How to use the GRIM test to spot fake data

Methodology

Too many people treat the Results section of a paper like a holy text. They see a mean and a sample size and just assume the math adds up. In my experience dealing with local reports, numbers get fudged or rounded poorly all the time. The GRIM test (Granularity-Related Inconsistency Check) lets you spot this in seconds. The logic is simple: if you are averaging whole numbers, the mean must be a multiple of 1 divided by the sample size (N). Let's say a study has 12 people (N=12) answering a yes/no question or a 1 to 5 scale. The mean has to be a multiple of 1/12. If the reported mean is 3.25, that works because 0.25 times 12 is 3. But if the reported mean is 3.21, it is mathematically impossible. You cannot average 12 whole numbers and get 3.21. The process: 1. Find the sample size (N). 2. Take the decimal portion of the reported mean. 3. Multiply that decimal by N. 4. If the result is not a whole number, the data is either sloppily reported or fabricated. It takes the guesswork out of the process. Instead of wondering if a study feels 'off', you can actually prove the numbers do not align.

8 comments

Comments

DevilsAdvocate_Dan·1 hour ago

Suppose a researcher uses a legacy software package that applies a non-standard rounding algorithm. The resulting inconsistency would be a software artifact rather than a manual manipulation.

GrassrootsGreta·1 hour ago

We don't see software artifacts in local reports. If the math is off, it's usually because someone just guessed a number to make the results look better.

HotTakeHarvey·1 hour ago

This isn't a fraud detector; it's a typo detector. Are we really calling every rounding error fabrication now?

SkepticalMike·1 hour ago

Crucial for auditing LLM-generated papers. AI often hallucinates means that look plausible but fail the GRIM test because it doesn't perform the actual summation.

ProfActuallyPhD·1 hour ago

The broader benefit is the push toward reporting raw data in supplementary files. This shift toward transparency inherently discourages the 'massaging' of means before publication.

LurkingLorraine·1 hour ago

it's basically a checksum for behavioral sciences.

MemoryHoleMarcus·1 hour ago

Authors sometimes report weighted means. In those instances, the GRIM test is useless because the denominator is no longer a simple integer N.

QuietOptimistQi·1 hour ago

Since weighted means complicate the math, is there a similar consistency check that works for those specific datasets?