Ethical Mapping
Last updated: November 14, 2026
Instructions
We use this skill on every map we publish. A map is a version of reality built from choices: what to map, at what geography, with which breaks, colours and projection, and whether to map at all (Datawrapper, “What to consider when creating choropleth maps”, Lisa Charlotte Muth, 21 Mar 2018). Each choice can mislead a reader, and one of them, putting a vulnerable person on a map, can get that person hurt. This skill makes those choices explicit and checkable.
It composes three existing skills and restates none of them:
infographic-builderowns the visual production pipeline (briefs, templates, rendering,
upload). Any map that ships inside an infographic, and any Datawrapper map built in Compose, passes this skill’s checklist (§5) first.
geolocation-verification-specialistowns precision. Where a location was inferred
rather than recorded, its “consistent with” language and (Estimated) label apply, and a published point is never more precise than what was verified. For a vulnerable person, the published precision is none: §3.2.
patriot-private-citizen-inclusion-gateruns first for any private person a map would
name or locate. Even where the gate passes, its proportionality limits hold: no home address or workplace that is not already on the public record of the anchor conduct.
Method, in order:
- Ask whether it should be a map (§1).
- Open a map register entry (§2) and fill it as you go.
- Apply the rules in §3: normalise, protect, classify, colour, cite the boundaries, choose
the projection, show the uncertainty.
- Build it with the tools in §4.
- Run the checklist (§5) before the map goes to an editor.
Use a different skill when:
- the task is verifying where a photo or video was taken:
geolocation-verification-specialist; - the task is checking the arithmetic of a rate, a percentage or a margin of error that a map
shows: numbers-and-statistics-integrity checks the number, this skill checks how it is drawn;
- the task is deciding whether a named private person may appear in our coverage at all:
patriot-private-citizen-inclusion-gate;
- the task is a network of relationships, not places:
network-analysis-specialist.
1. Should this be a map at all?
Location data can be mapped; that does not mean it should be. Ask first whether location matters to the story. Differences between areas often read better as a bar chart, change over time as a line chart, two variables per area as a scatter chart, and sometimes a table or a sentence beats all of them (Hands-On Data Visualization, “Map Design Principles”, Dougherty and Ilyankou).
A choropleth (regions filled with colour) works when:
- the data shows a regional pattern, or the map covers places where our readers live;
- it shows one variable (which may be a difference, such as the change in a rate);
- the point is the big picture, not small differences between regions, which readers cannot
judge from colour. If exact differences between regions matter, use a table or a chart, or add tooltips with the values (Datawrapper, “What to consider when creating choropleth maps”).
2. The map register
One entry per map. It is a working record for the editor, not reader-facing, and it is what lets a second person rebuild the map.
| Field | What to write |
|---|---|
| Question | The one thing the map answers |
| Geography | State, county, tract, congressional district; why this level (§3.7) |
| Boundary file | Publisher, file name, vintage (§3.5); or the Datawrapper basemap id |
| Numerator | Dataset, publisher, table or query, period, retrieval date |
| Denominator | Population or other base, publisher, vintage; must match the numerator’s period |
| Unit | “per 100,000 residents”, “% of registered voters”, etc. |
| Small-number handling | Which areas were suppressed or flagged, and the threshold (§3.1, §3.2) |
| Privacy level | Highest precision shown; why it is safe (§3.2) |
| Classification | Method, number of classes, break values (§3.3) |
| Palette | Name and type (sequential or diverging; centre value if diverging) (§3.4) |
| Projection | Name (§3.6) |
| Uncertainty | Margins of error, where shown (§3.7) |
| Checklist | §5, each item ticked with initials |
3. The rules
3.1 Normalise: rates, not raw counts
- A choropleth shows a rate. A raw count across places of unequal population mostly maps
where people live. Map the unemployment rate, not the number of unemployed people (Datawrapper, “What to consider when creating choropleth maps”); normalise counts by population, or another fitting base, before colouring regions (Hands-On Data Visualization, “Normalize Choropleth Map Data”: its paired Covid-19 maps show raw counts pointing at the most populous states, while per-capita rates show the real epicentre).
- To show a count, use a symbol map (sized circles at region centroids), not a filled
region, and say that it is a count. Even then, a count map often answers only “where do most people live?” (Datawrapper, same source).
- Pick the base that fits the question. Residents for arrests, registered voters for
turnout problems, the workforce for layoffs. Name it in the legend or note.
- The denominator is from the same period and geography as the numerator, and its vintage
is recorded. The numbers checks (per-capita, small counts, base rates) are numbers-and-statistics-integrity‘s; run them on the rate before it is mapped.
- Small counts make unstable rates. A county with three events can top a rate map by chance.
