Sample size is the number of observations used to estimate, compare, or test something. In gambling analysis, an observation might be a hand, spin, roll, wager, session, player trip, machine-day, or transaction.
The symbol (n) is commonly used for sample size. A larger (n) usually improves precision, but only when the observations and measurement method fit the question.
Count the correct unit
“Ten sessions” is not the same sample as “10,000 spins.” A session can contain a different number of decisions and a different mix of wagers.
Before using a sample, define:
- the observation unit;
- the population of interest;
- the time window;
- inclusion and exclusion rules;
- whether observations are independent;
- whether game rules or configuration changed;
- what outcome is being measured.
A casino comparing dealer error rates might use completed hands, supervised hours, or incidents. Each denominator answers a different question.
Why larger samples are more stable
For many estimators, standard error decreases approximately with the square root of sample size:
[ \text{Standard error}\propto\frac{1}{\sqrt{n}} ]
This means quadrupling the sample roughly halves the standard error. Doubling the sample does not double precision.
If an estimate has a standard error of 4 units at (n=100), a similar process may have a standard error near 2 units at (n=400), assuming independence and stable variability.
A sample does not force the expected result
A roulette wheel does not owe a balanced color count after 100 spins. A slot does not have to return its theoretical RTP over one session. Expected value describes the average across repeated trials under the model, not a schedule the game must follow.
A larger sample makes extreme percentage differences less common, but it never guarantees the exact expectation.
For the underlying distinction, see Long Run and Short-Term Variance.
Sample size and a proportion
Suppose 52 of 1,000 observed transactions require correction. The observed correction rate is:
[ \hat p=\frac{52}{1{,}000}=5.2% ]
A rough large-sample standard error for a proportion is:
[ SE(\hat p)=\sqrt{\frac{\hat p(1-\hat p)}{n}} ]
Using (\hat p=0.052) and (n=1{,}000):
[ SE\approx\sqrt{\frac{0.052\times0.948}{1{,}000}}\approx0.0070 ]
A rough 95% interval using (1.96\times SE) is about 3.8% to 6.6%. Exact or improved methods may be preferable for small counts or extreme proportions. NIST explains exact binomial confidence intervals and cautions that simple symmetric approximations can be inaccurate for small samples in its confidence-interval guidance.
Independence can matter more than the raw count
One thousand observations generated by the same unresolved system fault may not provide the information of one thousand independent events. Clustered data can exaggerate effective sample size.
Examples include:
- many transactions from one cashier and one shift;
- repeated spins from the same machine configuration;
- players nested within one promotion;
- multiple hands from one high-limit session;
- hourly readings from a strongly autocorrelated process.
Analysts may need clustering, stratification, time-series methods, or an effective sample size smaller than the raw count.
Mixing populations can hide the answer
Combining unlike games, denominations, shifts, or customer groups can produce a large but misleading sample.
A floor-wide slot hold percentage may look normal while one software version is abnormal. A cage-wide transaction time may hide a severe delay in marker settlement. A combined dealer-error rate may punish a team assigned to the most complex game.
Stratify when the groups have different rules, risk, variability, or operating conditions.
How much sample is enough?
There is no universal number. Required sample size depends on:
- desired margin of error;
- confidence level;
- variability;
- minimum effect worth detecting;
- event rarity;
- sampling design;
- cost and feasibility.
For estimating a mean with known standard deviation (\sigma), a planning formula is:
[ n\geq\left(\frac{z_{1-\alpha/2}\sigma}{E}\right)^2 ]
where (E) is the desired margin of error.
If (\sigma=20), the desired 95% margin is 5, and (z=1.96):
[ n\geq\left(\frac{1.96\times20}{5}\right)^2\approx61.47 ]
Round up to at least 62 observations, assuming the planning assumptions are reasonable.
Rare events need much larger samples. Observing zero incidents in 100 cases does not prove the true rate is zero.
Casino examples
Game performance
A machine’s realized hold should be reviewed with coin-in, number of decisions, volatility, jackpot exposure, paytable history, and meter quality. Percentage alone is incomplete.
Player rating
A short rating based on a few wagers can misstate average bet and theo. More observed decisions can improve the estimate, but only if the rating method is consistent.
Operational quality
A department with two errors in 50 transactions has a 4% observed rate; another with 30 errors in 10,000 has 0.3%. The first percentage is based on much less evidence.
Promotion testing
A campaign may appear successful because a few high-value players responded. A holdout group and enough qualified customers are needed to estimate incremental effect.
Bigger data does not repair bad data
A million duplicated or misclassified records can be worse than a smaller clean sample. Review:
- missing observations;
- selection bias;
- changed definitions;
- duplicate records;
- measurement error;
- survivorship bias;
- excluded failures;
- data entered after the outcome was known.
Sample size is one dimension of evidence quality, not the whole standard.
The practical definition
Sample size is the number of observations supporting a result. Larger samples generally improve precision, but the improvement depends on independence, stable definitions, correct grouping, and reliable measurement.
A short sample can produce almost any plausible gambling result. A large clean sample can estimate performance more precisely, but it still cannot make a random outcome “due” or guarantee that the next result follows the average.
Stopping after a favorable result biases the evidence
An analyst can create a misleading conclusion by repeatedly checking the data and stopping when the percentage looks impressive. A player does something similar by choosing the exact beginning and end of a streak after seeing the outcomes.
The observation window, primary metric, and stopping rule should be defined before analysis where practical. If the team explores many segments and periods, the final report should label exploratory findings and seek confirmation in new data.
Precision and detection are different planning goals
A sample-size plan may aim to estimate a rate within a chosen margin of error, or to detect a meaningful difference between groups. The second goal also depends on statistical power and the minimum effect worth acting on.
A huge sample can find a tiny difference that has no operational value. A smaller study may miss a material effect. Good planning states both statistical and business significance before collecting more data.