Does GRIM produce a p-value?

No. It checks arithmetic compatibility under specified assumptions.

Does passing prove the mean is correct?

No. Many different datasets and erroneous reports can be compatible with the same rounded mean.

Can missing responses matter?

Yes. The denominator must match the observations actually used for that particular mean.

AI Glossary / Statistics and probability / Intermediate

GRIM test

Definition

The GRIM test, short for granularity-related inconsistency of means, checks whether a reported mean is mathematically compatible with its stated sample size, measurement granularity, and rounding precision. For unweighted integer observations, the total must be an integer, so the mean can only take particular values. GRIM identifies violations of that arithmetic constraint. It is a consistency check, not a conventional significance test and not evidence by itself of intentional misconduct.

What it means

If twelve people each provide a whole-number response, the sum of their responses must be a whole number. Dividing that sum by twelve creates a discrete set of possible means. Rounding those means still leaves some reported decimal values impossible.

The check can therefore reveal a mismatch among a mean, sample size, and stated scoring method without access to the original observations. It cannot identify which component is wrong. A typographical mistake, variable-specific missing data, undocumented weighting, or a different calculation method can explain an apparent inconsistency.

For AI-generated reports, GRIM provides an arithmetic safeguard when measurement rules are known. Passing only establishes compatibility with those rules.

How it works

Identify the number of observations actually included in the mean, the allowed score increments, and the number of reported decimal places. For n unweighted integer observations, a possible exact mean has form S/n for an integer total S within any applicable score bounds.

Check whether at least one possible exact mean rounds to the reported value using the relevant rounding convention. An interval formulation asks whether multiplying the mean's rounding interval by n contains an admissible integer total.

For an equally weighted average across k integer-valued items per participant, with complete responses, the relevant total can have denominator n*k. That simplification does not automatically apply when item counts differ, scores are transformed, or weights vary.

Preserve displayed precision, including trailing zeros, and handle boundary cases and floating-point arithmetic carefully. If the reporting convention is unknown, test reasonable conventions and describe the uncertainty instead of declaring a definitive inconsistency prematurely.

Three practical examples

Research review: check means from small samples using integer rating scales, then verify sample sizes and scoring details before requesting a correction.
Business reporting: examine whether a reported average of whole-number customer ratings is arithmetically possible for the stated number of responses.
AI content quality: screen a generated statistical summary for incompatible counts and means, while also checking its source data, calculations, and interpretation through other methods.

Think of it this way

Think of dividing a collection of whole coins equally among a fixed number of people on paper. The resulting average can have fractions, but the original total must still be a whole number of coins. A reported average that cannot arise from any allowed total signals a mismatch somewhere in the description or arithmetic.

Worked example

Take an original example of twelve unweighted responses, each an integer from one to five. A report gives the mean as 3.43 to two decimal places, using ordinary nearest rounding with positive halfway values rounded upward.

An exact mean rounding to 3.43 must lie in [3.425, 3.435). Multiplying by twelve gives the required total interval [41.10, 41.22). This interval contains no integer, so no allowed response total produces the reported mean under these assumptions.

The neighboring integer total 41 produces mean 41/12 = 3.416666..., which rounds to 3.42. Total 42 produces 3.50. A compatible dataset for 3.42 is seven responses of three and five responses of four: the total is 7*3 + 5*4 = 41.

Exact-decimal enumeration of integer totals from 12 through 60 independently confirms that 3.42 is possible and 3.43 is not. The check cannot tell whether the mean should be 3.42, the sample size is wrong, or the scoring description is incomplete. It establishes incompatibility of the stated combination. No correction should be chosen from this calculation alone.

Illustrative case study

In a fictional feedback project, a report lists a mean rating and a respondent count that fail a GRIM check. The analyst examines the source table rather than accusing the survey team of fabrication. The investigation reveals that the displayed count describes all survey participants, while the rating mean excludes respondents who skipped that item.

The team replaces the count with the variable-specific denominator and reruns its checks. It also updates the reporting template so future summaries include the relevant response count. The fictional case illustrates how an arithmetic inconsistency can reveal an ordinary documentation error and improve reporting discipline.

Common misconceptions and mistakes

Common mistakes include treating a failed check as proof of fraud, assuming a passing check validates the data, and applying the integer-total rule to genuinely continuous measurements without justification.

Weighted averages, prorated scales, transformed scores, and variable item counts can change the allowable values. Repeated rounding at intermediate steps can also complicate reconstruction. The analyst must understand what was averaged before applying a constraint.

GRIM becomes less restrictive when sample sizes or item counts make possible means closer together relative to the reported precision. A report can therefore pass easily while containing other errors. The check also says nothing directly about sample representativeness, causal interpretation, standard deviations, or whether the underlying observations were collected as described.

Business use

Use GRIM as a targeted reporting-quality check where the required arithmetic assumptions are credible. Record the mean as printed, its precision, the effective denominator, scoring rules, and rounding convention. Route inconsistencies to source review and preserve the explanation of any correction. It is particularly useful as part of a broader reproducible reporting workflow rather than an isolated credibility score.

Check your understanding

Answer: A reported mean fails for the overall survey count.

What should you check before concluding that the mean is impossible?

Answer: Verify the variable-specific sample size, item count, weighting, scoring transformations, and rounding convention. A mismatch in any of those assumptions can change the set of admissible reported values.

Related terms

Arithmetic mean
rounding
measurement granularity
summary statistics
reporting error
data validation
GRIMMER test
reproducibility

Learn next

Learn how integer totals constrain averages and how rounding intervals work. Then study consistency checks for other summary statistics, keeping mathematical incompatibility separate from conclusions about its cause.

Sources

Brown and Heathers, The GRIM Test, original publication abstract and record: https://journals.sagepub.com/doi/10.1177/1948550616673876 — foundational purpose and scope; full article retrieval was unavailable. Lukas Jung, scrutiny grim documentation: https://lhdjung.github.io/scrutiny/reference/grim.html — primary implementation guidance on sample size, item count, precision, and rounding. The interval argument and exact-decimal enumeration above are original checked teaching calculations.

Previous: Grid search Next: Grokking