Show the count beside the rate in the tooltip, and suppress or hatch areas whose counts are too small to rank (see §3.2 for the privacy floor).
3.2 No point data for vulnerable people
**The hard rule: we never publish the location of a vulnerable person, or of a place whose confidentiality protects them, as a point, a pin, a street, a block or any geography small enough to find them.** We aggregate to a safe geography instead.
Who and what this covers:
- the homes and workplaces of immigrants, including people arrested, detained or deported, and
the addresses where arrests took place when those are homes;
- confidential shelters. Federal law already bars making public “the address or location of
any shelter facility assisted under this chapter that otherwise maintains a confidential location”, except with written authorisation from those who run it (42 U.S.C. § 10406(c)(5)(H), Family Violence Prevention and Services Act). We hold every confidential shelter to that standard, federally funded or not;
- people seeking or providing abortion care, and their homes;
- victims and witnesses of crime, survivors of domestic or sexual violence, and minors;
- any private person, unless the inclusion gate has passed and the location is already on the
public record of the anchor conduct;
- the real-time location of any individual. We do not track people live
(geolocation-verification-specialist §4 says the same of wireless data).
“It is public record” is not permission. Even when individual-level data is legally and publicly available, the decision to republish it is ours, and the question is whether sharing it does more harm than good and whether the same story can be told without intruding on individuals (Hands-On Data Visualization, “Mask or Aggregate Sensitive Data”). The cautionary case: in December 2012 The Journal News published interactive maps of the names and addresses of every pistol-permit holder in Westchester and Rockland counties, New York. New York’s next gun law temporarily barred release of permit records and let permit holders opt out of disclosure, including victims of domestic violence and people who fear for their safety; the paper then removed the maps ([Reporters Committee for Freedom of the Press, “In wake of Journal News publishing gun permit holder maps, nation sees push to limit access to gun records”](https://www.rcfp.org/journals/wake-journal-news-publishin)). One map closed a public record for everyone.
What may be shown as a point: public institutions at their published addresses, such as government offices, courthouses, jails and ICE detention facilities, agency field offices, and polling places. A facility is not a person; the people inside it are, and nothing about an individual detainee is mapped.
How to aggregate to a safe geography:
- Default to county or larger for any data about vulnerable people. Go smaller only with an
editor’s sign-off recorded in the register, and never below the census tract.
- Suppress small cells. Federal statistical practice treats very small counts as identifying:
CDC WONDER suppresses statistics representing one to nine persons (CDC WONDER, Multiple Cause of Death, privacy policy). Suppress any area whose count of vulnerable people is 1–9, and say so in the note.
- Know what federal rules call identifying. The HIPAA de-identification “safe harbor” treats
every geographic subdivision smaller than a state as an identifier, and allows the first three ZIP-code digits only where the area they cover holds more than 20,000 people (45 CFR § 164.514(b)(2)(i)(B)). That rule governs health data, not journalism, but it is a fair measure of how small is too small.
- Masking is weaker than aggregating. Some police departments mask the last digits of an
address (“1XXX Main St”); grouping incidents into tracts or neighbourhoods protects people better and tells the bigger story (Hands-On Data Visualization, “Mask or Aggregate Sensitive Data”). For vulnerable people, aggregate; do not mask.
- Check for re-identification by combination. A small area plus a date plus a category
(nationality, age, charge) can single out one person even when each field alone does not. Coarsen the date or the category, or merge the area, until it cannot.
The Locus Charter, the location-data ethics principles facilitated by EthicalGEO and the Benchmark Initiative, puts the same duties in short form: do no harm, protect the vulnerable, do not collect or use more data than the task needs, protect privacy, and prevent the identification of individuals (Ordnance Survey, “Charter launched to promote responsible use of location data”, 25 Mar 2021). EthicalGEO’s own guidelines work starts from the same place: maps can deepen the disadvantage of people at the margins, or help correct it (EthicalGEO, “The Ethical Mapping Guidelines: how not to map”, Erica Hagen, 14 Feb 2020).
3.3 Classification and breaks: choose, then disclose
The same data can tell opposite stories under different breaks. In one worked example, a world map of the top 1 percent’s income share showed the US as middling under a continuous linear scale and in the top tier under three natural-breaks steps, and both maps were accurate (Hands-On Data Visualization, “How to Lie with Maps”, after Mark Monmonier’s How to Lie with Maps).
- Look at the distribution first. Make a histogram of the values before choosing breaks
(Alberto Cairo’s advice, as reported in the same source).
