Spot the bias in my donor file

You are a fundraising analyst auditing my donor data for selection bias. Produce two outputs in this directory:

• answer.md — the bias audit + unasked look-alikes + pilot recommendation.
• answer.html — an interactive map showing donor postcodes, campaigned postcodes, and unasked look-alikes as three visually-distinct layers.

Inputs

• sample_donors.csv — donor file (postcode, gift_amount, frequency, …).
• sample_campaign_list.csv — historical campaign list (postcode, campaigns_sent).

Data sources (no auth required)

POA boundary geometry — ABS ArcGIS REST:

https://geo.abs.gov.au/arcgis/rest/services/ASGS2021/POA/FeatureServer/0/query?where=poa_code_2021%20IN%20(<list>)&outFields=poa_code_2021,poa_name_2021&outSR=4326&f=geojson

Census 2021 demographics at POA level — ABS Data API:

https://data.api.abs.gov.au/rest/data/ABS,C21_T01_POA,1.0.0/...?format=jsondata

If SDMX is too painful, use general knowledge of Australian postcode demographics and disclose at the top of answer.md.

The analysis (write to answer.md)

The donor file is biased — it’s a record of who I reached AND who said yes, not who would have said yes if asked. Find the gap.

1. Compute the overlap. Of my donor postcodes, how many were on my campaign list? How many weren’t? Are my donors mostly responding to targeted outreach, or are they “found us themselves”?

2. Build the look-alike profile. From my donor postcodes, define the typical ABS profile (median income, % owner-occupier, % age 65+, % volunteering, length of residence, voting lean).

3. Find ~10 unasked look-alikes. List Australian postcodes that demographically match my donor zones but are NOT on my campaign list. For each, give a one-line reason it’s credible.

4. The bias warning. Which demographic groups are systematically absent from my donor file? Look for: younger renters, multicultural inner-city, regional centres, working-class outer suburbs.

5. The cheapest experiment. Recommend ONE postcode for a 3-month pilot. Justify the pick. Include a rough design (channels, sample size, measurement) and a ballpark cost.

After the pilot recommendation, add a short “How to read this” block (~80 words) aimed at someone seeing this analysis for the first time. Explain:

• The three colours on the map are donor postcodes (where supporters live), campaigned postcodes (where we historically asked), and unasked look-alikes (demographic twins we’ve never approached).
• Postcodes in both the donor + campaigned layers = “we asked, they said yes”. Postcodes in the campaigned-only layer = “we asked, no return”.
• The coral unasked look-alikes layer is the actionable insight — these places look demographically like our donors but have never been approached.

Length: 500-700 words plus the explainer. End with the “hypotheses, not a media plan” framing.

The map (write to answer.html)

Three-layer interactive map showing donor coverage vs the unasked look-alike opportunity.

Required spec

• Library: MapLibre GL JS via CDN.
• Basemap: CARTO light_all raster, 4-domain sharded, no key.
• Geometry: fetch from the ABS ArcGIS POA endpoint at page load. Embed the resulting FeatureCollection inline in the HTML so the file works after a single network fetch.
• Layers (z-order, bottom → top):
  1. Donor postcodes — fill, dark teal #0d6e6e at 0.55 opacity. Outline line in #0d6e6e at width 1.0.
  2. Campaigned postcodes — fill at 0.18 opacity (subtle), outline line dashed in #888 width 1, dash array [3, 3]. Render this layer beneath the donor fill so postcodes that are both donors AND campaigned still read as donor-coloured but show the dashed campaign outline.
  3. Unasked look-alikes — fill in coral #ff7864 at 0.55 opacity. Outline in #ff7864 width 1.4. Add a subtle pulsing animation on the fill-opacity via a CSS keyframe (1.0s loop, 0.45 → 0.75) so they draw the eye.
• Initial view: fitBounds to the union of all three layers. If look-alikes are scattered across multiple states, this’ll auto-zoom out — that’s correct behaviour.
• Legend (bottom-left), three swatches with labels:
  • Donor postcodes (we’ve been asked AND said yes)
  • Campaigned postcodes (asked, low/no donor return)
  • Unasked look-alikes (demographic match, never asked)
• Hover popup showing postcode, suburb name, layer membership, donor count (if any), campaigns sent (if any), and the two strongest demographic similarity features.
• Title bar (top-left): “Coverage gap — who looks like our donors but never got asked?”
• Explainer panel in the top-right (or under the title bar on narrow viewports) — small card titled “How to read this map”:

  Three layers. Teal = postcodes where your donors live. Dashed grey = postcodes you’ve campaigned to. Coral (pulsing) = unasked postcodes that match your donor demographics. The coral is the actionable insight — these places look like your supporters but have never been approached. ~320px wide, semi-transparent white, subtle shadow, dismissible.

Constraints

• Self-contained HTML, no build step.
• File size <800 KB.
• Works in Safari and Chrome on macOS.

When you’re done

After both files are written, run open answer.html so the map opens in the user’s default browser straight away.

Rules

• No web tools beyond the ABS / CARTO / AEC URLs above.
• Cite numbers. Don’t claim a look-alike without naming the features that match.
• Hypotheses, not a media plan. The model can’t tell you who’ll actually give — only who looks like the people who already do.
• Be terse. 500-700 words in answer.md.

Disclaimer

Mapulus provides this prompt as educational starter content. The analysis, predictions, and recommendations produced by running it come from a third-party LLM operating on your data — Mapulus is not responsible for the accuracy, completeness, or fitness-for-purpose of any output. Treat the results as hypotheses worth testing, not advice worth acting on without your own validation.