In June 2025 the Election Commission of India began a Special Intensive Revision of the electoral rolls in Bihar, and over the following year it took the revision to most of the country. Every elector on the roll had to fill in an enumeration form, and anyone whose form did not come back was left off the draft roll. By 30 September 2026, the draft rolls published in 31 states and union territories had left off 13,39,03,307 names, about 14 in every 100 electors on the rolls before the revision began.
Off the Roll is my attempt to set those numbers out state by state and seat by seat, and to say plainly what they show and what they cannot. This note explains how the site is put together, because a site that asks the Election Commission to publish its data has to show its own working.
Where the figures come from
The first rule is that every figure comes from the Chief Electoral Officer of the state or from the Election Commission itself. Where a state published its tables by assembly constituency, as Tamil Nadu, Kerala, West Bengal, Delhi and Uttarakhand did, the site works at that level. Where it published by district only, as Karnataka and Maharashtra did, the site works at district level and says why there is no seat page. Where the only copy of a table is a photograph printed in a newspaper, as in Uttar Pradesh and Gujarat, the site uses it and says on every page that the table is the Chief Electoral Officer’s as published in the press.
Some tables were transcribed by hand from scans and photographs, so each one has to pass the arithmetic it implies before the build will accept it. In every row, the roll before the revision less the names left off must equal the draft. Where reasons are given, they must add up to the names left off. The rows must add up to the district totals printed on the same page, and the districts to the state figure the Chief Electoral Officer announced. A misread digit breaks at least one of these identities, and the build stops until it is fixed.
The second rule is that no figure is typed into a sentence. Every finding on the site, including the headline count on the front page, is computed from the published data files when the page loads. If a state publishes a corrected table, I change one row of data and every sentence that depends on it changes with it.
What the reasons mean
Most states published the names left off under three or four reasons: dead, shifted or absent, and enrolled more than once. Across the 31 states, 21 per cent of the names were recorded as dead, 72 per cent as shifted, absent or another reason, and 8 per cent as duplicates. These are what a booth-level officer wrote down after a visit, and nobody has checked them. In Bihar I set two of the reasons against outside records. The names entered as dead come to about 3.4 years of adult deaths at the Sample Registration System’s death rates, and the ages are right, but 52 per cent of them are women, against 43 per cent of adult deaths. The excess of young women marked as shifted comes to about half a year of migration for marriage, but across districts the two run in opposite directions, with a rank correlation of -0.51. Neither test proves the entries wrong. Both are reasons to check them.
The comparison with winning margins
The comparison readers ask for most is whether the removals were large enough to change an election result. In Bihar, the names removed at the draft stage exceeded the 2025 winning margin in 174 of 243 seats. Measured against the final roll, which restored some names and added new ones, the net fall still exceeded the margin in 124 seats. A number larger than a margin means the removals could have decided the seat. It does not mean they did. To get closer, the site estimates how the removed electors would have voted, using three different assumptions, and on each of them between four and six seats change hands, in both directions, and the NDA keeps its majority. The removals also fell more heavily on women. Women were 55.5 per cent of the names removed at the draft and over-represented among them in all 38 districts.
The lists show none of the intent behind the removals, and no page on the site claims that they do.
What the site does not do
The site never names an elector. The row-level lists that the counts come from carry names and EPIC numbers, and they stay out of the repository; a check fails the build if any data file under Bihar carries a field that looks like an EPIC number. I do not name the people or organisations who compiled the elector lists the Bihar and Hyderabad analyses rest on.
States are never ranked against each other, because they published different measures. West Bengal and Kerala published no count of the names left off below the state, so their pages measure the fall in the roll between two dates, which includes ordinary deaths, moves and new voters, and the pages never call that fall the removals. Assam held a Special Revision, which asked existing electors for no form and no document, so it sits on its own page and outside every all-India total.
Languages and corrections
The site is in English, Hindi, Urdu, Telugu, Tamil and Bengali. No native reader has yet signed off the translations, so every translated page says that the English page governs. The full reports and the newer state pages are in English only for now, because legal detail should not be machine-translated without review.
Corrections go to the address in the footer. The terms page commits to acknowledging a correction within 24 hours and answering it within 15 days. Every file the site publishes can be downloaded, with a sheet explaining each column, so anyone can check the arithmetic for themselves.