In a village everyone already knows your caste. In a city a stranger has your name.
India is urbanising, and that moves the question of who can identify whom. In a village, caste is common knowledge: the settlement, the hamlet, whose son you are, which lane you live on. A city strips most of that away. What survives on a rental application or a job form is a name.
How much does a last name give away? Anyone acting on caste has to work it out first, and so does anyone auditing them for it: a study asking whether Dalits get fewer callbacks usually infers caste from the applicant's name.
This repo measures that inference. It measures what a surname reveals about caste, whom it reveals least about, and what has to be added to a name before it reveals a lot.
The operational question: pick an Indian adult at random, know only their last name, and how often are you wrong about their caste?
About 20 times in 100. Knowing nothing at all, you would be wrong 25 times. The name is worth about five mistakes, and almost all of that comes from a handful of names hardly anybody carries.
Everything below is measured in that one unit: of a hundred people, how many would you get wrong?
Error rates depend on two things besides the surname. The first is the number of categories: sorting people into Dalit, Adivasi or neither is easier than placing them among 141 Bihari jatis, and the same surname leaves 20 mistakes in the first case and 47 in the second. The second is what else the guesser knows: a surname with a village is a different predictor from a surname alone.
Given that both vary across the analyses below, each result is reported with its target and its cues. Error rates computed against different targets are not comparable.
Indian electoral rolls do not record caste, and the caste census does not record where a household sits relative to its neighbours. Given that no single source carries both, the analyses below combine five.
| what it gives | what it cannot give | source |
|---|---|---|
| caste composition of a surname, nationally | fine categories; who actually carries the name today | SECC 2011 |
| how common a name is, and who a person's father or husband is | any caste at all | electoral rolls, 2017 |
| jati and hamlet for every household | anywhere outside Bihar; the landless | Bihar land records, Mahadalit census |
| jati and village per tenant in a second state | a partial state; still being scraped | Odisha Record of Rights |
| a third population with its own category labels | a population sample of anything | Karnataka PSC lists |
The national picture needs the first two joined. Whether the place does the work the name gets credit for needs the third, because only there can you put a name and a village together and take the village away. Whether the premium that produces is a fact about villages or a fact about Bihar needs the fourth, which records the same two things in another state. Whether "last name" even means the same thing across India needs the second and fifth, because those are where a name can be checked against a relative name or against an initial.
Every national number joins two sources on the surname.
outkast supplies the caste side, from the 2011 Socio-Economic and Caste Census. For each surname it gives the share of its bearers who are Scheduled Caste, Scheduled Tribe, or neither. It ships only cells holding at least 100 records, which leaves 3,930 surnames.
instate supplies the frequency side, from the 2017 electoral rolls: how many people carry each surname.
The split is deliberate. SECC counts heads of household, who are mostly men, so it badly undercounts women's surnames: Devi appears 2.3M times in SECC and 45M times on the rolls, where it is the commonest surname in the country. So how common a name is comes from the rolls, and what it means comes from SECC.
Filling a room with a hundred people means drawing surnames at their roll frequency and reading each one's caste mix off SECC. That room is 19 Dalit, 6 Adivasi, 75 neither.
Read every national number here against one limit. Those 3,930 surnames are 59% of all the names people carry on the rolls. The other 41% are the rarer names, cut by the 100-record floor, and rare names are the informative ones. So the room is built from the commoner half of Indian naming, which is the half that reveals least. Every figure below is a floor on what a surname gives away, not a ceiling.
Upnaam resolves which token in a name
is the recorded surname, versioned, for Bihar, Rajasthan and Maharashtra
(analysis 05). jati, which is not a public repository, and
land hold the Bihar land records and the
Mahadalit census, with a jati and a hamlet for each household (analyses 02 and
06). pranaam holds the Karnataka
Public Service Commission select lists (analysis 08), and
odisha-ror the Odisha Record of
Rights (analysis 09). That last one publishes the scraper and not the records,
so analysis 09 needs a local fetch to reproduce.
Because analyses 02 and 06 read from a local clone of jati, those two cannot
be reproduced from a clean checkout. The other six can.
Take one Dalit and one non-Dalit at random and rank the pair by what their surnames say about caste. The Dalit ranks higher 80 times in 100. A surname carrying no information would give 50, and for Adivasi against everyone else the figure is 91. So a surname is far from silent about who is more likely to be Dalit.
Now ask the same data to name a single category for each person. Guess "neither" for everyone in that room of a hundred and you are wrong 25 times. Let yourself hear the last name and you are wrong 20.
Those two facts are both true and the gap between them is the point. Only about 19 people in that room are Dalit, so a surname can raise someone's odds a great deal and still leave them below an even chance, at which point a rule that names the most likely category goes on answering "not Dalit". The name moves the odds; it rarely moves the decision.
