Can both margins be random when n is fixed?

Yes; multinomial sampling provides a standard example.

Are observed margins uncertain after collection?

Their recorded values are known, but the statistical model describes how they vary across repetitions.

Does conditioning change the original recruitment?

No. It changes the reference analysis, not the history of collection.

AI Glossary / Statistics and probability / Intermediate

Random marginal totals

Definition

Random marginal totals are row or column totals in a contingency table whose values are not fixed by the sampling design and can vary across repeated data collection. A contingency table classifies observations using two or more categorical variables; its margins summarize counts for individual variables. Whether a margin is random depends on how observations are obtained, not simply on the appearance of the completed table. A fixed overall sample size can coexist with random row and column totals.

What it means

Suppose a survey selects a set number of people and records both their customer category and their response. The number falling into each category is discovered during collection, so the category totals vary from sample to sample. Those margins are random even though the survey's total size was chosen in advance.

By contrast, a study that deliberately recruits specified numbers from each category fixes those category totals by design. Recognizing this distinction helps identify the appropriate probability model and the population quantities that the data can estimate. It also prevents observed counts from being mistaken for design constraints.

How it works

Start with the collection plan. Identify which counts were set before observation and which emerged from the observed classifications. Under multinomial sampling, the overall total can be fixed while counts across cells, and therefore the margins, vary. Under independent Poisson sampling, even the overall total may vary.

If separate samples of fixed sizes are taken within row categories, row totals are fixed while outcomes within each row vary. This commonly leads to a product-multinomial model. An analysis can also condition on observed margins, as in familiar exact-table procedures, even when those margins were not physically fixed by design. Conditioning is an analytical choice that must be distinguished from the original sampling mechanism.

Choose estimands and uncertainty calculations that fit the design. A table's numerical entries alone cannot reveal whether its margins were fixed, random, or conditioned upon. Preserve recruitment rules and sampling documentation alongside the table, especially when data are reused for a different business question.

Three practical examples

1. A fixed-size random customer survey records subscription type and renewal intention; both sets of category totals emerge from the sampled customers.

2. A study recruits exactly 60 standard and 40 premium customers, fixing the subscription-type margin while allowing satisfaction counts to vary.

3. A support center counts all incoming tickets during a time window and classifies their topic and urgency; the total number of arrivals and the margins may all vary.

Think of it this way

Think of filling a tray with colored counters. Choosing 100 counters without prescribing their colors fixes the total but leaves the color counts uncertain. Demanding exactly 40 blue counters and 60 red counters fixes that margin. Looking at the final tray cannot tell you which selection rule was used.

Worked example

Consider an original model in which 100 independently sampled customer records are classified by plan type and satisfaction. The population cell probabilities are stipulated as standard/satisfied 0.30, standard/not satisfied 0.10, premium/satisfied 0.40, and premium/not satisfied 0.20. These sum to one.

The probability of a standard customer is 0.30 + 0.10 = 0.40. Consequently, the standard-row total has a binomial distribution with n = 100 and p = 0.40. Its expected count is 40 and its variance is 100 times 0.40 times 0.60 = 24, giving standard deviation approximately 4.899. The expected satisfied-column total is 100 times 0.70 = 70.

Neither total must equal its expectation in the observed sample. Because the overall total is fixed at 100, the premium-row count is 100 minus the standard-row count; the two row totals are not independent.

Now change the design to recruit exactly 40 standard and 60 premium customers. The standard-row count no longer has sampling variance under that design: it is always 40. Satisfaction within each row remains variable. This change concerns the sampling mechanism, even if one realized table happens to look identical under both designs.

Illustrative case study

In a fictional software company, analysts receive a table comparing satisfaction by subscription tier. They initially assume it came from a single random sample of users. The research team explains that it deliberately recruited equal numbers from each tier.

The analysts revise their interpretation. They can compare satisfaction within the recruited tiers under the study's assumptions, but the table's tier totals do not estimate how common each tier is in the customer population. They obtain population tier counts before constructing an overall estimate. The case illustrates why sampling documentation is part of the data rather than optional background information.

Common misconceptions and mistakes

Do not equate a fixed grand total with fixed margins. Do not call a margin random merely because its value was unknown to the analyst; the relevant question is whether it could vary under the design or model.

Random margins need not be mutually independent. Conditional tests do not prove that margins were fixed during recruitment. Similar-looking test statistics can arise under different sampling schemes, but that does not make all interpretations interchangeable. Sparse cells, dependence between observations, survey weights, and selective recruitment can also require methods beyond a simple independent contingency-table analysis.

Business use

Businesses use cross-tabulations to compare customers, products, incidents, and outcomes. Knowing which margins are random helps prevent incorrect claims about population composition and inappropriate uncertainty estimates. Store the sample design with dashboard definitions, and distinguish category-specific results from overall population results. When recruitment quotas are used, explain which totals were imposed and how any combined estimate was weighted.

Check your understanding

A study recruits exactly 25 participants from each of four regions. Is the regional margin random under that recruitment plan?

Answer: No. Those regional counts are fixed by design, although measured outcomes within each region can vary.

Related terms

Contingency table
row total
column total
multinomial sampling
product-multinomial sampling
Poisson sampling
conditional inference
Fisher's exact test

Learn next

Learn how contingency tables summarize categorical data, then compare sampling schemes before studying tests of independence and association.

Categorical data

Sources

David Howell, Statistical Methods for Psychology glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

Penn State, STAT 504, Two-Way Tables: Independence and Association: https://online.stat.psu.edu/stat504/Lesson03

The author glossary was read directly; Penn State's indexed sampling-scheme explanation was consulted, while direct page retrieval timed out. The numerical model and fictional case are original.

Previous: Random indexing Next: Random model Anova