2018 and 2024
These two maps show, for the 2018 and 2024 general elections, how women's participation in voting compared with men's in every constituency. They are two snapshots, measured the same way, so you can set one election beside the other. The colour is the gap between women and men, not the number of votes: a seat where few of either voted looks the same as one where many did, as long as women and men took part at the same level. Whether that gap widened or narrowed between the two elections is the separate map below.
The formula. For each constituency we take the votes cast at women-only stations and
measure them against the registered women (women's votes ÷ registered women, written
Vf/Rf), and we do the same for men
(Vm/Rm). We then express women's figure as a share of the two
together: (Vf/Rf) ÷ (Vf/Rf +
Vm/Rm), a number between 0 and 1.
What the number means. 0.5 is parity: women and men took part at the same level. Below 0.5, women's participation trailed men's, relative to how many of each were registered; above 0.5, women's ran ahead. For example, if women's stations returned a vote equal to 40 percent of registered women and men's stations 60 percent of registered men, women's share is 0.40 ÷ (0.40 + 0.60) = 0.40, below parity. The pale end is parity, and the colour deepens as women fall further behind. The value is always between 0 and 1; it is never negative.
Why we measure it this way. Two reasons. First, we measure against the registered numbers, not raw vote counts, because more men than women are on the electoral rolls across much of Pakistan; counts alone would show men ahead almost everywhere simply because more men are registered, and dividing by the registered number removes that. Second, we use women's share of the two figures rather than the figures on their own, so parity always sits at the same point, 0.5, on every seat. A busy seat and a quiet seat both read 0.5 when women and men take part equally, which lets the colour show the gender gap alone, independent of how heavy voting was.
What it is measured from. Only the women-only and men-only polling stations, because they are the only places the vote is recorded by gender; combined stations, which do not record it, are set aside. Every constituency figure is built from the bottom up, from the polling stations themselves: the votes are summed across the individual stations in that seat, and the seat's figure is formed from those sums; no constituency-level published total is used. Nothing is modelled or estimated: every coloured cell is a measured figure, and the identical formula is applied to 2018 and to 2024, so the two maps are directly comparable.
What it tells you, and what it does not. It tells you whether women took part at the same level as men in that seat, in that election. It does not tell you how many women voted, or whether voting was heavy or light: a deep-coloured seat can have high overall turnout, and a pale seat low. For the raw numbers behind any seat, open it.
Grey means we could not measure it. A seat is grey wherever a women-versus-men figure cannot be formed: no women-only or men-only stations, or, for 2018, a seat that was not digitised or an area that was not a constituency then. Hover any grey seat and it tells you which.
How we know a station's gender. Neither election's returns break the vote down by gender; we recover it by identifying the stations that served women only or men only. For 2018 that comes from the Commission's own polling scheme, which records how many women's and men's booths each station has; it settles 56.4 percent of 156,114 stations and agrees with the station-name tag 92.5 percent of the time where both exist. A building's name is never used to guess: among stations at a girls' school whose gender we actually know, only 40.9 percent are the women's station.
How the shapes are drawn. No machine-readable constituency boundaries exist for either election, so each seat is its real district split into cells, one per constituency, each cell's area proportional to registered voters. The district and the relative sizes are real; the dividing lines are schematic, not the official boundaries. The same construction is used for both years.
What the colour shows. By default, the raw gap: women's share of the votes cast, women against men, at the single-gender stations. Because fewer women than men are registered across much of Pakistan, part of this gap is registration itself. Switch on Adjust for registration differences to compare turnout rates instead and take that out. The pale centre is parity in both; deeper colour means women further behind.
Constituency shapes are schematic: each district split into one cell per seat, area proportional to registered voters, not official ECP boundaries; the same construction is used for both years. Azad Jammu & Kashmir and Gilgit-Baltistan are hatched: they elect their own assemblies, send no members to the National Assembly, and so return no Form 48; that is not participation of zero. Grey is a seat that was contested but where women's participation cannot be measured; 2018 has more of them (93 of 806 seats).