This is why accuracy is the wrong measure to lead with. In a population that is 70% one group, a rule naming that group is right 70 times in 100 whatever surnames reveal. The 25 and the 20 above are mostly a statement about how many Indians are not Dalit.
The five mistakes the decision saves are not spread evenly either. A surname is predictive when few of its bearers fall outside its largest category, and it changes the guess only when that largest category differs from the population's. The commonest surnames are often the first without being the second. Yadav is the fifth commonest surname in India at 2.6% of the roll, and 99% of its bearers fall outside the schedules, so it leaves 1 mistake per 100. It nonetheless changes nothing, because "outside the schedules" is what one would have guessed without it.
For 91% of people the surname does not change the guess at all. The surnames that do change it are concentrated in the tail:
| rank by frequency | names | share of India | how many change your answer |
|---|---|---|---|
| 1–10 | 10 | 32% | none |
| 11–25 | 15 | 13% | 2 |
| 26–50 | 25 | 10% | 8 |
| 1001–3930 | 2,930 | 6% | 625 |
The ten commonest surnames cover a third of the country, and none of them moves the guess off the base rate. The 2,930 surnames ranked 1001 and below cover 6% of people, and 625 of them do move it.
A further 16% of people carry a surname whose caste composition is more
evenly divided than the population's. ram is 45% Dalit and 47% not. That is
real information, and it does not change the decision: the best guess for
someone named Ram is still "neither", and it is wrong 53 times per 100, the same
rate as guessing without the name. Such a surname warrants less confidence in an
answer it does not alter.
Analysis: 01. The ranking statistic is the Mann-Whitney statistic, equivalently the area under the ROC curve, computed twice by different routes as a check.
Part of the answer to claim 1 is that the thing being measured is not stable. The last token of an Indian name is a family name in some places and something else in others, and three of the analyses here found it failing in three different ways.
It can mark sex rather than family. The commonest last name in India is devi, at 6.5% of the electoral roll. On the Bihar rolls 81% of women carry a name from this group against 10% of men, and the family names in the same state read as 84 to 89% male, because those families' women are on the roll as Devi. A woman and her brother do not share a last name, and a name assigned by sex cannot track a lineage. The clear cases cover 10.5% of the country, and adding Singh and Kumar takes it to 19%.
Eighteen names cover a quarter of India, and 103 if you count only names a brother and sister share. That gap is not a correction toward a truer figure. Indian naming is as concentrated as the first number says, and Devi really is what tens of millions of women are called on a roll. The two count different things, names and lineages, and the distance between them is the room the caste signal used to occupy.
It can be in the other position. Every electoral roll record carries the
elector's father or husband, and a token appearing in both names is one that
passed between two family members. That test confirms the ten sex-marking names,
which all transmit under 1% of the time, and splits the two the hand list called
ambiguous: Singh transmits 78% of the time and Kumar 8%. It also found that
Maharashtra writes the surname first, as in patil ashwini with father
patil ashok, so instate's last-token column holds a given name there.
Analysis 03's Maharashtra and Gujarat figures were withdrawn.
It can be an initial. In Karnataka's Public Service Commission select lists
the last token is a single letter 34% of the time, and the six commonest are
S, R, K, N, B, C. Drop single letters and the list becomes Kumar,
Patil, Naik, Manjunatha: two inventories of "Karnataka's commonest
surnames" from one set of 48,395 names.
And it can be an initial with nothing behind it. Kerala is the limiting
case. 78% of last tokens there are a single letter and only 27% of people
have a written surname at all: a name is a given name followed by initials
standing for a father and a house, as in KALPANA A or SHYLAJA E T. Scoring
Kerala PSC candidates on Scheduled Caste against the other caste categories, the
last token ranks a Dalit's higher 59 times in 100, a written surname 70, and
the given name 69, which everybody has. The informative part of a Kerala name
is the part no surname-based method reads.
Any pipeline that takes the last token and calls it a surname is therefore measuring something different in each of those places, and in Kerala measuring almost nothing.
The national figure averages over people whose location is unknown, so it cannot separate the surname from the place. Bihar's land records identify both. A surname and a village leave 17 mistakes per 100 across 141 jatis; the surname alone leaves 47.
Both figures are leave-one-out: when a household is scored, its own record is excluded from the table used to score it. That matters here because half the name-and-village cells contain only one household. If its own record is left in, such a household is matched against itself, and the surname-and-village figure comes out at 16 instead of 17. The correction is small, but it applies almost entirely to the surname-and-village rung, which is the rung this comparison depends on.
That comparison knows the village in advance, which is the position of someone who already knows the place and not of a stranger. Holding whole villages out of the fitting leaves only the surnames of the other households there as a cue. On the Bihar land records those neighbours save about four mistakes per 100. The average hides the case: Chaudhary goes from 80 mistakes to 56, Prasad from 72 to 54, while Paswan, which already identifies you at 6, gains nothing. The names that carry no caste information alone are the ones the neighbourhood rescues.
