RSE Calculator — Relative Standard Error

An estimate without its precision is half a statistic. The relative standard error (RSE) expresses a survey or sample estimate’s sampling error as a percentage of the estimate itself — the single number statistical agencies print (or suppress) beside every figure they release. This calculator computes it from the three ingredients every study reports: the mean, the standard deviation, and the sample size. Enter them and it returns the standard error, the RSE, and the verdict band your discipline would apply to it.

The Formula, and Why It Is a Percentage

The standard error of a mean is SE = SD/√n — the typical distance between your sample's mean and the truth, shrinking with the square root of effort. Divide that error by the estimate itself and multiply by 100 and you have the RSE: RSE = (SE ÷ estimate) × 100%. The percentage form is the point. An SE of 1.0 is excellent for a mean income of 50,000 and catastrophic for a mean count of 3; absolute errors cannot be judged without scale, and the RSE supplies the scale in one division. That is why official statistics — labour force surveys, health interview surveys, agricultural censuses — publish RSEs (or their cousin, the coefficient of variation of the estimate) as the quality label beside every cell, and why the same cell can be a headline in one survey and a suppressed blank in another: the estimate did not change, its RSE did.

Worked Example: A Mean of 50 With RSE 2%

The defaults: mean 50, standard deviation 10, n = 100. SE = 10/√100 = 1.0; RSE = 1.0/50 × 100 = 2.0% — comfortably in the high-precision band. The interval translation: roughly 95% confidence spans ±1.96 SE, i.e. the estimate is good to about ±3.9% of itself, 50 ± 2 in original units. Now move the levers and watch the economics: the same data at n = 25 gives SE 2.0, RSE 4% — still fine. At n = 10, SE 3.16, RSE 6.3% — serviceable. The mean's own size matters just as much: measure a rarer quantity with mean 5 and the same SD and n, and the RSE jumps to 20% — nothing about the fieldwork worsened; the base shrank. Subgroup estimates live this trap daily: a national survey precise to 1% overall can be at 15–25% for a small region or minority group, and the RSE is the number that should travel with any such figure into print.

The Bands: How Agencies Judge an RSE

Thresholds vary by agency and era, but the working consensus this calculator applies: under 5%, high precision, publish freely; 5–10%, good for most purposes; 10–20%, usable with an explicit caution flag; 20–30%, low precision, conclusions in pencil; above roughly 30%, many agencies (the Australian Bureau of Statistics' 50% suppression line and the US Census Bureau's flags live in this territory) mark the estimate unreliable or suppress it outright. Two qualifications keep the bands honest. First, they judge sampling error only — the RSE says nothing about biased questions, non-response, or misreporting, which in real surveys often exceed the sampling error entirely. Second, the bands are conventions, not physics: a 22% RSE on a unique, decision-relevant estimate may be worth publishing with its warning label, while a 3% RSE on a biased measure is precise nonsense. The number disciplines judgment; it does not replace it.

Buying Precision: The Square Root's Price List

Because SE falls with √n, precision is bought at a brutal exchange rate: halve the RSE and you must quadruple the sample; cut it to a third and the sample grows ninefold. This single fact designs every survey budget ever written. It explains why national samples cluster near 1,000–2,000 respondents (the knee where RSE gains flatten against cost), why small-area estimation resorts to modelling and borrowing strength across regions instead of brute sample, and why oversampling small subgroups — accepting a deliberately unrepresentative design, corrected by weights — is standard practice when subgroup RSEs matter. Run the calculator at your study's planned n and at four times it: the difference in RSE is the entire case for or against the larger budget, stated in one percentage. And note the other lever, quieter and sometimes cheaper: reducing SD itself — better questions, better measurement, a more homogeneous sampling frame — enters the formula at full strength, no square root applied.