Women's participation · 2018 to 2024
This map shows how women's participation in voting relative to men's changed between the 2018 and 2024 general elections, for every National and Provincial Assembly constituency, drawn on the 2024 boundaries with National and Provincial side by side. Teal marks seats where women's participation rose relative to men's; orange where it fell; grey where a comparison cannot be made, and every grey seat says why on hover.
Across the comparable seats the average share was 0.448 in 2018 and 0.445 in 2024; it fell in most seats and rose in the rest. Hover any constituency for its 2018 and 2024 figures and its match coverage.
The figure we compare. For each year we compute women's share of the two genders' participation, the same measure as the two maps above: the votes at women-only stations against registered women, the same for men, then women's figure as a share of the two, (Vf/Rf) ÷ (Vf/Rf + Vm/Rm), a number from 0 to 1 where 0.5 is parity. Each year's constituency figure is built from the bottom up, from the polling stations themselves: the votes are summed across the individual stations in that seat and the seat's figure formed from those sums; no constituency-level published total is used.
What the colour is: the change. The colour here is not that share, but the change in it between the elections: the 2024 share minus the 2018 share. This is a signed number. Where it is negative, women's participation fell relative to men's between the two elections; where positive, it rose; zero (the pale middle) is no change. Across the comparable seats the average share was 0.448 in 2018 and 0.445 in 2024, and it fell in most seats and rose in the rest.
Why we cannot simply subtract the two maps above. The constituencies were redrawn between 2018 and 2024 (the 2013 delimitation to the 2023 one), so a 2024 seat has no single 2018 seat beneath it; you cannot lay one map on the other and subtract. What does not move is the polling station: a fixed physical place, almost always a government school, that keeps its name and its women-only or men-only character from one election to the next. So we rebuild the 2018 side on the 2024 ground: we match each 2018 station to its place on the 2024 map by name, comparing the village or locality (a shared phrase like "Primary School" tells us nothing, so it is ignored) and requiring the match to stay in the same district and assembly. This placed about 95 percent of the 2018 polling stations into their 2024 constituencies; the few left unplaced closed or were renamed beyond recognition.
How each side of the difference is built. The 2024 share is native: 2024's own women-only and men-only stations, over 2024's own registered women and men. The 2018 share is rebuilt: we take the matched 2018 stations' votes exactly, and, because registered voters are known only per constituency, we share each old constituency's registered women and men into the new seat in proportion to how many of its stations moved there. Votes are exact; only registration is apportioned.
Grey means no comparison. Grey always means missing data, and the reason is stated on hover: no 2018 station matched, an area that was part of the Federally Administered Tribal Areas with no provincial seat in 2018, only one gender of station matched, no women-only or men-only stations in 2024, or incomplete 2024 figures. A comparable seat also reports on hover how completely its 2018 side could be reconstructed.
All 854 seats, ranked by women's turnout relative to men's, furthest behind first. This table has its own search and filters; click a row to open the seat's detail.
One number per seat: women's turnout rate divided by men's, from the single-gender polling stations only. 1.00 is parity; below it, fewer women turned out than men, relative to how many were registered. The combined-gender stations are set aside rather than modelled. This is the same measurement the maps shade.
Scroll the table sideways to see every column →
| Seat | Name | Province | Women ÷ men | All stations |
|---|
Method
Every figure on this page is drawn from the polling-station returns themselves. Here is how 9,390 pages of scanned paper became a dataset, how turnout by gender was estimated from it, and how far the estimate can be pushed.
What was read to build this page
The Election Commission publishes Form 48 (the Returning Officer's consolidated statement of the count, one row per polling station) as scanned images on its public website. There is no digital export and no usable text layer. We downloaded all 858 documents, rendered every one of their 9,390 pages at 300 dpi, normalised page orientation, and read each vote grid with Claude Sonnet 5, Anthropic's vision language model, over its Batch API, one request per page.
