When Data Began to Read People — Who Was Counted, Who Disappeared?

Binding Imperial Quantities without Paper — Inka Khipu

Inka record keepers, or khipukamayuq, encoded households, stored goods, labor, and tribute through the positions, knots, and colors of cords, reporting through the road network. Evidence that several keepers maintained matching accounts also suggests layers of verification and control. Cusco is an editorial anchor for imperial administration, not the production site of one surviving khipu, and not every meaning beyond numerical records has been deciphered.

Reading Urban Regularity from Weekly Death Lists — John Graunt

Commercial demand · lens spotlight

London parishes issued weekly mortality bills for epidemic surveillance; Graunt recombined many years of them. Comparing causes of death, seasons, sex counts, and city size revealed population-level regularities invisible in any one death. Cause labels relied on searchers rather than modern medical examiners and many people were missing, so regularity in a table was neither complete registration nor a causal law.

Joining Parish Tables into a Continuous National Record — Tabellverket

Sweden's Tabellverket regularly combined parish reports of population, births, and deaths, letting the state compare itself over time. Repeated forms, rather than one grand total, exposed rates of change and regional differences. The tables served health and administration but rested on church and state categories and omissions; they were not a neutral modern database.

Counting a Country to Ask How Many It Could Feed — The 1801 Census

Amid war, bad harvests, and debate after Malthus, Parliament passed the 1800 Act and conducted an official census in England, Wales, and Scotland in 1801. Counts of households, people, and broad occupations became numbers for national planning. Household schedules completed by residents took shape only in 1841, and groups such as some soldiers and sailors were excluded, so 1801 was not yet a fully modern census.

Turning an Average from Summary into a Human Figure — Quetelet's Average Man

Quetelet carried error curves from astronomy into averages and regularities in height, crime, and marriage, making the 'average man' central to his 1835 social physics. A powerful way to summarize groups appeared, but treating averages across many traits as one real person or desirable norm erases individual variation and institutional causes. An average answers a question; it is not humanity's master copy.

Layering Death Addresses on a Map to Suspect the Water — John Snow

Snow investigated cholera death addresses and water use in Soho, comparing the concentration around the Broad Street pump with distant drinkers and exceptions such as a workhouse with its own well. The map made the case visible but was not the whole evidence. The outbreak was already waning when the handle was removed, and the map alone did not finally prove germ theory or a single cause.

Seeing a Return in the Cloud of Parents and Children — Galton's Regression

Galton plotted family heights and noticed that children of exceptionally tall or short parents tended, on average, to lie closer to the overall mean, calling the pattern regression toward mediocrity. Reading how two variables move together became a new visual and mathematical project. Correlation does not establish causation, and Galton joined these tools to discriminatory political projects of hereditary ranking and eugenics. Useful mathematics and harmful purpose cannot be separated by celebration.

Turning Human Answers into Holes a Machine Could Read — Hollerith

The U.S. Census transferred each person's age, sex, race, marital status, citizenship, and other answers into positions on punched cards, then used electrical readers and tabulators. Faster processing made it possible to recount combinations of categories and helped launch a data-processing industry. The machine was no more neutral than the questionnaire: political choices had already determined which questions and racial categories became card fields.

Measuring the Mismatch between Observed and Expected Cells — Pearson's Chi-Square

Karl Pearson published a general method that combined differences between observed category counts and model-expected counts into one goodness-of-fit statistic. It moved judgment from 'looks close' toward comparison with sampling variation. It did not single-handedly invent every hypothesis test or today's culture of p-values, and Pearson's biometric methods and institutions were deeply entangled with a eugenic program.

Learning Small-Sample Uncertainty from a Pint of Beer — 'Student'

Commercial demand · lens spotlight

At Guinness, William Gosset faced barley and malt experiments where only a few observations were practical and population variance was unknown. His 1908 paper under the name 'Student' described a distribution whose heavier tails change with sample size. The resulting methods still rely on assumptions such as independence and a chosen population model; they do not turn a small convenience sample into a representative one.

Letting Chance Choose the Plot instead of the Researcher — Rothamsted

Fisher and Rothamsted colleagues combined replication, blocks, and deliberate random allocation to separate fertilizer or crop effects from irregular soil. Chance supplied a reference comparison instead of letting the investigator pick plots that looked favorable. The 1926 paper is a landmark formulation, not the absolute first randomized experiment, and Fisher's active support for eugenics remains part of the history.

When Millions of Replies Lost to a Smaller Poll — Literary Digest and Gallup

Commercial demand · lens spotlight

The Literary Digest mailed roughly ten million ballots and received more than 2.3 million replies, yet wrongly forecast a Landon victory. Its frame and response process were both biased. Gallup's much smaller quota sample, run from Princeton, correctly called Roosevelt's victory. This was not a simple triumph of modern probability sampling: the Gallup design used quotas and its method suffered a major failure in 1948.

Separating Treatment Hope from Comparison — The MRC Streptomycin Trial

The British MRC tuberculosis trial used central allocation based on random numbers to assign scarce streptomycin and compared a treatment group with a bed-rest control group. Radiograph readers assessed images without knowing allocation. It was a major transition in clinical trials, but not the first controlled trial and not double-blind in every respect. A statistical difference also does not promise the same benefit to every patient.

Trying to See Regional Difference without Counting an Entire Country — India's National Sample Survey

Commercial demand · lens spotlight

Mahalanobis and the Indian Statistical Institute network launched the National Sample Survey in 1950 to repeatedly measure consumption, work, land, and production across rural and urban India. Stratified multistage designs and field organization produced timely planning evidence for a large and diverse new nation. A sample survey is not merely a cheaper miniature census but a separate design carrying uncertainty, and a national average cannot stand in for every region or group.

When the Overall Rate and Department Rates Pointed Opposite Ways — Berkeley Admissions

In UC Berkeley graduate admissions data, women's overall admission rate appeared lower, while department-level comparisons greatly reduced or sometimes reversed the gap. Applicants had entered departments with very different selectivity at unequal rates, allowing aggregation to reverse a pattern. The analysis was not an automatic verdict that discrimination was absent. Which variables to condition on depends on a causal question, including constraints shaping department choice.

Recounting the Faces Hidden under Average Accuracy — Gender Shades

Joy Buolamwini and Timnit Gebru evaluated commercial gender classifiers separately across intersections of perceived skin type and gender. Large performance gaps and the lighter-skinned, male-heavy composition of existing benchmarks became visible beneath one average accuracy. The study audited four classifiers and constructed datasets at a particular time; it was neither a theory of every face-recognition system nor a justification for reducing gender identity to a binary classification task.

Using Mathematics to Obscure the People Revealed by Publication — Differentially Private Census Tables

The U.S. Census Bureau applied a differential-privacy-based disclosure avoidance system to parts of the 2020 Census release to limit the risk of combining tables to reidentify people. Adding controlled random noise makes a privacy budget explicit but creates tension with accuracy, especially for small places and groups. Differential privacy is a framework for designing risk and utility together, not one magic algorithm that guarantees anonymity.