Row-wise logistic regressions are applied to a matrix with counts.
For each row, an overall test comparing the column counts across
columns is performed. Optionally, chi-square permutation tests are
used when the expected counts are below 5 for some column.
If set to TRUE, an exact test is used whenever
some expected cell counts are 5 or less
p.adjust.method
p-value adjustment method, passed on to p.adjust
Details
For each column, the proportion of counts in each row (with respect to
the overall counts in that column) is computed. Then a statistical
comparison of these proportions across groups is performed via a
likelihood-ratio test (if exact==TRUE a permutation based
chi-square test is used whenever the expected counts in some column is
below 5).
Notice that data from column j can be viewed as a multinomial
distribution with probabilities pj, where pj is a vector of length
nrow(x).
rowLogRegLRT tests the null hypothesis p1[i]=...pc[i] for
i=1...nrow(x),
where c is ncol(x).
This actually ignores the multinomial sampling model and focuses on its
binomial margins, which is a reasonable approximation when the number
nrow(x) is large and substantially improves computation speed.
#The first two rows present different counts across columns#The last two columns do notx <- matrix(c(70,10,10,10,35,35,10,10),ncol=2)
x
rowLogRegLRT(x)