1890: Tabulating the state

A census turns descriptions of people into categories that an institution can count. Hollerith's machines used punched records to automate parts of that work for the 1890 United States census. The Census Bureau's account explains the operational setting; IBM's institutional history traces the later business lineage. Neither source makes the machine a neutral observer. Someone still decided which questions to ask, how to classify answers, and which totals would influence public decisions.

The technical boundary was an encoding contract. A location on a card represented a particular attribute or category. A sensing mechanism detected holes; electrical counting machinery accumulated selected events. Once records followed the same format, machines could process many people without understanding their names or circumstances. Sorting and repeated tabulation supported different summaries from the same captured data. This is an early example of separating a persistent representation from the immediate question asked of it.

Consider a deliberately simplified ten-card dataset with two binary fields: district, eitherA orB, and employment, either employed or not. Six records belong toA; four of those are employed. Four records belong toB; three are employed. The table reports employment counts 4 and 3, but rates 4/6 and 3/4. A machine can produce both accurately while a reader incorrectly concludes thatA has the better rate because four exceeds three. Fast aggregation does not resolve denominator choice or category validity.

Now add one incorrectly punched district field. The total population may still be ten, so a control total does not necessarily detect the error. Independent verification, range checks, duplicate detection, and reconciliation answer different questions. Record representation makes processing repeatable, but also makes a mistake repeatable at scale. An omitted person is invisible to a check performed only over the submitted cards.

Commercial adoption required more than a counter. Customers needed card preparation, installation, training, supplies, maintenance, and confidence that the workflow would remain supported. A lease could lower an initial purchase barrier while creating recurring dependence on the supplier's format and service. An open encoding specification might permit substitute equipment, but compatibility would still require mechanical tolerances, field definitions, and operational testing.

The company chronology matters. Charles Flint organized the 1911 combination that included Hollerith's enterprise; Thomas Watson joined C-T-R in 1914. Attributing the acquisition to Watson confuses later leadership with the formation event. Corporate retrospectives are useful records of dates and self-description, but claims about inevitable progress or superior management require additional evidence.

The transferable lesson is precise: an information system combines representation, machinery, labor, procurement, and authority. To evaluate it, trace who chooses categories, who corrects records, who can inspect the process, and who can replace a component. Throughput is only one property of a trustworthy census.

Reconstruction exercise

Using the ten-card example, calculate district employment rates and the overall rate. Then suppose one employedA record is miscoded asB. Recompute both rates and explain why the population control total still passes. Propose one check that could detect a coding error and one source of error it cannot detect. Finally describe what an open card format must specify beyond the physical placement of holes.

Worked solution

The rates are 4/6≈66.7%,3/4=75%, and 7/10=70% overall. After the miscoding,A has 3 employed among 5 andB has 4 among 5, giving 60% and 80%; the total remains ten and total employment remains seven. Comparing punched fields with the original questionnaire can reveal this coding error, but cannot prove that the original answer was correct or that every resident was enumerated. An open format needs field meanings, permitted values, missing-value conventions, card orientation, and mechanical tolerances, plus test cases that independent machines process consistently. Openness creates a possibility of replacement; it does not eliminate verification work.

Test the limits and transfer the idea

1. Calculate and explain · allow 25 minutes

When correct subgroup counts reverse the headline

A fictional census reports employed/residents for two districts. In the younger band, A has 81/90 and B has 19/20. In the older band, A has 1/10 and B has 24/80. Compute each band's rates and each district's aggregate rate. A report claims that A has better outcomes in every group because its aggregate rate is higher. Diagnose the claim, then calculate both districts' standardized rates using a population that is half younger and half older. These are constructed numbers, not 1890 census observations.

Get a starting hint

Keep the denominator for each group visible; then give both districts identical group weights.

Compare your reasoning with the worked solution

A's younger rate is 90%, below B's 95%; its older rate is 10%, below B's 30%. Aggregation gives A 82/100=82% and B 43/100=43%, reversing the comparison because A's records are concentrated in the higher-employment age band. Equal weighting gives A (90%+10%)/2=50% and B (95%+30%)/2=62.5%. The tabulation need not contain an arithmetic error: the inference changes the question from subgroup rates to rates in differently composed populations. Standardization answers a specified comparison, not every causal question. It does not establish that changing district would change a person's employment, because selection and other unmeasured differences remain.

Check your reasoning

• Calculates all four subgroup and both aggregate rates.
• Explains the reversal through different population weights.
• Computes standardized rates and avoids a causal inference.

2. Trace and source criticism · allow 25 minutes

A correct tabulator can count the wrong population

A card file contains 1000 cards. Ten people each appear twice, and ten different residents are omitted. All remaining residents appear once. The true population is also 1000. Explain why a population control total passes. Design a reconciliation that detects both defects and state which historical claim would need stronger evidence: that Hollerith equipment counted encoded cards, or that it counted every resident accurately. Use the Census Bureau machine history as a later institutional account and IBM's history as a corporate chronology, not as original enumeration logs.

Get a starting hint

Count distinct represented people before proposing checks; a duplicate check cannot inspect an absent record.

Compare your reasoning with the worked solution

The file represents 990 distinct people: 980 occur once and ten occur twice, producing 1000 cards. Ten true residents are absent. A sum of cards therefore matches the true total by cancellation. A stable-person-identifier duplicate check can expose the repeated records if identifiers themselves are trustworthy. Detecting omissions requires comparison with independent address, household, or field-enumeration evidence, followed by adjudication rather than blindly treating either list as complete. The machine history supports how encoded records were processed. Universal accuracy would require surviving enumeration instructions, coverage studies, reconciliation logs and evidence about classification errors. IBM's later business chronology cannot supply those operational facts merely because its corporate lineage includes Hollerith's company.

Check your reasoning

• Shows how ten duplicates and ten omissions cancel in the total.
• Separates duplicate detection from independent coverage checks.
• Distinguishes retrospective institutional descriptions from operational primary records.

Check your understanding

One census card is assigned to the wrong district, but the population total still matches. What does this show?

• Matching population proves every field is correct.
• The source questionnaire must have omitted a person.
• The tabulator must be counting incorrectly.
• A control total can pass while category assignments are wrong.

Explanation

Moving a record between districts preserves the total while changing subgroup counts and rates.

A competing tabulator reads holes in the same locations but interprets one field differently. What is missing from compatibility?

• A larger supply of blank cards only.
• A guarantee that every customer leases equipment.
• A faster counting mechanism only.
• A shared semantic contract for field meanings and allowed values.

Explanation

Physical format agreement does not establish agreement about what encoded facts mean.

Sources and further reading

• The Hollerith machine: United States Census Bureau
• The origins of IBM: IBM institutional history