vruvinv: Variance inflated RUV-inverse.
In dcgerard/vicar: Various Ideas for Confounder Adjustment in Regression

Description Usage Arguments Details Value Author(s) References

RUV-inverse as described in Gagnon-Bartsch et al (2013) is just RUV4 using the maximum number of confounders allowed by the model, followed by estimating the variances using a method-of-moments type approach. vruvinv is similar in spirit to this by using the maximum number of confounders allowed by the model, but we still estimate the variance inflation using the confounders. We force limmashrink = TRUE, otherwise there is only one degree of freedom to estimate the variance.

vruvinv(
  Y,
  X,
  ctl,
  cov_of_interest = ncol(X),
  likelihood = c("t", "normal"),
  include_intercept = TRUE,
  fa_func = pca_naive,
  fa_args = list(),
  adjust_bias = FALSE
)

`Y`	A matrix of numerics. These are the response variables where each column has its own variance. In a gene expression study, the rows are the individuals and the columns are the genes.
`X`	A matrix of numerics. The covariates of interest.
`ctl`	A vector of logicals of length `ncol(Y)`. If position i is `TRUE` then position i is considered a negative control.
`cov_of_interest`	A vector of positive integers. The column numbers of the covariates in X whose coefficients you are interested in. The rest are considered nuisance parameters and are regressed out by OLS.
`likelihood`	Either `"normal"` or `"t"`. If `likelihood = "t"`, then the user may provide the degrees of freedom via `degrees_freedom`.
`include_intercept`	A logical. If `TRUE`, then it will check `X` to see if it has an intercept term. If not, then it will add an intercept term. If `FALSE`, then `X` will be unchanged.
`fa_func`	A factor analysis function. The function must have as inputs a numeric matrix `Y` and a rank (numeric scalar) `r`. It must output numeric matrices `alpha` and `Z` and a numeric vector `sig_diag`. `alpha` is the estimate of the coefficients of the unobserved confounders, so it must be an `r` by `ncol(Y)` matrix. `Z` must be an `r` by `nrow(Y)` matrix. `sig_diag` is the estimate of the column-wise variances so it must be of length `ncol(Y)`. The default is the function `pca_naive` that just uses the first `r` singular vectors as the estimate of `alpha`. The estimated variances are just the column-wise mean square.
`fa_args`	A list. Additional arguments you want to pass to fa_func.
`adjust_bias`	A logical. Should we also use the control genes to adjust for bias (`TRUE`) or not (`FALSE`)

Hence, this is just a wrapper for vruv4, but with a special choice for the number of confounders and forcing limmashrink = TRUE.

See vruv4 for the elements returned.

David Gerard

Gagnon-Bartsch, J., Laurent Jacob, and Terence P. Speed, 2013. "Removing unwanted variation from high dimensional data with negative controls." Berkeley: Department of Statistics. University of California. https://statistics.berkeley.edu/tech-reports/820

dcgerard/vicar documentation built on July 7, 2021, 1:08 p.m.