A semester of quantitative-methods homework, several unrelated assignments, one recurring blind spot. None of these was the featured assignment — they were routine weekly problem sets, the kind you complete and move past. Reread together, years later, they’re the same lesson three times: a number travels badly across units, currencies, scales, and populations unless someone deliberately carries it across.
Everyone treats a number as if it explains itself. “$4.7 million.” “37% of the variance.” “90% of respondents.” Said with enough confidence, a number sounds like it’s already done the work of meaning something. It hasn’t. A number is only as honest as the unit, the scale, and the population attached to it — and those are exactly the details a default assumption erases first, usually without anyone noticing, including the person who ran the analysis.
Conversion was personal before it was statistical
I didn’t meet this problem in a statistics course. I met it in a grocery store, during an exchange semester in the United States: milk sold by the gallon, produce priced per pound, a bottle labeled in fluid ounces — every purchase a small arithmetic problem between the number on the shelf and the quantities I’d grown up trusting.
A few years later I was back in the U.S. as a chemical engineering graduate student, and unit conversion was — literally — the first thing we had to learn. Not a warm-up. The curriculum. In American engineering practice the same word pound names two different physical quantities: pound-mass (lbm) and pound-force (lbf), and fluid-mechanics equations carry a constant, g꜀ = 32.174 lbm·ft/(lbf·s²), whose entire job is to keep the two pounds from colliding. The BTU — still the working unit of American energy engineering — is defined as the heat needed to raise one pound-mass of water by one degree Fahrenheit. Even the gallon isn’t one thing: a U.S. gallon and a British imperial gallon are different volumes wearing the same name.
My computer science friends never carried this weight. A byte is a byte in Bangalore and in Boston; a millisecond has no imperial cousin. Software is young enough that its units were born standardized. Chemical engineering is older than the metric treaty, and its formulas still show the seams. By the time these problem sets found me, I’d spent a decade professionally suspicious of units. That should have inoculated me. It didn’t.
The housing price that didn’t add up
One weekly assignment worked with a housing dataset — 545 homes, a mean price of $4,766,729, a median of $4,340,000. Nothing about the numbers themselves signaled a problem; regression-ready datasets rarely announce their own units. My own answer to the question “how does this compare to your neighborhood” hedged in a way I didn’t notice until I reread it:
“If we consider that the dataset’s prices are in USD …”
— and my answer went on to dutifully compare those millions against a typical price from a neighborhood I know well. The comparison isn’t the point; the hedge is. If we consider. I wrote the conditional into my own sentence and then never resolved it. A median home price near $4.3 million is not a typical American housing market, and the dataset gave no currency field to check the assumption against. The exercise — mine included — treated “the numbers are in dollars” as the default reading, because that’s the reading that requires no further work. Nobody in the assignment was asked to verify it, so nobody did, including me, and I’m the one who left the hedge sitting unresolved in my own answer.
The error I made explaining my own numbers
The one place this should have been impossible for me to get wrong: my own final paper, where I was the one doing the explaining. Writing up a dataset on India’s rural employment guarantee scheme for an American classroom, I translated the currency unit myself, in my own words:
“1 lakh is equivalent to ten million, and 1 lakh Indian rupees, based on the current conversion rate, roughly amount to 1200 USD.”
One lakh is one hundred thousand. Ten million is a crore — one hundred lakh, not one. I was translating a number system I’ve used my entire life for an audience encountering it for the first time, and I still shipped the conversion wrong, in the one paragraph where precision was the entire point. It’s not a typo I can blame on the dataset this time; there was no ambiguous currency field to hide behind. I was the native speaker of these units, mid-translation, and the error made it into the final draft anyway.
The average that hid the real story
A smaller version of the same problem showed up in a consumer-behavior assignment: the mean income in that dataset was $52,247, the median was $51,381.50. Close enough to ignore — until the follow-up question asked what the gap implied, and the honest answer was that the mean sitting above the median meant the distribution had a long right tail: a shrinking number of very high earners pulling the “average” upward, away from what a typical person in the dataset actually made. Report only the mean, and “typical” quietly gets replaced by “average,” which is a different number describing a different question.
And one more, lighter version: a mental-health-in-tech survey I worked with that week turned out to draw over 90% of its respondents from five countries — Germany, the U.S., Australia, the U.K., and Canada. Nothing in the framing ever called the sample “global,” but nothing corrected the impression either. A dataset can be broad enough to feel representative and still be mostly the same five countries answering for everyone else.
What transfers
Sometimes it seems I have converted so much that, by sheer volume, some failure was simply due. Kilograms to pounds in the grocery store, pound-mass to pound-force in the fluid-mechanics homework, USD to INR whenever money crossed the ocean — I have converted enough for a lifetime. That’s the comfortable reading of the lakh-to-crore error: run enough conversions and eventually one ships broken. The less comfortable reading is that somewhere along the way I gave up converting. Fluency in two systems doesn’t make the seam between them disappear; it makes the seam familiar enough to stop watching — and at some point I stopped watching. The deliberate carrying-across became a wave-through. I think it’s time to get back to being fluent in the working sense: not the fluency that stops checking, but the fluency that still does the conversion every time.
My throughline is negotiating for clarity, and I’d always applied it to language — every sentence gets renegotiated in real time. Rereading those answers extended the same discipline to numbers. A number crossing a border needs exactly the same negotiation a sentence does: what’s the unit, what’s the scale, whose experience does “average” actually describe, and who’s missing from the 90%. I caught the assumption sitting quietly in someone else’s housing dataset back then. I did not catch it in my own paragraph until I went looking for this piece.
Thanks for reading. Next time a number lands with real confidence — a headline stat, a dataset’s summary row — it’s worth asking what unit it’s actually in, what currency, what scale, and whose average. I found my own blind spot doing exactly that. What was the first unit that ever betrayed you?
Comments