Column 3: evidence · the aggregate among the units
Open data, followed all the way down to one person inside it.
Every piece here starts with a public dataset and ends with a story. The numbers can be checked: any figure set in mono opens its own evidence in a column to the right, so you never have to leave the sentence you are reading to find out whether it is true.
open to work
work
files
introduction
Four tables covering the 34 provinces of Indonesia. The legal minimum wage a province may pay, once a year from 2002 to 2022. The poverty line, twice a year from 2013, split into food and non-food and into town and countryside. What people actually spend, split the same way. And the hourly wage employees actually receive.
The wage is per worker per month. The poverty line and the spending are per person per month. That mismatch of denominators is not a nuisance here — it is the whole instrument. One row is one province.
A minimum wage is an answer to a question: what does it cost to live here? Divide it by the cost of one person and the answer arrives in a unit anybody can picture. In the median province one minimum wage holds 4.77 people above the poverty line — and across the panel the figure runs from 2.79 to 8.86. The same legal category, meaning very different things in different places, and the spread has not narrowed in 7 years.
Then the harder result. Regressed with province fixed effects, the wage tracks its province’s own poverty line almost exactly (+1.146). Add year fixed effects and the relationship vanishes (−0.262, interval crossing zero). The apparent tracking was the national inflation trend the whole time. A wage set by a national formula cannot track a local cost, and this is that fact arriving as a coefficient.
Before any of it: the file carries a national aggregate among its provinces, and one of its wage columns is scrambled. Both are shown rather than silently repaired.
| file | rows | cols | one row is | columns used here |
|---|---|---|---|---|
ump.csv | 35 | 9–157 | a province, as 21 years of minimum wage | provinsi and every ump.<year> column, in the file's own row order |
gk.df.csv | 5,460 | 3–6 | a province-year-period-area-kind poverty line | provinsi, jenis, daerah, tahun, periode, gk |
peng.df.csv | 4,620 | 3–6 | a province-year-area-kind expenditure | provinsi, daerah, jenis, tahun, peng |
upah.df.csv | 280 | 3–6 | a province-year hourly wage | provinsi, tahun, upah |
Prose is written for anyone. Every figure set in mono opens its own evidence in a column to the right. Blocks marked advance hold the method, the confidence intervals, and the objections — open them if you want the working, skip them and the piece still reads.
piece 03 · the premise
A minimum wage is a number that claims to know what living costs. Put it next to the official price of staying alive and it starts answering in people.
Indonesia sets its minimum wage province by province, so there are 34 of them. It also publishes, province by province, the garis kemiskinan — the monthly rupiah below which one person counts as poor. Both are monthly. Both are in rupiah. One is per worker and one is per person, so dividing the first by the second gives a count of human beings: how many people one legal minimum wage can hold above the official floor.
In 2022 the median province came to . The best-off province managed 7.92. The worst managed 3.53. Across all 238 province-years in the clean panel, sit below five people and 8% below four.
before anything: what is wrong with the file
The file has 35 rows and calls them all provinces. One of them is INDONESIA — the national aggregate, sitting in the same column as the units it aggregates. Left in, it quietly bends every spread, every correlation and every extreme in this piece. , and the 34 that remain are the provinces.
Worse, the 2021 minimum-wage column is wrong for . Not wrong in a way that needs an outside source to catch: 9 of those provinces are holding, to within two rupiah, another province’s wage. The year is excluded from everything above. What survives the exclusion is the real 2021, and it is a good deal starker than a data error — 18 of the 21 uncorrupted provinces did not raise the minimum wage by a single rupiah.
the finding · a wage that stopped listening
If the minimum wage were answering the cost-of-living question, the ratio above would be roughly flat across provinces: expensive places would pay more, cheap places less, and the number of people a wage holds would come out similar everywhere. It is not flat. Its coefficient of variation is , essentially unchanged from 0.231 in 2015.
Two more measures say the same thing louder. Dispersion of the wage itself, as a Theil index, has been flat for 20 years — . And the ranking of provinces barely moves: the order in 2015 predicts the order in 2022 at +0.920. Whatever this is, it is a structure, not a drift.
