Standard Error Calculator - SEM, SD and CI Conversions

    The standard error of the mean from raw data, an SD and n, or a count, plus the reverse: SD from a reported SEM or from a 95% confidence interval, and the sample needed for a target precision.

    Parameters

    Enter data for calculations

    Only the boxes for your choice appear

    Form progress0 / 1 fields

    💡 Fill in all required fields to unlock the calculate button

    Standard error of the mean, from data, a report or a published interval

    Divide the standard deviation by the square root of the sample size and you have the standard error of the mean: 15 / √36 = 2.50. This standard error calculator does that from a pasted list, from a standard deviation and n, or from a count of successes for a proportion. It also works backward, which is what most people searching for SEM actually need: it turns a reported standard error or a published 95% confidence interval into the standard deviation of single values. An optional target tells you how many observations bring the standard error down to the precision you want, and an optional population size applies the finite population correction.

    Problem: bars on a chart that nobody labeled

    A bar chart shows a mean of 103 with error bars. If the bars are one standard deviation, they reach from 88 to 118 and describe how far single measurements spread. If they are one standard error from 36 measurements, they reach from 100.5 to 105.5 and describe how much the mean itself would wobble between samples. Same data, bars six times apart in length, because √36 = 6. Journals print both, often with nothing more than "±" in the legend, and readers routinely treat a small SEM as a small spread.

    The confusion runs the other way too. A meta-analysis or a sample size plan needs the standard deviation, but the paper reports "mean ± SEM" or only a 95% interval. Plugging the standard error in where the standard deviation belongs makes the data look far more consistent than it was, and a study planned on it ends up badly underpowered.

    Five entries, and the first one decides which boxes appear

    1. Pick what you have: raw measurements, a standard deviation with n, successes out of a sample, a standard error from a report, or a confidence interval from a paper.
    2. Type the numbers in plain digits with a dot for decimals. The sample mean is optional where it appears; with it, the result prints ranges and a relative standard error instead of bare half-widths.
    3. For an interval, give both ends, the group size and the printed confidence level, so the right t or z divisor is used.
    4. Optionally add the population size (for a sample that is a large share of a finite group) and a target standard error.
    5. Read the headline, then the table that sets mean ± SD, mean ± SE and the 95% interval side by side.

    Solution: keep three plus-or-minus numbers apart

    Every result prints the same sample three ways, so the difference is on the screen rather than in a footnote. For the chart above, with a mean of 103, a standard deviation of 15 and n = 36:

    Written as Half-width Range Describes
    mean ± SD15.0088.00 to 118.00about 68% of single measurements, if roughly normal
    mean ± SE2.50100.50 to 105.50the typical wobble of the mean between samples
    mean ± t × SE5.0897.92 to 108.0895% confidence interval for the mean, t = 2.030 on 35 df

    The standard deviation stays near 15 however many people you measure. The standard error falls with the square root of n: at 144 measurements it is 1.25, at 576 it is 0.63. That is why the result also prints a row for two, four, nine and sixteen times your sample, with the standard deviation held fixed next to it.

    Moving between SD, SE and a confidence interval

    You have You want Formula
    SD and nSE of the meanSE = SD / √n
    SE and nSDSD = SE × √n
    95% interval, n under 60SESE = (upper - lower) / (2 × t), df = n - 1
    95% interval, larger nSESE = (upper - lower) / 3.92
    x successes of nSE of the proportionSE = √(p(1 - p) / n), p = x / n
    Any of the above, population Ncorrected SESE × √((N - n) / (N - 1))

    The split at 60 observations comes from the Cochrane Handbook for Systematic Reviews, which recommends the t divisor for small groups. At n = 25 the two differ by about 5%; at n = 200 by well under 1%.

    Six inputs and the standard error each one gives

    Five measurements 12, 14, 15, 13, 16: SD 1.581, SE 0.707, relative standard error 5.1%.
    SD 15, n = 36: SE 2.50. To reach a standard error of 1.50 at the same spread takes 100 observations, 64 more.
    520 of 1,000 said yes: SE 1.58 percentage points. Getting it down to 1 point needs 2,496 responses.
    A paper reports 120 ± 2.5 (SEM), n = 25: the standard deviation is 2.5 × 5 = 12.50, and single values spread roughly from 107.5 to 132.5.
    A 95% interval of 4.2 to 7.8, n = 25: with t = 2.064 the SE is 0.872 and the SD 4.361. The 3.92 shortcut would give 0.918, about 5.3% too high.
    SD 620, n = 400 out of 1,000 staff: without the correction the SE is 31.0; with the factor 0.7750 it is 24.0.

    Standard error questions people type in

    Should error bars show the standard deviation or the standard error?
    It depends on the message. Use the standard deviation to show how variable the individuals are, and a 95% confidence interval to show how precisely the mean is known. Bare standard error bars are the hardest to read: two groups whose ±1 SE bars just touch differ by z = 2 / √2 = 1.41, a p-value of about 0.16, which is not significant. Whatever you draw, name it in the legend.
    How do I calculate the standard error in Excel or R?
    Excel has no SEM function. Use =STDEV.S(A2:A37)/SQRT(COUNT(A2:A37)). In R, sd(x) / sqrt(length(x)). Both use the n - 1 standard deviation, the same one this calculator uses.
    When is a standard error too large to report the estimate?
    A common yardstick is the relative standard error, SE divided by the estimate. The National Center for Health Statistics and other federal survey programs flag estimates with a relative standard error of 30% or more as unreliable. Three yes answers out of 40 give a rate of 7.5% with a standard error of 4.16 points, a relative standard error of 55.5%.
    Why does the standard error of zero out of 40 come out as zero?
    Because √(p(1 - p) / n) is zero at p = 0. That is a quirk of the formula, not proof that the true rate is zero. The linked confidence interval calculator gives a Wilson or exact interval that still has width in that case.
    Can I get the standard deviation from an interval that is not centered on the mean?
    Not from its width. Intervals for ratios, odds and skewed measurements are usually built on a log scale and are lopsided on the original one. Type the mean in the optional box: if it sits more than 5% of the half-width away from the midpoint, the result warns you instead of printing a misleading standard deviation.

    Related tools

    Grouped Data Quartile Calculator

    Quartiles and the quartile deviation when only class counts are available - See calculator

    Confidence Interval Calculator

    Turns the standard error into an interval at 80% to 99%, with Wilson and exact intervals for proportions - See calculator

    Standard Deviation Calculator

    Sample and population standard deviation and variance from a list - See calculator

    Sample Size Calculator

    Plans the number of responses for a margin of error, a mean or an A/B test - See calculator

    Coefficient of Variation Calculator

    The spread of single values as a share of the mean, the SD counterpart of the relative standard error - See calculator

    Two-Sample T-Test Calculator

    The proper test when two groups' error bars leave you unsure - See calculator

    Similar calculators from this section