export srs_diff_est

loo_subsample() can sometimes be over-complex. srs_diff_est() is a non-exported internal function used by loo_subsample. srs_diff_est() implements difference estimator Equations 7-9 of Magnusson et al., 2020. It would be useful to export this function and provide user targeted function with documentation.

Sounds good

Hello,

would this be a good first issue/contribution?
I.e. roxygenise, #‘ @export respectively, this:

and also add 1 or 2 examples for users as well as the Magnusson et al. (2020) citation?

Kind regards

Sounds good to me too.

@vinniott

@vinniott thanks for your interest. Yes I think this would be a good first issue and what you described adding sounds good. We also probably want a bit more detail in the documentation when moving this to an exported user-facing function. For example, we should mention what type/shape the inputs should be and what the contents of the returned list are.

@jgabry

@jgabry thank you for elaborating.
If you want, then I could give it a try and update you here in a week?

Great, thank you!

Hi again,

while I have made progress, I think it might be better if I look into stan-dev/bayesplot now and return to loo later on.
Again, thank you for explaining this issue to me.

Maybe you could tag this as "good-first-issue" now so that others can find it easier?

In case this is helpful for someone, here is my initial roxygen2 skeleton for the export:
master...vinniott:loo:export-srs-diff-est

@vinniott

Thanks @vinniott for the skeleton. I think this issue requires also contribution from me. I can write the technical description I'd like to have included in the documentation.

Here is my proposoal for the documentation

#' Difference estimator with simple random sampling without replacement.
#'
#' The difference estimator `srs_diff()` estimates
#' the expectation $nE[y]$ when we have $n$ approximate values $\tilde{y}_i$,
#' $i=1,\ldots,n$ and $m<n$ accurate values $y_j$, $j \in
#' \mathcal{S}$, where $m$ is the subsample size and $\mathcal{S}$ is
#' a simple random subsample without replacement.  The original
#' approach is by Cochran (1977) and we follow the equations 7--9 by
#' Magnusson et al. (2020).
#' 
#' @param y_approx (numeric) `n` approximated values.
#' @param y (numeric) `m<n` subsampled values.
#' @param y_idx (integerish) The index of `y` in `y_approx`.
#' @return A named list containing numeric values:
#' * `y_hat`: estimated mean of $y$ (Eq 7),
#' * `v_y_hat`: variance of the mean estimate (Eq 8), and
#' * `hat_v_y`: estimated variance of $y$ (Eq 9).
#' 
#' @details Magnusson et al. (2020) Eq (9) first row second `+` should
#'   be `-`. Supplementary material Eq (6) has this correct.
#'   As `srs_diff_est()` in `loo` package is used for $nE[y]$, there is 
#'   a proportional difference of $1/n$ compared to the paper.

As srs_diff_est() has been used for elpd_loo which is in sum scale, srs_diff_est() is actually estimating $nE[y]$, while the equations in the paper are for $E[y]$. The question whether we should change the exposed function to estimate $E[y]$ and then adjust other code in loo_subsample.R

added documentation

Thank you for the technical documentation! I have incorporated it and made a PR #340 .
However, there is no @example yet because I am not fully familiar with the package as a whole.

added documentation

Navigation Menu

Repository navigation

export srs_diff_est #333

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

export srs_diff_est #333

Description

Activity

jgabry commented on Mar 2, 2026

vinniott commented on Mar 11, 2026

MansMeg commented on Mar 11, 2026

jgabry commented on Mar 11, 2026

vinniott commented on Mar 12, 2026

jgabry commented on Mar 12, 2026

vinniott commented on Mar 19, 2026

avehtari commented on Mar 20, 2026

avehtari commented on Mar 20, 2026

vinniott commented on Mar 22, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions