# 06 — Minimarket proximity: were KDMP built on top of existing retail? **Reproduce**: `KOPDES_RAW=data/snapshots/2026-08-13 python reports/06-minimarket-proximity/run.py` · No network · **Last run**: 2026-08-15 **Source**: `data/osm/indonesia_minimarkets.gpkg` (10,580 POIs) + Kontur population + the road cell index from [05](../05-road-access/), with KDMP coordinates from the **2026-08-13** snapshot. Hashes in [`_source.json`](_source.json). > **This report is built around two known problems in its own data**: the source > is incomplete and urban-biased (§2), and it is not actually a minimarket > dataset until it is re-tiered (§1). Read both before quoting any number. Every > forward figure is a **lower bound**, never a point estimate. ## 1. What is actually in this dataset The file is named `minimarkets` but it is an Overpass pull of `shop=convenience|supermarket|department_store`. Before measuring anything, it has to be split into formats that a village cooperative actually competes with ([`retail_tiers.csv`](retail_tiers.csv)): | Tier | POIs | Share | Used for proximity? | | ------------------------------------ | --------- | --------- | ------------------- | | 0 traditional warung / toko | 858 | 8.1% | no — see §5 | | **1 convenience / minimarket** | **7,617** | **72.0%** | **yes, primary** | | 2 supermarket (town-level) | 1,838 | 17.4% | reported separately | | 3 department store (not food retail) | 267 | 2.5% | excluded | **28% of the file is not a minimarket.** A Matahari or a Ramayana is not competition for a village cooperative selling rice and LPG, and a Hypermart serves a town, not a desa. An earlier version of this report measured against all 10,580 POIs and consequently **overstated every forward figure by roughly a quarter** (e.g. 3.22% within 500 m, against 2.48% correctly tiered). `run.py` also repairs `scripts/download_osm.py`'s `classify_brand()`, which misses spelling variants and regional chains and dumps them in `other`: 76 Alfamart-group (`Alfa Express`, `Alfa Midi`), 54 Bali chains (Pepito, Coco), 33 Yomart, 17 Hero, 10 7-Eleven. ## 2. The coverage problem, measured OSM POI coverage in rural Indonesia is poor **and urban-biased** — mapper density tracks urbanisation, the opposite of where KDMP sit. A naive point estimate from this source would undercount by an unknown margin, in the direction that **exonerates the programme**. **Deficit against published outlet counts** ([`osm_brand_coverage.csv`](osm_brand_coverage.csv)): | Brand | In OSM | Published outlets | OSM share | | --------- | ------ | ----------------- | --------- | | Indomaret | 3,030 | ~22,000 | **13.8%** | | Alfamart | 2,181 | ~20,000 | **10.9%** | > Published totals are hardcoded in `run.py` as `PUBLISHED_OUTLETS`, from > company reporting. **Verify and cite before publication.** No downstream > number depends on them — they size the deficit statement only. **Urban bias, measured from the data itself**, needing no external benchmark ([`urban_bias_diagnostic.csv`](urban_bias_diagnostic.csv)): | | Median population in own 400 m cell | Share in a zero-population cell | | ----------------------- | ----------------------------------- | ------------------------------- | | KDMP | 261 | 21.3% | | OSM minimarket (tier 1) | **3,786** | **0.1%** | Mapped minimarkets sit in cells ~15× more populous than cooperatives do, and essentially never in an empty one. Provincial density says the same ([`osm_density_by_province.csv`](osm_density_by_province.csv)): 365 minimarkets per 100 KDMP in DKI Jakarta, 0.15 in Sulawesi Barat, **0 in Papua Tengah**. ## 3. Forward direction — lower bounds only **At least** this many KDMP have a mapped tier-1 minimarket nearby ([`kdmp_near_minimarket_lower_bounds.csv`](kdmp_near_minimarket_lower_bounds.csv)): | Within | Cooperatives (≥) | Share (≥) | | ------ | ---------------- | --------- | | ~500 m | 2,068 | **2.5%** | | ~1 km | 4,150 | 5.0% | | ~2 km | 7,425 | 8.9% | | ~5 km | 16,505 | 19.8% | Supermarkets separately ([`kdmp_near_supermarket_lower_bounds.csv`](kdmp_near_supermarket_lower_bounds.csv)): 1.0% within ~500 m, 2.6% within ~1 km, 12.6% within ~5 km. Restricting to the 27 provinces where OSM retail density is at least a quarter of the national median barely moves it — 2.9% / 5.8% / 10.4% / 23.1% ([`scope_restricted_lower_bounds.csv`](scope_restricted_lower_bounds.csv)). **These are floors and the true values are materially higher.** Write "at least 2.5%", never "only 2.5%". ## 4. Reverse direction, and the null models that make it interpretable Turning the question around removes most of the coverage bias, because the statistic is conditional on the store existing ([`minimarket_near_kdmp.csv`](minimarket_near_kdmp.csv)): **78.3% of mapped minimarkets have a KDMP within ~1 km**, 95.7% within ~2 km. On its own that proves nothing — KDMP are one-per-village and cover 95% of the population within ~1.4 km ([03](../03-population-coverage/)), so _any_ populated point has one nearby almost by construction. Three controls, each matched to the minimarket count ([`null_model_comparison.csv`](null_model_comparison.csv)): - **Population-weighted random** — a location sampled proportional to Kontur population. - **Road-constrained population-weighted random** — a location on a non-track road cell inside a populated area, i.e. a plausible retail site. - **Road-constrained, population-_matched_** — the same road-cell pool, but sampled to reproduce the minimarkets' own population distribution. **This is the primary control.** ### What this design is called It was built from the problem rather than from a textbook, so the names are recorded here for anyone who wants to place it in a literature. | Component | Standard term | | --------------------------------------- | --------------------------------------------------------------------------------- | | The comparison itself | **Null model**; formally an _inhomogeneous_ null, not Complete Spatial Randomness | | The statistic | **Cross-type nearest-neighbour distribution function** `G₁₂(r)`, at four radii | | The question being tested | **Independence null hypothesis** for a bivariate (two-type) point pattern | | Matching the null's density to the POIs | **Frequency matching** on a confounder, in deciles | | The 40 redraws and their range | **Monte Carlo envelope** | Read as a whole it is a **case-control point pattern analysis with frequency-matched controls**: minimarkets are the cases, the road-cell draws are the controls, and population density is the confounder being matched out. The closest literature is distance-based localization testing in economic geography, where the control set is drawn from **feasible sites** (Duranton & Overman 2005, _Testing for Localization Using Micro-Geographic Data_); the road-cell pool is that idea. Diggle's _Statistical Analysis of Spatial Point Patterns_ is the reference for the bivariate framing, Besag & Diggle (1977) for the Monte Carlo test. **Do not claim more than was run.** This is a simplified version of all of the above: four radii rather than a full `G₁₂` curve, no p-value, an envelope from 40 draws rather than the conventional 999, matching on one covariate, and no edge correction. Describe the design, do not name-drop the test. ### Why the third null exists (corrected 2026-08-15) The road+pop null was described in this README as answering "if minimarkets and KDMP both simply sit on the village road, that alone would manufacture apparent co-location." It only half answered it. Measured on the same grid ([`null_density_match.csv`](null_density_match.csv)): | Points | Median population in own 400 m cell | | ----------------------------- | ----------------------------------- | | Minimarket (tier 1) | 3,786 | | Null: road + pop-**weighted** | **2,374** | | Null: road + pop-**matched** | 3,786 | Weighting by population is not the same as matching it. The old null stood in for the minimarkets in places roughly a third less dense, and since KDMP coverage rises with density, that gap was scored as co-location. Matching the distribution removes it, and takes the ~500 m excess from **+9.6 to +6.7 pts**. | Within | Minimarkets | Pop null | Road+pop null | **Road+pop-matched** | **Excess (primary)** | | ------ | ----------- | -------- | ------------- | -------------------- | -------------------- | | ~500 m | 43.8% | 27.3% | 34.7% | 37.1% | **+6.7 pts** | | ~1 km | 78.3% | 62.2% | 70.9% | 74.6% | **+3.7 pts** | | ~2 km | 95.7% | 87.0% | 92.9% | 94.2% | +1.5 pts | | ~5 km | 99.6% | 98.2% | 99.4% | 99.4% | +0.2 pts | The residual is real but modest, and it **decays with distance** (+6.7 → +3.7 → +1.5 → +0.2) — the signature of genuine short-range catchment overlap rather than a global artefact. ### The nulls are draws, not constants Two further corrections from the same pass: - **The published figure was one sample.