Self-attribution from model

[domenico-simone]

Hi,

is there a way to get P of assignments of isolates to their own clusters? I.e., self-attribution results to assess the model reliability?

Thank you,

Domenico

[jmarshallnz]

If what you're after is P(Source | ST) from the 'genotype distribution' part of the model (asymmetric island or dirichlet) then you can do this by using the likelihood returned by st_fit() along with a prior. st_fit() for the AI model returns a leave one out estimate of log(P(ST | Source)). So you can convert that to P(Source | ST) via Bayes Thm.,

summary() gives you the posterior mean of each source given ST assuming apriori each source equally likely. You could adapt the code there to give you the whole posterior. One thing to watch is in the exponentiation of logP - it pays to do it by first factoring out the largest value so you don't lose precision.

Code:

library(islandR)
library(tidyverse)
st = st_fit(formula = Source ~ ST,
            non_primary = "Human",
            data = manawatu,
            method="island",
            sequences = ~ ASP + GLN + GLT + GLY + PGM + TKT + UNC)

prior <- tibble(source = setdiff(unique(manawatu$Source), "Human"),
                prior = 1/4)

exp_sum_log <- function(x) { y = x - max(x); exp(y)/sum(exp(y)) }

st |> as.data.frame() |>
  left_join(prior) |>
  mutate(log_pp = log(prior) + log_p) |>
  group_by(type, iteration) |>
  mutate(posterior = exp_sum_log(log_pp)) |>
  select(type, iteration, posterior)

[jmarshallnz]

Note that this is not including the 'attribution' part of the model, which infers the mixing (here the fixed prior) of the sources.

Other SA models may be utilising that 'inferred mixing' when doing self attribution. This might be particularly important for imbalanced sources, though you may be able to incorporate that via the prior above.

[jmarshallnz]

See 22e3215