Imputation delta

scripts/imputation_delta.py compares imputed_multilateral_by_purpose totals between an earlier oda-data release and the installed code, by year and donor, and splits the difference into named causes.

uv run python scripts/imputation_delta.py \
    --cache-dir /path/outside/repo/oda-data-cache \
    --output-dir /path/outside/repo/imputation-delta-out \
    --years 2015-2024

The earlier release (--old-version, default 2.7.0) runs in an isolated environment through uv run --isolated, against the same cache as the current code. Both directories must be outside the repository. --providers, --measure, --currency and --base-year set the comparison, --refresh re-downloads the current code's inputs, --skip-channel-remapping skips the channel diagnostic, and --dry-run prints the plan without reading data.

Output

imputation_delta.csv has one row per year and donor:

Column	Meaning
old_total	The earlier release's total.
new_total	The current total.
observed_delta	new_total - old_total.
bi_multi_exclusion	Current code with bi_multi == 2 rows excluded from the shares, minus current code with them kept.
share_basis	Current code with flow_types=("ODA", "OOF") minus current code with ("ODA",).
window_padding	Current code with max_share_age=5 minus current code with 0, plus, in years the earlier release left empty, the rest of observed_delta after unallocated_and_proxy.
unallocated_and_proxy	The value of current rows with allocation_status unallocated or proxy.
residual	observed_delta minus the four columns above.

channel_remapping.csv lists each (provider_code, agency_code) pair with its channel under both versions and the CRS money at stake. summary.json holds the grand totals.

Reading the table

Current output conserves core contributions. The rows for each year, donor and channel sum to the donor's MultiSystem core contribution, so a donor's total depends on its core contributions alone. The CRS shares only split that total across purposes and recipients. bi_multi_exclusion, share_basis and the max_share_age part of window_padding compare two current runs over the same contributions, so they are zero. Their effect falls on the purpose split, which this table does not show.

Totals move for three reasons. Releases before 2.8.0 return no rows for the first period_length - 1 years of a multi-year request, and the script puts that gap in window_padding. Earlier releases also dropped core contributions to channels without a spending share, which the current code reports as proxy or unallocated. Differences in channel assignment and share basis make up the rest and land in residual, along with stale_share money, which earlier releases also dropped. A negative residual means the current code reports money as unallocated or proxy that the earlier release allocated.