The RSE of a Proportion: Small Percentages, Big Percentages of Error

The FAQ's formula deserves its worked moment, because it ambushes real headlines. A survey of n = 1,000 finds 4% prevalence of something — a condition, a behaviour, an intention. SE = √(0.04 × 0.96 / 1,000) = 0.0062, and RSE = 0.0062 / 0.04 = 15.5% — the caution band, from a thousand interviews. The 95% interval is 4% ± 1.2 points, a relative span of ±30%: "prevalence between 2.8% and 5.2%" is the honest headline, and "4%" alone is the dishonest one. Halve the prevalence to 2% at the same n and the RSE climbs to about 22%; this is why rare-population statistics are pooled across years and why a small subgroup's dramatic percentage swing between surveys is so often just the RSE clearing its throat. Before quoting any small percentage from any sample, run its three numbers through this page — the calculation takes twenty seconds and has ended more bad headlines than any editor.

Design Effects: When n Is Not Really n

The formula assumes a simple random sample, and real surveys rarely enjoy one. Cluster sampling — interviewing households street by street, students classroom by classroom — correlates the answers within each cluster, and the effective sample is smaller than the counted one by the design effect (deff): the true SE is the calculator's SE × √deff. A deff of 1.5, unremarkable in clustered fieldwork, inflates every SE and RSE by about 22%; the study that reports precision from the raw formula alone has quietly upgraded its own grades. Stratification can push the other way (deff below 1), and weighting for non-response usually costs further effective sample. Treat this page's output as the simple-random benchmark — the precision the same n would buy under textbook conditions — and ask any real survey for its design effect before believing its decimals.

RSE Versus Margin of Error: Translating Between Dialects

Polls speak of margin of error; statistical agencies speak of RSE; they are the same uncertainty in different units, and the conversion is one multiplication. The 95% margin of error is 1.96 × SE, so the relative margin of error is 1.96 × RSE — the defaults' 2.00% RSE is a ±3.9% relative margin, which on a mean of 50 is the ±2.0 points a pollster would quote. Conversely, a poll's advertised ±3 points on a 50% figure is an RSE of about 3.06% (3 ÷ 1.96 ÷ 50). Keep the dialects straight and a shelf of apparently conflicting precision claims resolves into one comparable scale — and the figure whose RSE cannot be recovered from its report at all is the figure to trust least, whatever dialect it chose.

Related Tools

More from the Statistics Calculators hub: the Error Propagation Calculator combines uncertainties through arithmetic, the Descriptive Statistics Calculator produces the mean and SD this page consumes, and the Coefficient of Determination Calculator covers precision's regression-side relative.

Frequently Asked Questions

What is the difference between SE, RSE and coefficient of variation?

The standard error is the absolute sampling error of an estimate. The RSE is that error as a percentage of the estimate. The coefficient of variation (CV) is the SD as a percentage of the mean — a property of the data's spread, whereas RSE is a property of the estimate's precision; they differ by the √n factor, and confusing them overstates precision by exactly that root.

What RSE is acceptable for publication?

Most statistical agencies treat under 5% as solid, flag 10–25% with cautions of increasing sternness, and suppress or heavily caveat estimates above roughly 30–50%. Journals rarely fix a line, but reviewers increasingly ask for the RSE (or confidence interval) beside any survey estimate — computing it takes the three numbers this page asks for.

Can RSE exceed 100%?

Yes, for estimates near zero or wildly variable data — an estimated rare-event count can carry an RSE well past 100%, meaning the sampling error dwarfs the estimate itself. Such numbers should be reported as "too imprecise to use," or pooled across periods or groups until the base supports them.

How do I reduce a bad RSE?

Four routes, in usual order of practicality: increase n (quadruple it to halve the RSE); reduce the data's SD through better measurement or a tighter population definition; pool waves or neighbouring groups; or switch to model-based small-area methods that borrow strength statistically. There is no fifth route — precision is bought with data, design, or assumptions, and the RSE prices each honestly.

Does a low RSE mean the estimate is accurate?

It means the estimate is precise — repeatable under resampling. Accuracy also needs the survey to be unbiased: a question everyone misunderstands the same way yields a beautifully low RSE around the wrong number. Total survey error is sampling error plus everything else; the RSE certifies only its own term.

Why do proportions use a different-looking formula?

For a proportion p, the SD is √(p(1−p)), so SE = √(p(1−p)/n) and RSE = SE/p — the same definition, with the binary data's own SD substituted. Rare proportions (small p) therefore carry large RSEs at ordinary sample sizes, which is why survey reports flag small percentages even from big surveys.