The geometric mean is the average appropriate for values that multiply together rather than add — growth rates, ratios, index scores, and concentrations measured on a multiplicative scale. It is calculated as the nth root of the product of n values, and it is mathematically identical to exponentiating the arithmetic mean of the logged values. Researchers reach for the ordinary (arithmetic) mean by default, but for a specific and common class of data — anything involving compounding, rates of change, or log-normally distributed measurements — the arithmetic mean gives a number that is not just less useful, it is factually wrong. This guide covers the formula, the properties that make it necessary, the geometric standard deviation used to report log-normal data, a frequent back-transformation error, and how the related harmonic mean differs. For the standard measures of center and spread more broadly, see the descriptive statistics guide, which covers the arithmetic mean in full.
What is the geometric mean?
For a set of n positive values x₁, x₂, …, xₙ, the geometric mean (GM) is defined two equivalent ways:
GM = (x₁ × x₂ × … × xₙ)1/n
the nth root of the product of all the values — and, equivalently:
GM = exp( (ln x₁ + ln x₂ + … + ln xₙ) / n )
the exponentiated arithmetic mean of the natural logs. These give identical results because taking a log converts multiplication into addition: the product of the raw values becomes a sum once logged, an arithmetic mean of that sum is taken, and exponentiating reverses the log to return to the original scale. This log-based form is also how every statistical package actually computes it, since multiplying many raw values directly risks numeric overflow or underflow.
Illustrative example: the geometric mean of 4 and 9 is √(4 × 9) = √36 = 6. Using the log form: (ln 4 + ln 9) / 2 = (1.386 + 2.197) / 2 = 1.792, and exp(1.792) = 6. Both routes agree, as they always will.
Why the arithmetic mean is the wrong tool for growth rates
This is the single most important reason to reach for the geometric mean, and it is worth working through a concrete illustrative example rather than stating it abstractly.
Suppose a research output — a journal’s citation count, a lab’s grant income, a population count, it does not matter which — grows 100% in year one, then falls 50% in year two. What is the average annual growth rate?
The arithmetic mean of the two percentage changes is (100% + (−50%)) / 2 = 25%, suggesting healthy average growth. But trace the actual value: starting at 100, a 100% increase brings it to 200; a 50% decrease from 200 brings it back to 100. After two years of “25% average growth,” the value is exactly where it started — zero net growth. The arithmetic mean of the growth rates is not just imprecise here, it is wrong in direction: it implies growth where none occurred.
The geometric mean gets this right. Growth rates are converted to growth factors (a 100% increase is a factor of 2.00; a 50% decrease is a factor of 0.50), and the geometric mean of the factors is calculated: √(2.00 × 0.50) = √1.00 = 1.00, a growth factor of exactly 1 — that is, 0% average annual growth, matching reality. This is precisely why compound annual growth rate (CAGR), average population growth rate, and average citation-growth rate are all geometric means of period-over-period ratios, never arithmetic means of period-over-period percentage changes. Whenever a quantity compounds — each period’s value depends multiplicatively on the last — the geometric mean is the only average that reconstructs the actual end-to-end change.
When the geometric mean is the correct choice
Ratios, rates of change, and compounding
Any sequence of multiplicative changes — investment returns, citation growth year over year, population growth, inflation-adjusted price indices — should be averaged geometrically, for the reason worked through above. A related and equally common case is averaging ratios or relative values (fold-changes in gene expression, relative risk across several studies, index numbers relative to a baseline year): because a ratio of 2 (doubling) and a ratio of 0.5 (halving) are equal and opposite on a multiplicative scale but not on an additive one, only the geometric mean treats them symmetrically. The arithmetic mean of 2 and 0.5 is 1.25, which is not neutral; the geometric mean is 1.0, which correctly represents “no net change on average.”
Log-normally distributed data
Many biological and clinical measurements are not normally distributed on their raw scale but are approximately normal once log-transformed — that is, they are log-normally distributed. Common examples include antibody titres, viral load, enzyme activity levels, drug concentrations, and reaction times. These variables are typically bounded at zero, right-skewed (a long tail of high values), and span several orders of magnitude across a sample. For this shape of data, the arithmetic mean is pulled upward by the high-value tail and is a poor summary of the “typical” value; the geometric mean, calculated on the log scale and back-transformed, is the more representative central-tendency measure and is the conventional one reported in the biomedical literature for such variables. See the guide to skewness for more on right-skewed distributions, and normality of distribution for how to test whether log-transforming a variable actually achieves approximate normality before relying on this approach.
