Fast imputation

Fast imputation algorithm based on local XGBoost models.

snp_fastImpute(
  Gna,
  infos.chr,
  alpha = 1e-04,
  size = 200,
  p.train = 0.8,
  n.cor = nrow(Gna),
  seed = NA,
  ncores = 1
)

Arguments

Gna

A FBM.code256 (typically <bigSNP>$genotypes).
You can have missing values in these data.

infos.chr

Vector of integers specifying each SNP's chromosome.
Typically <bigSNP>$map$chromosome.

alpha

Type-I error for testing correlations. Default is 1e-4.

size

Number of neighbor SNPs to be possibly included in the model imputing this particular SNP. Default is 200.

p.train

Proportion of non missing genotypes that are used for training the imputation model while the rest is used to assess the accuracy of this imputation model. Default is 0.8.

n.cor

Number of rows that are used to estimate correlations. Default uses them all.

seed

An integer, for reproducibility. Default doesn't use seeds.

ncores

Number of cores used. Default doesn't use parallelism. You may use bigstatsr::nb_cores().

Value

An FBM with

• the proportion of missing values by SNP (first row),

• the estimated proportion of imputation errors by SNP (second row).

See also

snp_fastImputeSimple()

Examples

if (FALSE) {

fake <- snp_attachExtdata("example-missing.bed")
G <- fake$genotypes
CHR <- fake$map$chromosome
infos <- snp_fastImpute(G, CHR)
infos[, 1:5]

# Still missing values
big_counts(G, ind.col = 1:10)
# You need to change the code of G
# To make this permanent, you need to save (modify) the file on disk
fake$genotypes$code256 <- CODE_IMPUTE_PRED
fake <- snp_save(fake)
big_counts(fake$genotypes, ind.col = 1:10)

# Plot for post-checking
## Here there is no SNP with more than 1% error (estimated)
pvals <- c(0.01, 0.005, 0.002, 0.001); colvals <- 2:5
df <- data.frame(pNA = infos[1, ], pError = infos[2, ])

# base R
plot(subset(df, pNA > 0.001), pch = 20)
idc <- lapply(seq_along(pvals), function(i) {
  curve(pvals[i] / x, from = 0, lwd = 2,
        col = colvals[i], add = TRUE)
})
legend("topright", legend = pvals, title = "p(NA & Error)",
       col = colvals, lty = 1, lwd = 2)

# ggplot2
library(ggplot2)
Reduce(function(p, i) {
  p + stat_function(fun = function(x) pvals[i] / x, color = colvals[i])
}, x = seq_along(pvals), init = ggplot(df, aes(pNA, pError))) +
  geom_point() +
  coord_cartesian(ylim = range(df$pError, na.rm = TRUE)) +
  theme_bigstatsr()
}