From 65f8fe48d57e65d181a66e49840188784140c096 Mon Sep 17 00:00:00 2001
From: Victor Lin <13424970+victorlin@users.noreply.github.com>
Date: Thu, 15 Aug 2024 14:57:14 -0700
Subject: [PATCH] Add sections on sampling methods

Reword the subsampling introduction with *what* it is, followed by
examples on *why* paired with *how*.

This also allows future sampling methods such as weighted sampling to be
added by simply including a new section.
---
 .../filtering-and-subsampling.rst             | 66 +++++++++++++++++--
 1 file changed, 61 insertions(+), 5 deletions(-)

diff --git a/src/guides/bioinformatics/filtering-and-subsampling.rst b/src/guides/bioinformatics/filtering-and-subsampling.rst
index cfdbefab..40afd2a9 100644
--- a/src/guides/bioinformatics/filtering-and-subsampling.rst
+++ b/src/guides/bioinformatics/filtering-and-subsampling.rst
@@ -114,11 +114,56 @@ Subsampling within ``augur filter``
    Subsampling is applied after all standard filter options and before
    force-inclusive filter options.
 
-Another common filtering operation is subsetting of data to achieve a more
-even spatio-temporal distribution or to cut-down data set size to more
-manageable numbers. The filter command allows you to partition the data into
-groups based on column values and sample uniformly. For example, target one
-sequence per month from each country:
+Another common filtering operation is **subsampling**: selection of data using
+rules based on output size rather than individual sequence attributes. These are
+the sampling methods supported by ``augur filter`` and a final section for caveats:
+
+.. contents::
+   :local:
+
+Random sampling
+---------------
+
+The simplest scenario is a reduction of dataset size to more manageable numbers.
+For example, limit the output to 100 sequences:
+
+.. code-block:: bash
+
+   augur filter \
+     --sequences data/sequences.fasta \
+     --metadata data/metadata.tsv \
+     --min-date 2012 \
+     --exclude exclude.txt \
+     --subsample-max-sequences 100 \
+     --output-sequences subsampled_sequences.fasta \
+     --output-metadata subsampled_metadata.tsv
+
+Random sampling is easy to define but can expose sampling bias in some datasets.
+Consider uniform sampling to reduce sampling bias.
+
+Uniform sampling
+----------------
+
+``--group-by`` allows you to partition the data into groups based on column
+values and sample uniformly. For example, sample evenly across countries over
+time:
+
+.. code-block:: bash
+
+   augur filter \
+     --sequences data/sequences.fasta \
+     --metadata data/metadata.tsv \
+     --min-date 2012 \
+     --exclude exclude.txt \
+     --group-by country year month \
+     --subsample-max-sequences 100 \
+     --output-sequences subsampled_sequences.fasta \
+     --output-metadata subsampled_metadata.tsv
+
+An alternative to ``--subsample-max-sequences`` is ``--sequences-per-group``.
+This is useful if you care less about total sample size and more about having
+a fixed number of sequences from each group. For example, target one sequence
+per month from each country:
 
 .. code-block:: bash
 
@@ -132,6 +177,17 @@ sequence per month from each country:
      --output-sequences subsampled_sequences.fasta \
      --output-metadata subsampled_metadata.tsv
 
+Caveats
+-------
+
+For these sampling methods, the number of targeted sequences per group does not
+take into account the actual number of sequences available in the input data.
+For example, consider a dataset with 200 sequences available from 2023 and 100
+sequences available from 2024. ``--group-by year --subsample-max-sequences 300``
+is equivalent to ``--group-by year --sequences-per-group 150``. This will take
+150 sequences from 2023 and all 100 sequences from 2024 for a total of 250
+sequences, which is less than the target of 300.
+
 Subsampling using multiple ``augur filter`` commands
 ====================================================