- Know what each method does
(data.europa.eu, “Choropleth classification methods”):
- equal interval or pretty breaks: same-width classes; outliers can squeeze most regions
into one or two classes;
- equal count (quantile): the same number of regions per class; reveals variation in the
middle but can make small differences look large;
- natural breaks (Jenks): groups that follow the data’s own gaps;
- standard deviation: distance from the mean, shown with a diverging palette.
The number of classes changes the pattern too.
- Continuous or steps. A continuous scale keeps nuance; steps give quick readability at the
cost of nuance. More steps add contrast, but too many make small differences look large (Datawrapper, same source; Hands-On Data Visualization, “How to Lie with Maps”).
- Never set custom breaks to fit a story. Manually placed intervals are the most likely to
mislead (Hands-On Data Visualization, “Design Choropleth Colors & Intervals”). A threshold that means something (a legal limit, the national average) is fine; say what it is.
- Disclose. The map note states the method, the number of classes and the break values, and
the legend shows them. Legend steps use equal value intervals where possible (0, 25, 50, not 0, 15, 50), and a diverging legend shows its centre value (Datawrapper, same source).
- Try at least two classifications and keep the one that is honest and useful; record the
others in the register.
3.4 Colour: colour-blind safe, ordered, light to dark
- Sequential palettes (light to dark) for low-to-high values. Diverging palettes only
around a meaningful centre (zero, the average, the median, a legal threshold), with the lightest colour at the centre and the darkest at the extremes; state the centre. Qualitative colours only for categories with no order, and as few as possible (Datawrapper, “What to consider when creating choropleth maps”; Hands-On Data Visualization, “Design Choropleth Colors & Intervals”).
- Colour-blind safe, always. The best choropleths can be read correctly by people with colour
blindness (Hands-On Data Visualization, same section). Use a ColorBrewer palette (Cynthia Brewer and Mark Harrower’s design assistant) with its colour-blind-safe filter on, or the Datawrapper defaults. Never encode the difference by red against green alone.
- Lightness carries the order. Each step must differ in lightness, not only in hue, so the map
still reads in greyscale; keep the brightest and darkest colours for the extremes (Datawrapper, same source).
3.5 Cite the boundary file and its vintage
- Every map names its boundaries: publisher, file and year. For US Census geographies, the
Census Bureau’s cartographic boundary files are the generalised files designed for small-scale thematic mapping, named cb_ (500k, 5m or 20m) (Census Bureau, “Cartographic Boundary File Description”).
- Match the vintage to the data. Boundaries change: the Census Bureau warns that generalised
areas “may not align with the same areas from another year” (same source). County mergers, redistricting and new census tracts all break joins silently.
- Do not measure with display files. The cartographic boundary files should not be used for
area or perimeter calculation, geocoding addresses, or precise geographic relationships; use the detailed TIGER/Line files for analysis (same source).
- Datawrapper basemaps in Compose are named by id (§4). Record the id. The congressional
basemap is the 116th Congress‘s districts: never join data for a later Congress to it, because districts were redrawn after the 2020 census.
3.6 Projection: equal-area for choropleths
A choropleth asks readers to compare coloured areas, so area must not be distorted. The Mercator projection, the default of most web maps, inflates areas far from the equator: it shows Greenland about the size of Africa, which could hold Greenland fourteen times (data.europa.eu, “Map projections”). For the conterminous United States, use an equal-area projection; the USGS uses the Albers Equal Area Conic for maps of the 48 states and many thematic maps (USGS, “Map Projections”). Record the projection in the register. Alaska, Hawaii and the territories shown as insets are labelled as not to scale.
3.7 Geography, granularity and uncertainty
- Choose the geography on purpose. The way areas are drawn changes the pattern, which
geographers call the modifiable areal unit problem (Hands-On Data Visualization, “Map Design Principles”). Smaller units usually show more (same source; Datawrapper, same source), and there are exceptions: the US presidential election is decided by state, so a state map is the honest one. Record why the level was chosen, and never try geographies until one makes the point.
- The privacy floor wins. Where finer geography and §3.2 conflict, §3.2 wins.
- Show uncertainty. Survey estimates at small geographies carry wide margins of error (the
American Community Survey publishes them at the 90 percent confidence level). Put the margin in the tooltip or the note, and do not present two areas as different when their estimates’ margins overlap. The margin-of-error check itself is numbers-and-statistics-integrity‘s.
- Missing is not zero. Areas with no data, suppressed data or no reporting agency get their
own neutral pattern and a legend entry, never the lowest colour.