Index construction
Composite indices that combine several normalized indicators measured in different units — a research-output index blending publication counts, citation rates, and funding totals, for example — commonly use a geometric mean rather than an arithmetic mean to aggregate the sub-indicators. The reason is the same multiplicative logic: an arithmetic mean lets a very high score on one indicator compensate freely for a very low score on another (full substitutability), while a geometric mean penalizes imbalance — a score of zero (or one) on any single component pulls the whole index down sharply, which is usually the more defensible property when “balanced performance across dimensions” is what the index is meant to capture.
Properties of the geometric mean
- It is always less than or equal to the arithmetic mean for any set of positive values, with equality only when all the values are identical. This is the AM-GM inequality, a basic and well-established result: for any positive x₁,…,xₙ, (x₁+…+xₙ)/n ≥ (x₁×…×xₙ)1/n. The gap between the two widens as the values become more spread out (higher variance) — for identical values the two means coincide exactly.
- It is undefined for any dataset containing a zero or a negative value. The product becomes zero or the root of a negative number becomes undefined for even n, or of inconsistent sign for odd n. Ratio data occasionally including a zero (no citations at all in a given year, say) has to be handled explicitly — commonly by adding a small constant before taking logs, or by excluding zeros with the exclusion reported — rather than silently breaking the calculation.
- It is less sensitive to large outliers than the arithmetic mean. Because the geometric mean operates on the log scale, a single very large value is compressed before it is averaged, so it pulls the result up far less than it would pull an arithmetic mean.
The geometric standard deviation, and why log-normal data are reported as mean ×÷ GSD
For a normally distributed variable, “mean ± SD” is a symmetric interval around the mean, appropriate because the distribution itself is symmetric. Log-normal data are not symmetric on their raw scale, so an additive ± SD interval built from the raw values is misleading — it can even extend below zero for a variable that is strictly positive by definition (a concentration, for instance).
The correct analogue is the geometric standard deviation (GSD), calculated as GSD = exp(SD of the log-transformed values). Because it comes from a multiplicative scale, the GSD is combined with the geometric mean multiplicatively rather than additively: the reported interval is GM × GSD to GM ÷ GSD, not GM ± GSD. This interval is asymmetric around the geometric mean on the raw scale (wider above than below), which correctly reflects the underlying right skew, and it stays strictly positive — unlike an additive interval built on raw log-normal data, which does not.
Back-transformation: a common and consequential analysis error
A frequent mistake when working with log-transformed data: taking the arithmetic mean of the logged values and then exponentiating it does not give the arithmetic mean of the original, untransformed values — it gives the geometric mean of the original values. exp(mean(log(x))) = geometric mean of x, always; it is only equal to the arithmetic mean of x when every value in x is identical. Reporting a back-transformed value as though it were an arithmetic mean (a common slip when a manuscript log-transforms a skewed outcome for a valid statistical reason — to satisfy a model’s normality assumption — and then back-transforms the model’s estimated mean for the results section) silently changes which quantity is being reported, and understates the true arithmetic average of a right-skewed variable. The result itself is not wrong — it is a legitimate geometric mean — but it must be labelled as one, not presented as if it were the ordinary average.
The harmonic mean: a related but different tool
The harmonic mean (HM) is a third type of average, calculated as the number of values divided by the sum of their reciprocals: HM = n / (1/x₁ + 1/x₂ + … + 1/xₙ). It is the correct average specifically for rates expressed as a ratio of two quantities where the numerator (not the denominator) is held fixed — the textbook case is averaging speeds over a fixed distance. A trip covered at 60 mph for the first half of a fixed distance and 30 mph for the second half averages 40 mph overall (the harmonic mean of 60 and 30), not 45 mph (their arithmetic mean), because more time is spent at the slower speed over that fixed distance.
