# 02 — Zero inflation: there is almost no outcome variable

**Reproduce**: `KOPDES_RAW=data/snapshots/2026-08-13 python reports/02-zero-inflation/run.py` ·
No network, deterministic · **Last run**: 2026-08-13
**Source**: `kopdes_stats_village.csv` from the **2026-08-13** snapshot (**83,069
villages after deduplication**). Hashes in [`_source.json`](_source.json). Running
without `KOPDES_RAW` reads the 08-05 export instead and every figure below moves.

> **Corrected 2026-08-09.** The 2026-08-05 export contains **1,555 duplicate
> `village_id`s**, found when the 2026-08-09 snapshot came back 1,555 rows
> shorter with an identical id set. Left in, they double-count: the
> village-level transaction total came out IDR 18.8M above the province-level
> total. All figures below are now deduplicated. The headline barely moved
> (97.02% → 96.97% zero; top-100 concentration unchanged at 37.31%), but the
> denominators did.

## Finding 1 — activity fields are ~97% zero; administrative fields are ~96% filled

[`zero_inflation_by_field.csv`](zero_inflation_by_field.csv)

| Field                   | Kind           | Non-zero  |
| ----------------------- | -------------- | --------- |
| `npwp_count`            | administrative | **97.2%** |
| `accounts_count`        | administrative | **95.8%** |
| `nib_count`             | administrative | 73.1%     |
| `savings_total_amount`  | activity       | 12.5%     |
| `simpanan_pokok_amount` | activity       | 11.9%     |
| `simpanan_wajib_amount` | activity       | 9.2%      |
| `transaction_value`     | activity       | **3.3%**  |
| `transaction_volume`    | activity       | **3.3%**  |

The split is the interesting part. Fields written **once at registration** are
nearly complete. Fields that require **ongoing operational reporting** are
nearly empty. Conditioning on the village having an account doesn't change it —
of the 79,579 villages that have one, 98.9% have an NPWP but only 3.4% have ever
recorded a transaction ([`zero_inflation_given_account.csv`](zero_inflation_given_account.csv)).

**This is consistent with both explanations** and does not settle the question:

- _Genuine inactivity_: registration completed, operations never started.
- _Reporting lag_: registration is a one-off form that got filled; transaction
  reporting needs a working POS/reporting channel that isn't live yet.

See [01-snapshot-drift](../01-snapshot-drift/) for the only test that can
separate them, and its limits.

## Finding 2 — the consequence for the analytics plan

`transaction_value` **cannot be used as a continuous outcome**. Kruskal-Wallis
or ANOVA across accessibility bands — the method
[`analytics-plan.md`](../../analytics-plan.md) specifies for A4, and the basis
of H1 and H2 — is comparing two distributions that are each 97% ties at zero.

Workable reframings, in order of preference:

1. **Binary**: `any_activity = transaction_value > 0`, 2,726 positives out of
   83,069. Rare-event logistic regression with province fixed effects.
2. **Savings as secondary outcome**: 12.5% non-zero is far less degenerate, and
   `simpanan_pokok` vs `simpanan_wajib` is a real dormancy signal.
3. **Conditional magnitude**: among the 2,726 active villages only, model the
   amount. Small n, but it answers "where activity happens, what shapes it".

## Finding 3 — activity is extraordinarily concentrated

[`transaction_concentration.csv`](transaction_concentration.csv)

| Top N villages       | Share of all national transaction value |
| -------------------- | --------------------------------------- |
| 10                   | 11.8%                                   |
| 50                   | 24.4%                                   |
| **100**              | **34.8%**                               |
| 500                  | 73.5%                                   |
| 1,000                | 90.6%                                   |
| 2,726 (all non-zero) | 100%                                    |

**100 villages out of 83,069 — 0.12% — carry 34.8% of the program's entire
reported economic output.** A thousand carry 90.6%. (The shares eased from 37.3%/92.9% on 05-08 as more villages began reporting on 08-13 — concentration is still extreme, but it is no longer the whole story.)

This is the finding that should lead the investigation. Whatever the zeros
turn out to mean, the program's _reported_ output is not "low across the
board"; it is absent everywhere except a few hundred sites. The research
question that follows is what those sites have that the other 82,000 don't —
and that is a question the placement data can actually answer.

## Finding 4 — the province gradient runs the wrong way for a simple story

[`zero_inflation_by_province.csv`](zero_inflation_by_province.csv). Non-zero
transaction rate ranges from **17.2% (DKI Jakarta)** and 16.5% (Banten) down to
**0.0%** (Kepulauan Riau) and ~0.04–0.1% in the Papua provinces. Median province: 1.4%.

Note this gradient tracks urbanisation and connectivity — which is exactly what
_both_ explanations predict (richer, better-connected places both transact more
**and** report better). It is not evidence for either. Don't cite it as such.

## Caveats

- Village rows total 84,624 against 83,382 cooperatives; 84,291 villages have
  exactly one cooperative, so village ≈ cooperative but not exactly.
- These village rows cannot be joined to coordinates for ~21% of cooperatives —
  see §1.5 of [`analytics-plan-review.md`](../../analytics-plan-review.md).