Roulette Lab · Lab · Statistics · about 4 min read

Chi-Square Test

The standard test for “are these 37 counts what a fair wheel would produce?”

Short answer

χ² = Σ (observed − expected)² / expected over all 37 pockets. Compare with the chi-square distribution on 36 degrees of freedom to get a p-value.

statistic
Σ (O−E)²/E
df
36
critical at 5%
51.0
critical at 1%
58.6

Computing it

Expected count per pocketequipmentPocketA compartment on the rotor in which the ball can settle. The number of pockets depends on the wheel.Read full definition → is n/37. For each pocket, square the difference between observed and expected and divide by expected; add the 37 terms. Under the fair, independent-spin model, the statistic has mean 36 (the degrees of freedomlabDegrees of freedomThe number of unconstrained components used in a statistical reference distribution. For fixed-total counts in 37 specified categories, it is 36.Read full definition →). With adequate expected counts, its reference distribution is approximately chi-square, with a standard deviationmathStandard deviationThe square root of variance: a measure of dispersion in the same units as the result.Read full definition → of about 8.5. The approximate upper-tail cutoffs are 51.0 at the 5% level and 58.6 at 1%. The Wheel Bias Analyzer computes the statistic and a numerically computed approximate p-valuelabp-valueUnder a specified null model, the probability of a test statistic at least as extreme as the observed one, using the chosen test.Read full definition → from the regularized incomplete gamma function.

n spinsexpected per pocketσ per pocketa pocket at +2σa pocket at +3σ
370103.11619
1,000275.13742
3,7001009.9120130
10,00027016.2303319
37,0001,00031.21,0621,094
For one pocket under the independent-spin model, σ = √(n·p·(1−p)) with p = 1/37. These counts describe individual pockets; selecting the largest count after inspecting all 37 requires a multiple-testing adjustment or an independent confirmation sample.

A worked example

Consider a hypothetical sample of 1,000 independent spinsmathIndependent spinsSpins for which learning an earlier outcome does not change the probabilities assigned to a later one.Read full definition →. Each pocket has an expected count of 1,000/37 ≈ 27.03. If pocket 8 appears 41 times, its contribution to the statistic is about 7.22. Suppose the total statistic is approximately 45.2. On 36 degrees of freedom, the approximate p-value is 0.14: under the fair-wheel model, a statistic at least this large has a probabilitymathProbabilityA measure of how likely an event is, from zero to one. Percentages express the same scale from 0% to 100%.Read full definition → of about 14%. This is not significant at the 5% level, and it does not prove the wheel is fair. A pocket selected after examining these counts needs independent confirmation; another 1,000 spins is not guaranteed to resolve a small departure.

Assumptions

The usual approximation assumes independent observations and adequate expected counts. The common five-per-cell rule gives n ≥ 185 for 37 equal pockets; it is a guideline for the approximation, not a hard boundary for all possible tests. Pooling wheels changes the population being tested. Dependence may invalidate the reference distribution without necessarily increasing the statistic.

Roulette Dealer SignatureIndependent Roulette Spins

What the test establishes

The test is named for Karl Pearson and is often illustrated with roulette counts. Historical stories about particular records require separate source verification; the calculation does not depend on accepting those stories.

An imagined record-keeper illustrates why statistical tests need reliable observations. An unusual result alone does not identify its cause.

Monte Carlo Roulette

Interpreting an automatic-wheel study

A published analysis of specific automatic wheels is evidence about those recorded samples, not all automatic roulette. A non-significant uniformity test cannot prove the absence of steering, and combining several wheels does not identify the behavior of each one. Check the original study’s data and design before drawing that conclusion.

Beyond chi-square

A test of one pocket chosen in advance can be more sensitive to that specific alternative than an omnibus test. SectorwheelSectorA continuous arc of neighboring pockets around the wheel, rather than a group defined by the betting layout.Read full definition → tests examine a different hypothesis. If you choose a pocket or sector after a broad test, account for selection or use independent confirmation data.

Sector Analysis

Formula

χ² = Σᵢ (Oᵢ − Eᵢ)² / Eᵢ
Eᵢ = n / 37 · df = 36
critical values: 51.0 (5%) · 58.6 (1%)

Questions

Why 36 degrees of freedom?

The 37 counts sum to the fixed sample size. Once 36 counts are known, the last is determined. With the pocket probabilities specified in advance, that leaves 36 degrees of freedom.

Does a large statistic prove physical bias?

It challenges the specified uniform independent model; data errors and other model failures also need consideration.

Can I run it on 200 spins?

The expected count is about 5.4 per pocket, meeting the common approximation guideline. Power against small departures remains limited.

Comic relief

For 36 degrees of freedom, the approximate 5% upper-tail cutoff is 51.0.

For 36 degrees of freedom, the approximate 5% upper-tail cutoff is 51.0.

Sources

  1. NIST, chi-square goodness-of-fit test https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/chsqgood.htm
  2. American Statistical Association, statement on p-values (2016) https://www.amstat.org/asa/files/pdfs/P-ValueStatement.pdf
Default model: European, single zero; exceptions are stated in the articlemethodology ↗sources ↗