Now the measurement that matters. Regress the log wage on the log poverty line with a fixed effect for each province — the standard way to ask whether a province’s wage follows its own costs — and the answer looks like a resounding yes: , a near-perfect one-for-one. Then add a fixed effect for each year, which removes whatever was happening to the whole country at once. The coefficient collapses to , with an interval that comfortably contains zero.
The first estimate was measuring inflation. Both series climb together year after year, and a province fixed effect does nothing to separate a common national climb from a local response. Once the year is held constant, a province’s wage carries no information about what that province costs. Which is exactly what you would predict of a wage set by national formula rather than by local price — and Indonesia has set it that way, by formula, since 2016.
a check that draws no line at all
Every figure so far leans on the poverty line, and a poverty line is something somebody chose. There is one way to check welfare here without choosing anything: Engel’s law, the most durable regularity in the whole of household economics. The poorer a household, the larger the share of its spending that goes on food. No line required — just a ratio.
The file lets that ratio be computed twice: for what people actually spend, and for the poverty line itself. In 2022 the median urban household spent of its money on food, and the median rural household 56.5%. The poverty line for the same places and the same year is 72.8% and 76.1% food.
The line is describing a basket nobody buys, and the gap has held near 26 points for as long as both series exist. That is not drift; it is construction. The poverty line is not a measurement of how the poor live. It is a stipulation of what surviving would cost if you spent like nobody does — which is worth knowing before treating a poverty count as an observation about the world.
move it yourself
The formula Indonesia adopted in 2021 sets the minimum wage from consumption per head, multiplied by household size, divided by how many people in the household work. Those are three levers, and the law pulls them using national averages. Pull them yourself and see where they land — the province, the number of mouths, the number of wages, and whether the line being cleared is the town line or the village one.
A household of 4 in DI YOGYAKARTA with 1 earning the minimum wage lives on Rp 460,229 per person per month — that is 0.88× the poverty line where they are.
below the line, by the arithmetic of the law that set the wage
…where the wage is DI YOGYAKARTA’s legal minimum for 2022, Rp 1,840,916 a month, and the poverty line is Rp 521,672 per person per month for town and village together. Of the 24 household shapes available here, 6 fall below the line.
DKI JAKARTA has no rural poverty line, having no countryside, so it has no row here. Every cell above was computed in Python and committed before this page was built; the controls index an array and divide nothing.
Two things tend to come out of playing with it. The first is that a single wage almost never clears a household on its own by much: a four-person household with one earner falls under the line in and clears it narrowly in most of the rest. The second is stranger. Switch the line from town to village and the same wage suddenly holds more people above it — not because anything improved, but because the line itself is lower there. The measure moves under the thing it is measuring.
the floor, and what is happening to it
One last series. Alongside the legal minimum, the file carries the wage employees are actually paid, by the hour. Put it on the same monthly footing — 173 hours, the statutory full-time month — and compare.
In 2015 the median province’s actual wage stood at its legal minimum, and no province was below it. By 2022 the median had fallen to 1.13× and had an average employee wage beneath their own legal floor. A minimum is becoming a maximum.
upah and does not say what it is denominated in. It is read as rupiah per hour. The only evidence is internal: multiplied by the statutory 173-hour month it lands within a quarter of the legal minimum in every province-year, which no other plausible unit does.one province
Somebody lives at the bottom of that spread, and it is the same somebody every year. DI YOGYAKARTA holds the fewest people above the line in — 2.79 people at its worst, in 2016, and 3.53 in 2022. Its minimum wage that year was Rp 1,840,916 a month against a poverty line of Rp 521,672 a person.
A household there of four, with one wage coming in, is below the official poverty line by the arithmetic of the law that set the wage. This is not a discovery about DI YOGYAKARTA. It is the oldest question in wage economics, arriving as division. Adam Smith held that a wage must at least let a worker raise a family or the supply of workers fails. Marx put the same thing as the cost of reproducing the worker and their replacements. Rowntree, counting York in 1901, invented the poverty line and immediately discovered that family size drove everything.
