Can different quantities have different ESS values from the same data?

Yes, because precision depends on the quantity being estimated.

Does duplicating records create information?

No; a correct analysis must account for that dependence.

Is ESS the same as effective degrees of freedom?

No. ESS concerns information relative to a reference sample; degrees of freedom concern a different aspect of model or inferential complexity.

AI Glossary / Statistics and probability / Intermediate

Effective sample size

Definition

Effective sample size is the size of a specified reference sample that would provide comparable statistical information or precision for a particular analysis. It distinguishes the number of recorded observations from the information they contribute. The definition depends on the estimator, outcome, sampling design, dependence, and comparison being used. There is no single formula that applies to every dataset.

What it means

A file containing many rows may contain less independent information than its size suggests. Repeated measurements from the same customer, strongly unequal analysis weights, or highly correlated simulation draws can reduce precision relative to a simple independent reference sample.
The supplied glossary also uses effective sample size to compare information across outcome types, including losses associated with ties or censoring. These uses share an information-equivalence idea but are not interchangeable numerical calculations. A reported ESS should therefore name the quantity being estimated and the reference standard.

How it works

First choose the target, such as a population mean, a treatment comparison, or a posterior expectation. Then identify the uncertainty or power under the actual analysis and under an appropriate reference analysis. Effective sample size summarizes their relationship rather than replacing the underlying calculation.
For fixed nonnegative weights applied to independent observations with common variance, a weighted mean has variance proportional to sum(w_i squared)/(sum w_i) squared. Matching this to an equally weighted mean gives the familiar weight-based expression ESS = (sum w_i) squared/sum(w_i squared). Equal positive weights give the actual sample count; concentrated weights reduce this diagnostic.
In complex surveys, the ratio of the actual variance to a reference simple-random-sampling variance is a design effect. The corresponding variance-equivalent sample size is often expressed as n divided by that effect. A weights-only calculation does not account for all clustering, stratification, calibration, or relationships between weights and outcomes.
For a stationary simulation sequence, a mean-specific asymptotic ESS can be written N/[1 + 2 sum of lag autocorrelations], when the relevant limit and sum are well behaved. Positive serial dependence commonly reduces it, while useful negative dependence can produce a value above N. Estimated bulk and tail diagnostics address different aspects of simulation performance.
For censored survival outcomes, the number of observed events is an important information benchmark in common proportional-hazards comparisons. It is not a universal replacement for sample size in every survival estimand or design. Outcome ties can likewise reduce information for some rank-based comparisons.

Three practical examples

A survey team reports the actual respondent count and a weight-based ESS, alongside its full survey variance estimates.
A simulation analyst checks ESS for important posterior quantities to assess Monte Carlo precision.
A study planner distinguishes enrolled participants from observed outcome events when assessing information for a survival comparison.

Think of it this way

Think of listening to several people describe an event. Ten independent accounts can tell you more than ten repetitions of one account. Effective sample size asks how much information the collection provides for a particular question, rather than simply counting how many statements were recorded.

Worked example

Consider four independent measurements with the same mean and common variance nine. Assign fixed weights 1, 1, 2, and 4. Their sum is eight, and their squared values sum to 1 + 1 + 4 + 16 = 22.
The weight-based ESS is 8 squared/22 = 64/22, approximately 2.91. The actual number of measurements remains four.
The weighted mean's variance is 9(22/64) = 3.09375. Four equally weighted measurements would have variance 9/4 = 2.25. Their variance ratio is 3.09375/2.25 = 1.375, and the standard-error ratio is the square root of 1.375, approximately 1.173.
Thus the weighted mean has about 17.3% greater standard error under these assumptions. Multiplying every weight by ten leaves the ESS unchanged because the same squared factor cancels. These calculations concern fixed weights and equal independent variances; they do not describe every weighted business survey.

Illustrative case study

Fictional case: Northline Insights receives a large customer survey but applies highly variable weights to represent under-sampled groups. Its dashboard initially emphasizes only the response count.
The analysis team adds a weight-based ESS and estimates uncertainty using the actual survey design. Some subgroup estimates have much less precision than managers expected. The team improves recruitment for poorly represented groups and explains that removing weights merely to increase the diagnostic could distort the target population. The resulting report separates representation, precision, and raw volume.

Common misconceptions and mistakes

Mistake: reporting ESS as a corrected count of real people. It is an analysis-dependent equivalent, not a revised headcount.
Mistake: assuming ESS must always be smaller than N. Anticorrelated simulation draws or an efficient design can improve precision relative to the chosen reference.
Mistake: using the weights-only formula to capture arbitrary dependence. It does not.
Mistake: treating high simulation ESS as proof that all relevant posterior regions were explored. Convergence and other diagnostics still matter.
Mistake: combining different ESS values without explaining their definitions.

Business use

Businesses can use ESS to communicate why large datasets sometimes yield uncertain estimates and to prioritize better data collection or more efficient computation. Reports should retain the actual count, the estimator, the ESS method, and the resulting uncertainty. Increasing independent coverage or improving a sampling design may be more valuable than adding many similar observations. ESS alone does not measure bias, representativeness, or decision value.

Check your understanding

Why does rescaling all four example weights leave ESS unchanged?

Answer: Both numerator and denominator acquire the same squared multiplier.

What must accompany an ESS number?

Answer: Its definition, target estimator, reference comparison, and relevant assumptions or limitations.

Related terms

Sample size
Design effect
Autocorrelation
Censoring
survey weights
Monte Carlo standard error
statistical power
and effective degrees of freedom

Learn next

derive the variance of a weighted mean, compare survey design effects with weights-only diagnostics, and distinguish simulation precision from uncertainty in the underlying posterior distribution.

Sources

Sources: Harrell, Glossary of Statistical Terms: https://hbiostat.org/glossary/ . svyweight package documentation, weight-based effective sample size: https://www.stat.ethz.ch/CRAN/web/packages/svyweight/refman/svyweight.html . Stan Reference Manual, Posterior Analysis: https://mc-stan.org/docs/reference-manual/analysis.html . The worked example is an original fixed-weight calculation.

Previous: Effective model dimensionality Next: Efficiently updatable neural network