How much the place is worth varies enormously by state. Sorting people into Dalit, Adivasi or neither, the same guess closes 67% of the gap in Assam and 0% in Haryana. That measure carries claim 1's problem, so it comes with the ranking measure beside it, and the two disagree in a way worth seeing. Kerala closes under 1% of the gap while its surnames separate Dalit from non-Dalit about as well as Maharashtra's, 0.85 against 0.86, because only 8% of Kerala's extract is Scheduled Caste. Punjab and Haryana are the states whose surnames genuinely carry little, at 0.55 and 0.59 against a floor of 0.50.
Punjab's roll says why, and the reason is not simply that its names are concentrated. Bihar's three commonest names touch 78% of its pairs and Bihar ranks 0.96. What matters is whether the dominant names sit where the state sits.
The clearest form of it is one surname doing opposite work in two states. In
Bihar, singh is carried by 9% of people and almost none of them are Dalit,
against a state that is 18% Dalit; knowing someone is called Singh there tells
you a great deal. In Punjab it is carried by 73% and its Dalit share is
0.38 against a state of 0.38; there it tells you what you already knew.
Across ten states, the more of a state a name covers the closer its composition
sits to that state's own.
A name that spreads to everyone stops distinguishing anyone. That is a
process, not a measurement problem. Singh and Kaur were adopted across castes,
and the caste information the surname carried went with the adoption. It is the
same mechanism as claim 2's sex-marking names, seen from the other end: Devi
spread to women of every caste, and stopped meaning anything about lineage for
the same reason. What carries a state is one or two large, decisive names rather
than many of them: Assam has two majority-Dalit surnames among its commonest and
closes two thirds of the gap, because one of them is das. Punjab has five and
closes 3%, because its biggest are ram at 62% and lal at 52%. Haryana has
none at all. Every state's coverage is printed beside its result, because these
are surnames that cleared a 100-record disclosure floor and they cover between
3% and 19% of a state.
And the village premium is smaller outside Bihar. The Odisha Record of Rights records a jati and a village for every tenant, so the same measurement is possible elsewhere. Across 30 districts and 10,604 villages, on the same protocol, a village adds 24 points where in Bihar it adds 30. Put as a share of the errors the surname leaves behind, adding the village removes 55% of them in Odisha and 64% in Bihar.
That pooled figure hides a spread. District by district the premium runs from 2.2 to 22.8 points, median 10.7, and the pooled number is larger than any single district: naming a village in the pooled problem also names a district and a region, so it carries geography a district-level score has already held fixed. Bihar's 30 is pooled too, which is what makes the pooled rows comparable.
The two levels are not comparable: Bihar sorts people among 141 curated jatis and Odisha among 1,397 labels — of 6,507 strings recorded, kept when they reach more than one village and either 50 households or 5 villages, of which 387 carry 99% of tenants — and a harder target costs mistakes whatever the surname does. Given that, the comparable quantity is the distance between the rungs. The two ladders cross: Odisha is the easier target at every size of place except the finest, and Bihar overtakes it exactly when the village arrives.
Three of the analyst's choices could have produced that difference, so each was
varied. Two spellings are merged only when the bearers' surnames agree as well
as the strings, because ଗଣ୍ଡ/ଗଣ୍ଡା and ଚମାର/କମାର are one edit apart and
two different jatis each; moving that gate across its whole defensible range
moves the premium by 0.14. Stripping the religion suffix from the jati label
moves it from 24.1 to 23.3. The last token is the surname for 92% of tenants who
share one with the relative named beside them, so that rule is measured rather
than assumed. The scrape is unfinished and its districts were not entered at random,
so this is a partial Odisha; a completed village is now censused rather than
sampled, but villages still in progress are partial and bias the premium
downward.
An average error rate conceals who bears it. If the uninformative names are not spread evenly, every name-based caste method has differential error in a knowable direction.
They are not. Nationally, the guess is wrong about 66 of every 100 Dalits and 4 of every 100 people outside the schedules, a seventeen-fold gap. Those weight up to the 20 mistakes per hundred it makes across everybody, so this is one guess split by who it lands on rather than a second estimator.
This is claim 1 seen from the other side. Because a surname rarely pushes anyone past an even chance of being Dalit, the rule almost never answers "Dalit", so it misses most Dalits and is right about nearly everyone else. The lopsidedness is not a separate finding. It is what naming the most likely category does to a minority when the cue is informative without being decisive.
The name is not useless for Dalits. It takes them from wrong about all of them to wrong about 66 in a hundred, and it does more still for Adivasis, 100 down to 43. But a large gain still leaves Dalits far and away the worst served.