The twist is that Indonesia agrees. The wage formula introduced in 2021 sets the minimum from average consumption per head, multiplied by average household size, divided by the average number of household members who work. That is this ratio, written into law. So none of the above is a standard imported from outside — it is the state’s own arithmetic, handed back.
what the file is called
The dataset is named pekerja sejahtera — the prosperous worker. Nothing in it measures prosperity. It measures distance from destitution, which is a different quantity and a much smaller one. Aristotle separated zēn, living, from eu zēn, living well, and held that only the second was worth organising a city around. A poverty line is an instrument for the first. Amartya Sen made the modern version of the objection: income sits in the space of commodities and well-being sits in the space of what a person is actually able to do, and the rate of exchange between them differs by place — which is one way to read a spread of 2.79 to 8.86 that has not closed in 7 years.
So the honest reading of this file is narrow, and worth stating plainly. It cannot tell you whether an Indonesian worker is prospering. It can tell you that the number meant to answer that question has stopped tracking the thing it is named after, that the legal floor is sinking toward the going rate, and that the basket used to define poverty is one nobody has ever filled. , which is the only claim here the file makes entirely against itself.
the story
A fable. It contains no figures at all — the numbers have had their say already, and what is left of them here is only their shape.
On an archipelago of many islands there were two instruments, and they were built to speak to one another.
The first was kept by the surveyors. Every spring and every autumn they went out to the islands and asked what a month of staying alive costs, and they came back and wrote a number for each island, and the number was different on each one, because the islands were not alike.
The second was kept in the capital, in a bureau with a good view. It said what a month of work was worth. It also gave a number for each island.
The arrangement was supposed to be simple. The surveyors would say what living cost, and the bureau would make sure work paid more than that. One instrument listened; the other answered.
What happened was that the bureau found a rule.
The rule was elegant and it saved a great deal of argument. Take last year’s number, it said, and add to it the amount by which the whole archipelago had grown, and the amount by which everything had got dearer across all the islands together. Do this for every island. The rule was fair in the way that giving everybody the same size of shoe is fair.
And so the bureau stopped needing the surveyors. Not by decree — nobody ever announced it — but arithmetically, which is quieter. Every island’s wage still rose every year, and rose by about the same proportion, and if you laid the wages against the costs you would see them climbing together and conclude that the two instruments were still in conversation. They were not. They were both simply going up.
The wage on each island was no longer an answer to a question about that island. It was an answer to a question about last year.
There was a woman in a small inland city, in a house with four mouths in it and one wage coming in.
Her city was famous. It had a palace and a great many students and a way of speaking that people elsewhere found beautiful, and the rent was accordingly not cheap. Her island’s wage, though, was among the smallest in the archipelago, and had been for as long as anyone had been counting. Not because of any decision anyone could point to. Because of last year, and the year before that, all the way back to a first number nobody now remembered choosing.
She went to the bureau and asked a reasonable question. Her wage was meant to cover a month of living. Whose month? Because there were four of them in the house and only one wage, and if the wage was one person’s month of living then three of them were living on nothing.
The clerk was kind about it. The wage is for a worker, he said. The cost is for a person.
Yes, she said. I know. That is what I am asking about.
Here is the part about the surveyors’ instrument, which was not innocent either.
To say what a month of staying alive cost, they had first to decide what staying alive consisted of. So long ago they had written down a basket. So much rice, so much oil, so much salt and fish and sugar, enough to keep a body upright. Then a smaller allowance for everything that is not eaten — a roof, a lamp, a bus, a school shoe.
The basket was mostly food. Overwhelmingly food. And no household on any island in the archipelago spent its money like that. Even the poorest spent something under half on what it ate, because a roof cannot be declined and a bus fare cannot be haggled down to rice.
So the line marked the cost of surviving as a person who did not exist: someone who ate like the poor and paid rent like nobody. It was not a measurement. It was a decision, costed. And every count of how many people were poor was a count of people standing on one side or the other of that decision.
The surveyors knew this. It was written in their own notes, for anyone who read that far. Almost nobody read that far, because the number at the front was so much easier to quote than the note at the back.
One year — and this is the part of the story that is hardest to believe, and the easiest to prove — the column went wrong.
It was the plague year, when almost every island had held its wage exactly where it was, which was itself a decision worth a great deal of argument and got none. Somewhere between the bureau and the ledger the figures for that year came apart from the islands they belonged to. A third of the islands ended up holding a neighbour’s number. The great port in the west was given the wage of a small inland city and became, on paper, one of the poorest places in the archipelago. A quiet island in the east was handed the port’s wage and became, on paper, the richest place anyone had ever recorded.
