A cross-tabulation is the table that answers "did the groups differ?" — the single most common question asked of survey data, and the one a frequency table cannot answer. Everything else about the technique follows from that.

This is a working guide: what the table is made of, how to read the significance marks that make it defensible, and the four mistakes that produce a table that is wrong in a way nobody notices.

What a cross-tabulation actually is

A frequency table tells you 62% chose "reliability". A cross-tabulation splits that 62% by a second variable — gender, region, customer tenure — and shows you that it is 71% among long-tenure customers and 48% among new ones. The first number describes your sample. The second tells you something you can act on.

The vocabulary is worth fixing early, because the same table has different names in different rooms. Statisticians call it a contingency table. Market researchers call it a crosstab, and call the whole deliverable banner tables or a tab book. They are the same object.

Two parts define it:

  • The banner — the variable across the columns. This is the cut: who you are comparing. Gender, age bracket, segment, wave.
  • The stub — the variable down the rows. This is what you are measuring: the answers to one question, one row per category.

A commercial tab book is one banner applied to every question in the study, so it is one table per question, all cut the same way. That is why a fifty-question survey produces a fifty-sheet workbook rather than fifty separate analyses.

How to read one

Here is the shape of a single table, with the parts labelled:

Q7. Main reason for choosing the brand
Significance: 95% confidence

                    Total      Male      Female     18-34      35+
                                 (A)        (B)       (C)      (D)
Base (N)              850       410        440       295       555
Reliability         62.0%     58.0%      65.7%     51.2%    67.7% C
Price               21.4%     26.1% B     17.0%     29.5% D   17.1%
Design               9.9%      8.5%      11.1%     13.2%      8.1%
Other                6.7%      7.4%       6.2%      6.1%      7.1%

The base is how many people the percentages are computed on — not how many people are in the study. A column whose base is 40 is describing forty people, however confident the percentage looks.

The letters are the significance test. Each column gets one. A letter inside a cell means this column is significantly higher than the column that owns that letter, at the stated confidence. So 26.1% B in the Price row means men chose price significantly more often than women. And 67.7% C means the 35+ group beat the 18-34 group on reliability.

Read the letters, not the gap. A four-point difference on a base of 800 can be real; a fifteen-point difference on a base of 30 usually is not. The letters encode that arithmetic so you do not have to redo it for every cell.

What the significance test is doing

Between two columns, the standard test is a two-proportion z-test: it asks how likely it is that two samples this different came from populations that are the same. At 95% confidence, a mark means that likelihood is under one in twenty.

Three things about that number are worth knowing before you put it in a deck.

It is a statement about sampling, not importance. With a large enough base, a two-point difference becomes significant. Significant and meaningful are different words; the table can only tell you the first.

Small bases should not carry letters at all. Below roughly thirty cases the test has almost no power, and any mark it produces is noise with a decoration. Good tabulation suppresses letters under a threshold and flags the column instead.

Many comparisons manufacture findings. A fifty-question book with a six-column banner runs thousands of tests. At 95%, one in twenty comes back significant by chance alone — so a handful of the marks in any large tab book are accidents. This is why researchers form the hypothesis before reading the tables, not after. Some houses apply a Bonferroni correction; most do not, and instead treat isolated marks with suspicion and repeated patterns as findings.

Four mistakes that survive review

1. Reading percentages off the wrong base

Column percentages and row percentages answer different questions. "62% of women chose reliability" and "54% of those who chose reliability are women" are both true and are not interchangeable. Most tab books are column-percentage by convention; a table that mixes the two without saying so is unreadable.

2. Treating multi-response questions like single-response ones

When respondents can pick several answers, the column sums past 100% and the base is respondents, not mentions. A table that divides by mentions produces percentages that look tidy and describe nothing anyone asked about.

3. Letting the base drift without noticing

Filtered questions change the base — "asked only of those aware of the brand" is a different universe from the full sample. The base row is the only thing on the sheet that reveals it. When the base moves between two tables, comparing their percentages directly is comparing two different populations.

4. Cutting until something turns significant

Adding banner points until a difference appears is the most common way to manufacture a finding from clean data. The remedy is procedural rather than statistical: decide the banner before the fieldwork closes, and treat anything found by extra cutting as a hypothesis for the next wave, not a result in this one.

Weighting, briefly

When the sample does not match the population — too many young respondents, too few in one region — weights correct the imbalance by giving each respondent a multiplier. The percentages then describe the population rather than the sample.

The cost is precision. A weighted base of 850 behaves statistically like a smaller unweighted one, and the significance test has to use that effective base rather than the headline number. A tab book that weights the percentages but tests on the raw base overstates its own certainty. If a supplier cannot tell you which base their letters used, that is the question to press.

Producing them

SPSS, R and Stata all cross-tabulate, and for one table they are fine. The reason market-research shops use dedicated tabulation software is the shape of the real job: one banner, fifty questions, significance on every cell, delivered as a formatted workbook a client can read without the analyst present. Doing that by hand is not analysis, it is assembly.

If you have a .sav file and want to see the shape of that deliverable, our free cross-tabulation tool builds it in the browser: pick the banner variable, and it returns an Excel workbook with one sheet per question, bases, column percentages and 95% significance letters. No install, no account, and nothing is stored.

Where this fits

Cross-tabulation is what happens after the data is coded. Closed questions arrive ready; open-ended ones do not — "what do you like least about the brand?" has to be turned into categories before it can be a stub at all. That step is its own discipline, with its own failure modes, and it is covered in our guide to survey coding and in open-end coding in market research.

A tab book is only as good as the codeframe underneath it. A beautifully tested table built on categories that merge two different complaints is a precise answer to the wrong question.