The check on that reading is the form's own arithmetic: candidate votes must sum to the printed valid total, and valid plus invalid must equal the printed total. 95.4% of all rows reconcile: 97.8% in Khyber Pakhtunkhwa, falling to 83% in Balochistan, whose scans are the weakest. No poor scan was discarded; it was flagged and kept, and every correction is logged and reversible.
The result is 179,202 polling stations across 854 constituencies, read from the paper the Commission publishes. We are not aware of an earlier machine-readable record of the 2024 count at this level. The full procedure is set out in the digitization methodology note that accompanies this release.
Form 48 records votes per station, not per voter, so gender is not a column in the source. What it does carry is the station's name, and Pakistan's station plans designate each one for women, for men, or combined. We refer to these as single-gender and combined-gender stations (the Election Commission's own label for the mixed ones is simply combined). Votes cast at single-gender stations are therefore directly attributable to women or to men.
So turnout by gender is measured the plainest way there is: for each seat, women's turnout rate (women's votes at women's stations ÷ registered women) is set against men's, and the page reports the ratio of the two. 45.4% of stations are combined-gender, and their return does not record the voter's gender; those stations (with a further 622 whose designation could not be read) are set aside, not modelled. Because combined-gender stations are larger than average, that sets aside 48.5% of the vote, not 45.8% of it.
Nothing on this page is modelled, and no 2018 information enters the numbers. The one correction that could be applied (for the fact that men and women are not equally likely to be assigned to a combined station) is deliberately left out, because measuring it would import the 2018 station plans into a 2024 figure. Every seat reports the share of its vote that was set aside, so you can see exactly how much of any single ratio rests on how few stations.
The 2023 delimitation redrew Pakistan's constituencies, and we could not find its boundaries published in machine-readable form anywhere we looked: not with the Election Commission, HDX, Datawrapper, ArcGIS, GitHub or Wikimedia. The only public record we found is scanned paper.
So the geometry is derived from official administrative boundaries by a specified procedure. Every seat is resolved to its real district using the Commission's own delimitation sheets. Multi-district seats become the union of their whole districts. The four districts created after 2020 are lifted from their genuine tehsil polygons. Each district is then partitioned into one cell per constituency, with cell area proportional to registered voters.
What is surveyed and what is derived, stated plainly: the country, province, district and tehsil outlines are official boundaries from OCHA. Each seat's district is official. Each cell's area is a true function of its registered electorate. What is derived is the position of the dividing line within a district; the Commission's exact internal boundary exists on paper only.
The construction is verified rather than asserted: tiling error 0% (the cells exactly fill their district, no gaps, no overlaps) and the correlation between cell area and registered voters is 1.
The 2018 map is built by the identical procedure: the 2018 election ran on the older (2013) delimitation, so its seats are traced by splitting the same real districts into one cell per 2018 constituency, area proportional to registered voters. Because both years are constructed the same way, the two maps are directly comparable. A 2018 seat we could not place in a district, and any seat with no single-gender basis, is left grey.
Candidate figures are published for 737 of 854 seats; the remaining 117 are held back until the Form 48 scan has been re-read by hand and their figures can be finalised; open the Candidates tab on one of them and you will see the notice rather than a number.
This does not touch the turnout figures. Turnout is built from each station's totals, not from candidate identity, so a seat whose candidate rows are still being checked still carries a sound turnout measurement; every seat's Quality tab shows the checks it passed.
Stated rather than quietly dropped:
Every step of this work is reproducible. The pipeline is a numbered sequence run in order; the raw data is never written to, only read. Every correction applied to a digitized cell is logged and reversible, every external input carries its URL, retrieval time and SHA-256, and the geometry is deterministic: the same shapefiles and the same code produce identical polygons, byte for byte.