Fit GLMM For Many Features

Fit regression model y ~ design + X_features[,j] for each feature j

Usage

glmmFitFeatures(
  y,
  design,
  data,
  family,
  U,
  s,
  weights = NULL,
  offset = NULL,
  dcmpMethod = c("general", "categorical"),
  delta = NULL,
  delta.range = c(-10, 10),
  tol = 0.001,
  tol.eta = 0.001,
  detail = 1,
  nthreads = 1,
  fastApprox = FALSE,
  lambda = 0,
  verbose = TRUE,
  ...
)

# S4 method for class 'ANY,ANY,matrix'
glmmFitFeatures(
  y,
  design,
  data,
  family,
  U,
  s,
  weights = NULL,
  offset = NULL,
  dcmpMethod = c("general", "categorical"),
  delta = NULL,
  delta.range = c(-10, 10),
  tol = 0.001,
  tol.eta = 0.001,
  detail = 1,
  nthreads = 1,
  fastApprox = FALSE,
  lambda = 0,
  verbose = TRUE,
  ...
)

Arguments

y

response vector

design

design matrix shared across all models

data

feature matrix with model j using feature j

family

a description of the error distribution and link function to be used in the modelm just like for glm(). Also supports negative binomial as string "nb:theta".

U

eigen-vectors of random effect

s

eigen-values of random effect

weights

sample-level weights

offset

sample-level offset values

dcmpMethod

use a "general" method (default) for SVD of Z. If Z is a categorical design matrix, used faster method "categorical"

delta

if NULL estimate delta, if value is given use this fixed value

delta.range

min and max values (in log space), of the search space for delta to fit the random effect

tol

convergence criterion for the 1D search of the delta space

tol.eta

convergence criterion in the PQL iteration

detail

level of model detail returned, with LEAST = 0, LOW = 1, MEDIUM = 2, HIGH = 3, MOST = 4, MAX = 5. LEAST (beta), LOW (beta, se, sigSq, rdf), MEDIUM (vcov), HIGH (residuals), MOST (hatvalues), MAX (deviance residuals)

nthreads

number of threads. Each model is fit in serial, analysis is parallelized across features

fastApprox

default false. if true, use pre-projection on the working response from an initial regression fit on only the design. Under the null for data, this is a very good approximation and _much_ faster

lambda

ridge shrinkage parameter

verbose

show progress

...

other args

Value

List of parameter estimates with entries coef, se, sigSq, rdf and other depending on detail

Examples

library(fastglmm)
#> Loading required package: nlme
library(lme4)
#> Loading required package: Matrix
#> 
#> Attaching package: ‘lme4’
#> The following object is masked from ‘package:nlme’:
#> 
#>     lmList

set.seed(1)
sleepstudy$V = rnorm(nrow(sleepstudy))

# lmer
fit <- lmer( Reaction ~ Days + V + (1 | Subject), sleepstudy)
fit
#> Linear mixed model fit by REML ['lmerMod']
#> Formula: Reaction ~ Days + V + (1 | Subject)
#>    Data: sleepstudy
#> REML criterion at convergence: 1782.453
#> Random effects:
#>  Groups   Name        Std.Dev.
#>  Subject  (Intercept) 37.12   
#>  Residual             31.06   
#> Number of obs: 180, groups:  Subject, 18
#> Fixed Effects:
#> (Intercept)         Days            V  
#>     251.457       10.472       -1.295  

# prepare response, design and random efect
y = sleepstudy$Reaction
design = model.matrix(~ Days, sleepstudy)
dcmp = indicator_decomp(sleepstudy$Subject)

data = as.matrix(sleepstudy$V)
rownames(data) = rownames(sleepstudy)
colnames(data) = paste0("SNP_", seq(ncol(data)))

fit1 = glmmFitFeatures(y, design, data, dcmp$vectors, dcmp$values, family=gaussian())
coef(fit1)
#>       (Intercept)     Days         x
#> SNP_1    251.4575 10.47224 -1.295992