Nobody noticed for a long time. This is the crucial thing, and it is not because the clerks were careless.
They did not notice because the middle of the column was still right. The islands had swapped numbers with each other, so the numbers were all still there, and if you asked the ledger what a typical island paid that year it gave very nearly the correct answer. Every summary was fine. Every average was fine. The error was invisible to anyone who asked the ledger a question about all the islands at once, and visible immediately to anyone who asked it a question about one island in particular.
Which is to say the error was invisible to everyone who used the ledger the way the ledger was usually used.
It was found, in the end, the dull way. Someone laid the year against the year before it and the year after it and saw that those two agreed with each other more closely than either agreed with the year in between — and a year cannot do that. A year that sits between two years which match must match them too. That is not a discovery about wages. It is a property of sitting between things.
And once you knew to look, the rest was almost embarrassing. Wages on the archipelago never fell. So a year lower than the year before it and lower than the year after it was not a surprising wage. It was an impossible one, and there were a great many of them.
Some of the displaced numbers were even traceable, sitting one rupiah away from the island they had come from — a rupiah being far too small a sum for a person to have chosen, and exactly the size of the dust a number picks up passing through a machine.
The woman with four mouths and one wage never heard about the column. It made no difference to her; her own island’s figure that year happened to be one of the correct ones, and it was too small either way.
But the two things are the same thing, and it took me a long while to see it.
An instrument that is only ever asked about everyone at once will go wrong about any particular person and give no sign. The bureau’s rule was of that kind: it was built out of quantities that describe an archipelago — how much it had grown, how much dearer it had become — and it produced, faithfully, a number for each island that was about none of them. The surveyors’ basket was of that kind too: a costing of an average that no household anywhere actually is.
The scrambled column was not a different sort of failure from the rule and the basket. It was the same failure, briefly made obvious, because for once the thing that had gone wrong was small enough and local enough to be caught.
The instruments were built to speak to one another about places. They had been left speaking to one another about the archipelago. And an archipelago has no mouths to feed.
Source. Pekerja Sejahtera, Kaggle — rezkyyayang/pekerja-sejahtera. Retrieved 2026-07-30. A mirror of two BPS series: provincial minimum wages 2002–2022, poverty lines 2013–2022, per-capita expenditure 2007–2022, hourly wages 2015–2022. 34 provinces, plus a INDONESIA row among them that is removed. Figures use the maret poverty line, perdesaanperkotaan, with 2021 excluded. Pipeline and committed data: analysis/pipelines/03_pekerja_sejahtera/build.py and web/content/03-pekerja-sejahtera/data.json, both in the repository.
Limits. Dividing a per-worker wage by a per-person poverty line is my analytical decision, not a claim the dataset makes — though it is the arithmetic Indonesian law adopted in 2021. A household of four shares a roof and a stove, so a per-person line multiplied by four overstates what four people cost, and the ratio therefore understates how many a wage can hold; it also assumes one earner, which is why the number of earners is a lever the reader can move. Average household size is not in this file and is never multiplied in anywhere. Nothing here identifies a causal direction: a poverty line is itself partly a function of prices that wages help set, so the fixed-effects results describe whether two published series move together and no more. The two-way estimate rests on 7 usable years, and its interval is wide because the panel is thin, not because the effect is precisely zero. The hourly wage is read as rupiah per hour on internal evidence only. Four of the 13 corrupted 2021 rows cannot be traced from inside the file and are left unattributed rather than guessed. Every figure set in mono is traceable to the committed data file; the fable is fiction, and deliberately contains no figures at all.
The file has 35 rows in its province column. One of them is the country. It is not flagged, not last, and not named differently from the rest — in ump.csv it is simply another row.
Every cross-province figure in this piece — spread, Theil, correlation, the minimum and the maximum — is computed on 34 rows, not 35.
why it matters more than it sounds
A national figure is not an extreme, so it does not announce itself as an outlier. It sits near the middle and quietly pulls every measure of spread inward, makes every correlation look a little tidier, and adds one fake province to every count. The pipeline asserts the row is present and removes it by name; if a future version of the file drops it, the build fails rather than silently changing every number.