The harmonic mean also underlies the F1 score used to evaluate classification and information-retrieval models: F1 is the harmonic mean of precision and recall, F1 = 2 × (precision × recall) / (precision + recall). The harmonic mean is used there specifically because it penalizes a large imbalance between precision and recall much more heavily than an arithmetic mean would — a model with 100% precision and 1% recall scores an arithmetic mean of ~50%, which overstates its usefulness, but a harmonic mean of just under 2%, which reflects it accurately. Note the ordering that always holds for positive values: harmonic mean ≤ geometric mean ≤ arithmetic mean, with equality across all three only when every value is identical.
Reporting conventions for log-transformed outcomes
- Report the geometric mean, not the arithmetic mean, as the central-tendency statistic for variables that were log-transformed for analysis (titres, concentrations, reaction times, and similarly right-skewed positive measures).
- Pair it with the geometric standard deviation or a back-transformed confidence interval (exponentiate the CI bounds calculated on the log scale), not an additive ± SD/SE built from raw values.
- State explicitly in methods that values were log-transformed and that reported averages are geometric means — this single sentence prevents the back-transformation ambiguity described above.
- When comparing groups on log-transformed outcomes, note that the difference in means on the log scale, once exponentiated, becomes a ratio of geometric means between groups, not a difference — a common point of confusion when writing up results.
For the broader toolkit these conventions sit inside — central tendency, dispersion, shape, and general APA-style reporting — see the descriptive statistics guide. For the unitless companion measure of relative spread, see coefficient of variation, and for a robust, outlier-resistant spread measure see mean absolute deviation. A histogram is usually the fastest way to spot the right-skewed shape that signals log-normal data in the first place.
Calculating the geometric mean: R, Python, and Excel
- R has no single built-in function; the standard approach is
exp(mean(log(x))), or install a package such aspsychorDescToolsfor a dedicatedgeometric.mean()/Gmean()function. - Python:
scipy.stats.gmean(x)(from SciPy) computes it directly;numpy.exp(numpy.mean(numpy.log(x)))is the equivalent manual form. - Excel / Google Sheets: the built-in
=GEOMEAN(range)function.
All three require every input value to be strictly positive — a zero or negative value in the range will produce an error or an undefined result, consistent with the property described above.
Frequently asked questions
Is the geometric mean always smaller than the arithmetic mean?
Yes, for any set of positive values that are not all identical, the geometric mean is strictly less than the arithmetic mean (the AM-GM inequality). The two are equal only when every value in the set is the same.
Can the geometric mean be negative or zero?
Not for real-valued data with an even count of values: the geometric mean is undefined if any value in the set is zero or negative, because the calculation requires taking a root of a product that would be zero, negative, or undefined. Analysts working with data that can include zero or negative values (net returns that include a loss year, for example) need a different approach — commonly adding a constant before log-transforming, with that adjustment reported — rather than applying the geometric mean formula directly.
What is a simple real-world example of when to use the geometric mean?
Average annual investment return over several years is the clearest case: if an investment returns +20% one year and −20% the next, the arithmetic mean of the two returns is 0%, but the actual outcome is a net loss (1.20 × 0.80 = 0.96, a 4% overall decline), which the geometric mean of the growth factors correctly captures as roughly −2% average annual return. The same logic applies to any other compounding or ratio-based quantity, including citation growth or population growth over time.
Why not just use the arithmetic mean for everything?
Because for ratio-scale, multiplicative, or log-normally distributed data, the arithmetic mean answers a different, less relevant question than the one usually being asked. It correctly summarizes an additive process (values that sum, such as a set of independent measurement errors) but not a multiplicative one (values that compound), and for right-skewed positive data it is inflated by the upper tail relative to where most of the data actually sits.
How is the geometric mean different from the harmonic mean?
Both are calculated via a transformation-then-back-transformation pattern (logs for the geometric mean, reciprocals for the harmonic mean), and both are less than or equal to the arithmetic mean. But they answer different questions: the geometric mean is the right average for multiplicative/compounding quantities and ratios; the harmonic mean is the right average specifically for rates where a numerator quantity is held fixed across the observations being averaged, such as speed over a fixed distance, or the balance between precision and recall in an F1 score.
