scaleMeanToSamples = fitToSamples,
  scaleMeanOfSamples = NULL,
  getOLSAdjustedData = TRUE,
  getResiduals = TRUE,
  getFittedValues = TRUE,
  getWeights = TRUE,
  getEBadjustedData = TRUE,

  verbose = 0, indent = 0)

Arguments

data	A 2-dimensional matrix or data frame of numeric data to be adjusted. Variables (for example, genes or methylation profiles) should be in columns and observations (samples) should be in rows.
removedCovariates	A vector or two-dimensional object (matrix or data frame) giving the covariates whose effect on the data is to be removed. At least one such covariate must be given.
retainedCovariates	A vector or two-dimensional object (matrix or data frame) giving the covariates whose effect on the data is to be retained. May be NULL if there are no such "retained" covariates.
initialFitFunction	Function name to perform the initial fit. The default is to use the internal implementation of linear model fitting. The function must take arguments formula and data or x and y, plus possibly additional arguments. The return value must be a list with component coefficients, either scale or residuals, and weights must be returned in component specified by initialFit.returnWeightName. See lm, rlm and other standard fit functions for examples of suitable functions.
initialFitOptions	Optional specifications of extra arguments for initialFitFunction, apart from formula and data or x and y. Defaults are provided for function rlm, i.e., if this function is used as initialFitFunction, suitable initial fit options will be chosen automatically.
initialFitRequiresFormula	Logical: does the initial fit function need formula and data arguments? If TRUE, initialFitFunction will be called with arguments formula and data, otherwise with arguments x and y.
initialFit.returnWeightName	Name of the component of the return value of initialFitFunction that contains the weights used in the fit. Suitable default value will be chosen automatically for rlm.
fitToSamples	Optional index of samples from which the linear model fits should be calculated. Defaults to all samples. If given, the models will be only fit to the specified samples but all samples will be transformed using the calculated coefficients.
weights	Optional 2-dimensional matrix or data frame of the same dimensions as data giving weights for each entry in data. These weights will be used in the initial fit and are are separate from the ones returned by initialFitFunction if it is specified.
automaticWeights	One of (unique abrreviations of) "none" or "bicov", instructing the function to calculate weights from the given data. Value "none" will result in trivial weights; value "bicov" will result in biweight midcovariance weights being used.
aw.maxPOutliers	If automaticWeights above is "bicov", this argument gets passed to the function bicovWeights and determines the maximum proportion of outliers in calculating the weights. See bicovWeights for more details.
weightType	One of (unique abbreviations of) "apriori" or "empirical". Determines whether a standard ("apriori") or a modified ("empirical") weighted regression is used. The "apriori" choice is suitable for weights that have been determined without knowledge of the actual data, while "empirical" is appropriate for situations where one wants to down-weigh cartain entries of data because they may be outliers. In either case, the weights should be determined in a way that is independent of the covariates (both retained and removed).
stopOnSmallWeights	Logical: should presence of small "apriori" weights trigger an error? Because standard weighted regression assumes that all weights are non-zero (otherwise estimates of standard errors will be biased), this function will by default complain about the presence of too small "apriori" weights.
minDesignDeviation	Minimum standard deviation for columns of the design matrix to be retained. Columns with standard deviations below this number will be removed (effectively removing the corresponding terms from the design).
robustPriors	Logical: should robust priors be used? This essentially means replacing mean by median and covariance by biweight mid-covariance.
tol	Convergence criterion used in the numerical equation solver. When the relative change in coefficients falls below this threshold, the system will be considered to have converged.
maxIterations	Maximum number of iterations to use.
garbageCollectInterval	Number of variables after which to call garbage collection.
scaleMeanToSamples	Optional specification of samples (given as a vector of indices) to whose means the resulting adjusted data should be scaled (more precisely, shifted).
scaleMeanOfSamples	Optional specification of samples (given as a vector of indices) that will be used in calculating the shift. Specifically, the shift is such that the mean of samples given in scaleMeanOfSamples will equal the mean of samples given in scaleMeanToSamples. Defaults to all samples.
getOLSAdjustedData	Logical: should data adjusted by ordinary least squares or by initialFitFunction, if specified, be returned?
getResiduals	Logical: should the residuals (adjusted values without the means) be returned?
getFittedValues	Logical: should fitted values be returned?
getWeights	Logical: should the final weights be returned?
getEBadjustedData	Logical: should the EB step be performed and the adjusted data returned? If this is FALSE, the function acts as a rather slow but still potentially useful adjustment using standard fit functions.
verbose	Level of verbosity. Zero means silent, higher values result in more diagnostic messages being printed.
indent	Indentation of diagnostic messages. Each unit adds two spaces.