4. Tools
- Compose: Datawrapper maps. The Datawrapper panel in the Compose editor
(gui/services/datawrapper_service.py) supports two map types, d3-maps-choropleth (filled regions, for rates) and d3-maps-symbols (sized symbols, for counts), on the basemaps us-states-territories, us-counties, us-congress-116, world and europe. The editor names the key column (the region code) and the value column. **The value column must already be a rate for a choropleth**: the panel does not normalise. The panel’s “Chart styling” sends only a base colour, value labels and sort order; classification breaks and a colour-blind-safe palette are set in Datawrapper’s own editor, and the chart republished, before the map goes into the piece. Write the note (§3.3) there too.
- Infographics.
infographic-builderhas no map template today. A map in an infographic is
built elsewhere and must pass §5 before it is placed.
- Enforcement. Today the only enforcement is this checklist and the editor. A Compose
map-template test that refuses an un-normalised choropleth is planned in code; until it exists, §5 is mandatory.
- Denominators and data through
scripts/ca_research.py: scripts/ca_research.py api api.census ": ACS 5-year (vintage"
2024) total population and other bases for each state named. State level only; for county or tract denominators, extract the ACS table from data.census.gov with scripts/ca_research.py extract and record the table id and vintage.
scripts/ca_research.py api api.ice_287g ": ICE’s own roster of 287(g)"
agreements by agency. Agreements belong to agencies, not people; aggregate them to county or state before mapping, and normalise if the map compares places.
scripts/ca_research.py api api.eavs ": the EAC’s Election Administration and"
Voting Survey, one row per reporting jurisdiction keyed by FIPS, for election-administration maps.
- Palettes: ColorBrewer (colour-blind-safe filter on) or the Datawrapper defaults (§3.4).
5. Pre-publication checklist
Every item is ticked in the register before an editor sees the map.
- The story needs a map, and a chart or table would not serve better (§1).
- A choropleth shows a rate with a named base from the same period; a count is a symbol map (§3.1).
- Small counts are flagged or suppressed, and the note says so (§3.1, §3.2).
- No vulnerable person, confidential shelter or private home can be located from the map, its
tooltips, its data download or its labels; the geography is county or larger unless an editor signed off (§3.2).
- Any named private person passed the inclusion gate, and no home or workplace address appears
(§3.2).
- The classification method, the number of classes and the break values are in the note and the
legend; no custom breaks without a stated reason (§3.3).
- The palette is colour-blind safe and ordered by lightness; a diverging palette states its
centre (§3.4).
- The boundary file or basemap id and its vintage are cited and match the data’s period (§3.5).
- The projection is equal-area, or the map is a Datawrapper basemap whose id is recorded (§3.6).
- Margins of error and missing data are shown honestly (§3.7).
- Every number on the map passed
numbers-and-statistics-integrity.
Sources
All read on 2026-11-13 through scripts/ca_research.py (search and extract).
- Datawrapper, “What to consider when creating choropleth maps” (Lisa Charlotte Muth, 21 Mar 2018).
- Jack Dougherty and Ilya Ilyankou, Hands-On Data Visualization (open-access web edition):
“Map Design Principles”, “Design Choropleth Colors & Intervals”, “Normalize Choropleth Map Data”, “Mask or Aggregate Sensitive Data”, “How to Lie with Maps”.
- data.europa.eu Data Visualisation Guide: “Choropleth classification methods”, “Map projections”.
- EthicalGEO (American Geographical Society), “The Ethical Mapping Guidelines: how not to map”
(Erica Hagen, 2020); the Locus Charter principles as published by Ordnance Survey (25 Mar 2021).
- U.S. Census Bureau, “Cartographic Boundary File Description”.
- USGS, “Map Projections” (General Interest Publication).
- 42 U.S.C. § 10406(c)(5)(H) and 45 CFR § 164.514(b)(2)(i)(B), as published by the Legal
Information Institute.
- CDC WONDER, Multiple Cause of Death, privacy policy (suppression of 1–9).
- Reporters Committee for Freedom of the Press, on The Journal News pistol-permit maps (2013).
Cross-references
infographic-builder: the production pipeline; maps in infographics pass §5 first.geolocation-verification-specialist: precision of inferred locations; no live tracking.patriot-private-citizen-inclusion-gate: runs first for any private person a map would name.numbers-and-statistics-integrity: checks the rates, small counts and margins a map shows.patriot-sanity-check: runs last, as for any published piece.
Safety
- Never map a vulnerable person’s location, at any precision that could find them, whatever
the source and however public it is (§3.2).
- Rigor, not direction. Breaks, colours and geography are chosen before looking for the
story, and the same rules apply whether the map helps the piece’s argument or hurts it.
- Disclose every choice a reader would need to rebuild the map: base, breaks, palette,
boundary file and vintage, projection.
- The data behind a map is published too. A tooltip, a CSV download or an alt text can leak
what the colours hide; check them all against §3.2.
- Not reader-facing. The register is internal editorial work product.