The same pattern appears on independent data. Karnataka is absent from the census extract the rest of this repo runs on, so its Public Service Commission select lists are the only caste-linked name data available for the state. The guess there is wrong about 62 of every 100 Scheduled Caste candidates and 17 of every 100 General ones, an error 3.6 times larger for the group the quota exists for. That is a different state, a different source and a different set of labels, pointing the same way. Overall the surname closes 24% of the gap there, from 52.5 mistakes per 100 to 40.0.
This is the result with consequences. Use a name-based method to test whether Dalits get fewer callbacks, and two thirds of the real Dalits sit in your comparison group, on both sides of the gap you are trying to measure. The discrimination is unchanged; your estimate of it shrinks.
The same question asked about sex has no such answer. Women's names cost 3.3 more mistakes per hundred than men's in Bihar, but 1.4 fewer in Rajasthan and 0.2 fewer in Maharashtra. Punjab, where Kaur and Singh make the most sex-marked naming system in the country, adds 0.6. Four states, no direction.
Read them with care: after matching Upnaam's resolved surnames to the caste table they rest on 57% of the Bihar roll, 36% of Maharashtra's, 20% of Rajasthan's and 19% of Punjab's. Only Bihar is built on most of a state.
Caste is printed on no electoral roll, so the question a roll actually poses is what its cues are worth to someone who can match them against a caste register of the same population. Across 887,512 Scheduled Caste households in 8,307 held-out villages, sorted into 22 jatis: knowing nothing leaves 59 mistakes per 100, the surname alone leaves 30, neighbours bring that to 26, and the name, the father's name and the hamlet together with a caste register leave 9.
A surname is a weak instrument. A surname joined to a register is a strong one.
Analysis: 06.
Each owns its pipeline, data loading, figures, note and out/.
| what it does | |
|---|---|
| 01 surname to category | what a surname alone reveals, nationally |
| 02 jati by geography | surname against surname-plus-village, Bihar, 141 jatis |
| 03 how few names | how concentrated Indian naming is, and what a sex-marking name does to that count |
| 04 which token is the surname | which token transmits between family members |
| 05 who has an uninformative name | who the guess fails, by their own caste and by sex |
| 06 neighbours | held-out villages, the neighbour cue, and the ceiling with a caste register |
| 07 where the name works | the spread across fifteen states |
| 08 Karnataka PSC | the only caste-linked name data for a state absent from SECC |
| 09 Odisha village premium | whether the village premium travels outside Bihar |
| 10 Kerala PSC | the state where most people have no surname to read |
- Analysis 01 covers SC / ST / Other only, because SECC has no OBC category. The finer categories exist only in analysis 02, and only for Bihar.
- Analysis 02 is Bihar landowners, so it under-represents the landless, who are disproportionately Dalit and EBC.
- Analysis 03's title list is a judgment call, published in
analyses/03_how_few_names/titles.pywith a reason beside each token and reported at three levels so you can take the conservative half. Patronymics and OCR debris are further non-surnames it does not quantify, so even 19%, the widest of the three levels, is a floor. - The scoring analyses assume a perfect guesser with one or two cues. That is a floor on what is knowable, not a ceiling on what someone can work out about you: a real person also has your first name, your father's or husband's name, and your neighbourhood.
The quiz ships the caste composition of the 1,000 commonest surnames, which is 93% of the names people carry. Until it existed this repo published no per-name table at all, on the reasoning that a ranked list of how much each name gives away is a screening tool. Shipping the thousand is a deliberate change to that position, not an oversight, and it is worth being plain about what it costs: those names sorted by composition are such a list.
What still holds. The analyses publish results by frequency band, by geography and by group, never a ranked diagnosticity table. Nothing here infers a caste for a named individual, and the app is built so it cannot be used to: it shows a pair and asks a question, with no lookup box and no way to ask about a name of your choosing.
The quiz also drops two kinds of name that would make it a worse instrument than
it looks. Religion-marked surnames are near-perfect "not Dalit" tells, carrying
4.8% of non-Dalit draws against 0.1% of Dalit ones, so a player who recognises
them wins without knowing anything about caste. And one surname spelled two
ways, sing against singh, asks about transliteration. Removing both costs
one point of the score the data achieves.
uv venv .venv
uv pip install --python .venv/bin/python -e '.[dev]'
make all # the eight scripted analyses: tables, figures, notes
make a01 # just the first, and so on through a09
make a04 # runs the notebook
make test
make lint
Each analysis owns its own pipeline, data loading, figures, note and out/.
Shared code is only the scoring measure, the drawing style and one data reader:
| path | what |
|---|---|
src/last_name_basis/scoring.py |
mistakes per hundred; ladder scoring; leave-one-out |
src/last_name_basis/style.py |
palette, hundred-square grid, band definitions |
src/last_name_basis/upnaam.py |
validated reader for resolved roll surnames |
MIT licensed.