** `null_stability.csv` redraws each null 40 times: the ~500 m excess spans **5.7–7.7** for the matched null and **8.2–10.4** for the old road null. Quote the mean and the range, never a single draw to one decimal. The figures the site prints come from [`null_excess_published.csv`](null_excess_published.csv) for exactly this reason; the ~2 km and ~5 km rows there are still single draws, and say so. - **The fixed seed never made this reproducible.** The nulls draw integer indices into DuckDB result sets, and DuckDB does not guarantee row order, so re-running with no code change to the null moved the ~500 m road null from 34.41% to 33.79%. The candidate queries now carry `order by`, and two consecutive runs are byte-identical. An earlier robustness note claimed these figures were "essentially unchanged by the §1 re-tiering (previously +9.4 / +6.7 / +2.3 / +0.1 on the untiered set)". That comparison is withdrawn: the untiered run exists nowhere in the repo, and the numbers it quoted are indistinguishable from ordinary redraw noise. ### What this does and does not establish **Does**: KDMP and existing modern retail overlap at short range more than chance placement on the road network would produce. **Does not**: establish intent, or cannibalisation. Both formats plausibly target the same village focal point — market, junction, balai desa — a sub-cell feature no null at this resolution can control for. And cannibalisation is a claim about _trade_, which needs the transaction data that is 97% zero ([02](../02-zero-inflation/)). Proximity is a precondition for competition, not evidence of it. ## 5. The limitation that matters most: the real incumbent is invisible The competitor a village cooperative actually displaces is the **warung / toko kelontong**, not an Alfamart. OSM has **858 of them nationally** — against a true population in the millions (BPS counts micro and small retail units in the millions, not thousands). So "cannibalisation of existing retail", properly framed, is **largely untestable with this data**. What §3 and §4 measure is one visible slice — branded modern retail — of a much larger informal-retail question that OSM cannot reach. This belongs in the write-up as a **scope statement**, not as a buried caveat: otherwise the paper implicitly claims to have tested competition when it has tested a minority of it. Sourcing traditional retail would need a different instrument entirely — BPS PODES village-facility counts, or field survey. ## Caveats - Every forward number is a lower bound of unknown tightness (§2). - Ring distance is a band: adjacent r10 cell centres are ~132 m apart, so `k × 0.132 km` is approximate and degrades slightly with latitude. - Province is assigned to each POI from its **nearest KDMP**, not a boundary polygon — adequate for a density diagnostic, not for exact attribution. - The nulls are population- and road-weighted but otherwise uniform; they do not model the commercial siting logic a retail chain actually uses, so the residual excess is an **upper bound** on deliberate co-location. - Tier assignment leans on the `shop` tag and brand names; the 3,450 unbranded `other` POIs are assigned by `shop` tag alone. - OSM extract is a Geofabrik/Overpass snapshot from 2026-08-07. ## Outputs | File | Contents | | ---------------------------------------------------------------------------------- | ------------------------------------------------ | | [`retail_tiers.csv`](retail_tiers.csv) | what the dataset actually contains | | [`osm_brand_coverage.csv`](osm_brand_coverage.csv) | OSM counts vs published outlets | | [`urban_bias_diagnostic.csv`](urban_bias_diagnostic.csv) | population context, KDMP vs minimarket | | [`osm_density_by_province.csv`](osm_density_by_province.csv) | minimarkets per 100 KDMP | | [`kdmp_near_minimarket_lower_bounds.csv`](kdmp_near_minimarket_lower_bounds.csv) | forward, tier 1 | | [`kdmp_near_supermarket_lower_bounds.csv`](kdmp_near_supermarket_lower_bounds.csv) | forward, tier 2 | | [`minimarket_near_kdmp.csv`](minimarket_near_kdmp.csv) | reverse direction | | [`null_model_comparison.csv`](null_model_comparison.csv) | both nulls | | [`scope_restricted_lower_bounds.csv`](scope_restricted_lower_bounds.csv) | 27-province subset | | `kopdes_minimarket_distance.csv` | per-cooperative (gitignored, rebuilds in ~1 min) |