Details

This function uses Empirical Bayes-moderated (EB) linear regression to remove variation in data due to the variables in removedCovariates while retaining variation due to variables in retainedCovariates, if any are given. The EB step uses simple normal priors on the regression coefficients and inverse gamma priors on the variances. The procedure starts with multivariate ordinary linear regression of individual columns in data on retainedCovariates and removedCovariates. Alternatively, the user may specify an intial fit function (e.g., robust linear regression). To make the coefficients comparable, columns of data are scaled to (weighted if weights are given) mean 0 and variance 1. The resulting regression coefficients are used to determine the parameters of the normal prior (mean, covariance, and inverse gamma or median and biweight mid-covariance if robust priors are used), and the variances are used to determine the parameters of the inverse gamma prior. The EB step then essentially shrinks the coefficients toward their means, with the amount of shrinkage determined by the prior covariance.

Using appropriate weights can make the data adjustment robust to outliers. This can be achieved automatically by using the argument automaticWeights = "bicov". When bicov weights are used, we also recommend setting the argument maxPOutliers to a maximum proportion of samples that could be outliers. This is especially important if some of the design variables are binary and can be expected to have a strong effect on some of the columns in data, since standard biweight midcorrelation (and its weights) do not work well on bimodal data.

The automatic bicov weights are determined from data only. It is implicitly assumed that there are no outliers in the retained and removed covariates. Outliers in the covariates are more difficult to work with since, even if the regression is made robust to them, they can influence the adjusted values for the sample in which they appear. Unless the the covariate outliers can be attributed to a relevant variation in experimental conditions, samples with covariate outliers are best removed entirely before calling this function.

Value

A list with the following components (some of which may be missing depending on input options):

adjustedData	A matrix of the same dimensions as the input data, giving the adjusted data. If input data has non-NULL dimnames, these are copied.
residuals	A matrix of the same dimensions as the input data, giving the residuals, that is, adjusted data with zero means.
coefficients	A matrix of regression coefficients. Rows correspond to the design matrix variables (mean, retained and removed covariates) and columns correspond to the variables (columns) in data.
coefficiens.scaled	A matrix of regression coefficients corresponding to columns in data scaled to mean 0 and variance 1.
sigmaSq	Estimated error variances (one for each column of input data.
sigmaSq.scaled	Estimated error variances corresponding to columns in data scaled to mean 0 and variance 1.
fittedValues	Fitted values calculated from the means and coefficients corresponding to the removed covariates, i.e., roughly the values that are subtracted out of the data.
adjustedData.OLS	A matrix of the same dimensions as the input data, giving the data adjusted by ordinary least squares. This component should only be used for diagnostic purposes, not as input for further downstream analyses, as the OLS adjustment is inferior to EB adjustment.
residuals.OLS	A matrix of the same dimensions as the input data, giving the residuals obtained from ordinary least squares regression, that is, OLS-adjusted data with zero means.
coefficients.OLS	A matrix of ordinary least squares regression coefficients. Rows correspond to the design matrix variables (mean, retained and removed covariates) and columns correspond to the variables (columns) in data.
coefficiens.OLS.scaled	A matrix of ordinary least squares regression coefficients corresponding to columns in data scaled to mean 0 and variance 1. These coefficients are used to calculate priors for the EB step.
sigmaSq.OLS	Estimated OLS error variances (one for each column of input data.
sigmaSq.OLS.scaled	Estimated OLS error variances corresponding to columns in data scaled to mean 0 and variance 1. These are used to calculate variance priors for the EB step.
fittedValues.OLS	OLS fitted values calculated from the means and coefficients corresponding to the removed covariates.
weights	A matrix of weights used in the regression models. The matrix has the same dimension as the input data.
dataColumnValid	Logical vector with one element per column of input data, indicating whether the column was adjusted. Columns with zero variance or too many missing data cannot be adjusted.
dataColumnWithZeroVariance	Logical vector with one element per column of input data, indicating whether the column had zero variance.
coefficientValid	Logical matrix of the dimension (number of covariates +1) times (number of variables in data), indicating whether the corresponding regression coefficient is valid. Invalid regression coefficients may be returned as missing values or as zeroes.

Author(s)

Peter Langfelder

See Also

bicovWeights for suitable weights that make the adjustment robust to outliers.

WGCNA documentation built on Sept. 18, 2024, 5:08 p.m.