Written and maintained by CASRAI Editorial Board
Published
On this page
The method of moments (also called the moment method, moments method or method of moments estimation) is a way to estimate a distribution’s unknown parameters by setting its theoretical moments equal to the corresponding sample moments and solving for the parameters. If a distribution has one unknown parameter, you match the mean; with two, you match the mean and the variance, and so on. The result is a closed-form estimate you can usually compute by hand.
Method of moments definition
A UC Berkeley statistics course (Stat 135, Spring 2026) defines the method as estimating parameters “by finding expressions for them in terms of the moments of the lowest possible order.” The NIST/SEMATECH e-Handbook of Statistical Methods describes the same idea more compactly: the method “equates sample moments to parameter estimates.” Both sources treat it as one of several general-purpose estimation approaches alongside maximum likelihood and least squares.
Two practical points frame everything below. First, NIST notes that “when moment methods are available, they have the advantage of simplicity,” but that they often lack the optimality properties of maximum likelihood and least squares estimators. Second, NIST says moment estimates mainly serve as starting values for those more precise estimators. That is the honest positioning: a fast, transparent first estimate, not necessarily the final one.
Moment method vs moments method: same thing
“Method of moment”, “method of moments”, “moment method”, “moment methods” and “moments method” all refer to this single technique. It should not be confused with the generalized method of moments (GMM), an econometric estimator built on moment conditions and a weighting matrix, which is a different procedure with its own theory.
Formula and symbols
Following the Berkeley notes, for a random variable X the k-th moment is
μk = E(Xk)
and, for a sample X1, …, Xn, the k-th sample moment is
μ̂k = (1/n) Σ Xik
- X: the random variable whose distribution depends on unknown parameter(s) θ.
- E(·): expected value under the model.
- k: the moment order (1 for the mean, 2 for the mean of squares).
- n: the sample size; Xi: the i-th observation.
- μk: the theoretical moment, a function of θ; μ̂k: its sample counterpart, a number computed from data.
The variance links the first two moments: Var(X) = μ2 − μ12. The procedure has three steps, as laid out in the Berkeley lecture:
- Find E(Xk) as a function of the parameters, for as many orders k as there are unknown parameters.
- Solve for the parameters: θ = h(μ1, …, μk).
- Substitute the sample moments: θ̂ = h(μ̂1, …, μ̂k).
Standard one- and two-parameter cases
The data table below applies the three steps to four common distributions. Bernoulli, exponential and Poisson cases appear in the Berkeley lecture; the gamma row follows from the same steps using the gamma mean (αθ) and variance (αθ²), which our gamma distribution formula guide derives.
Worked example: gamma shape and scale from six measurements
Suppose six positive, right-skewed measurements (an illustrative dataset, not drawn from a study): 2.1, 3.4, 1.8, 4.5, 2.9, 3.7. We model them as gamma(α, θ) with shape α and scale θ, so the mean is αθ and the variance is αθ². Two unknowns need two moments.
Step 1, theoretical moments. μ1 = αθ and μ2 − μ12 = αθ².
Step 2, solve. Dividing the variance by the mean gives θ = Var/mean. Squaring the mean and dividing by the variance gives α = mean²/Var.
Step 3, sample moments.
- Sum: 2.1 + 3.4 + 1.8 + 4.5 + 2.9 + 3.7 = 18.4, so x̄ = 18.4 / 6 = 3.0667.
- Sum of squares: 4.41 + 11.56 + 3.24 + 20.25 + 8.41 + 13.69 = 61.56, so μ̂2 = 61.56 / 6 = 10.26.
- Moment-matching variance: s² = μ̂2 − x̄² = 10.26 − 9.4044 = 0.8556. Check by deviations: Σ(x − x̄)² = 5.1333, and 5.1333 / 6 = 0.8556.
Result. θ̂ = 0.8556 / 3.0667 = 0.2790 and α̂ = 3.0667² / 0.8556 = 9.4044 / 0.8556 = 10.99. As a consistency check, α̂ × θ̂ = 10.99 × 0.2790 = 3.067, the sample mean, and α̂ × θ̂² = 0.856, the sample variance (both by construction).
A one-parameter example from the Berkeley notes
For an exponential model with rate λ, μ1 = 1/λ, so λ̂ = 1/x̄. In the Berkeley alpha-particle example, the observed mean count was 0.8392 per second, and with a Poisson model the rate per 10-second interval is λ̂ = 0.8392 × 10 = 8.392.
Interpretation: how to read a moment estimate
The estimates above are point estimates only. They say which parameter values reproduce the sample mean and variance exactly; they do not say how uncertain those values are. With n = 6, treat 10.99 as a rough number.
| What you see | How to read it |
|---|---|
| Estimates reproduce sample mean and variance exactly | This is by construction, not evidence the model fits. Check fit separately (histogram or probability plot). |
| Estimate used as a final answer | NIST notes that moment methods lack the optimality properties of maximum likelihood and least squares; report uncertainty or refine. |
| Estimate used as a start | NIST describes this as the main use: starting values for maximum likelihood or least squares routines. |
| Small n | Sampling variability is large; the estimate of a higher-moment-dependent parameter (such as α) is especially unstable. |
For the connection between estimators and bias, see what an unbiased estimator is. For the maximum-likelihood alternative, see the likelihood function guide.
Assumptions and common mistakes
- The moments must exist and be expressible in the parameters. You need a model whose first few moments are finite functions of the parameters; the method works from those moments, not from the full density.
- Use as many moments as parameters. One unknown, one equation; two unknowns, two equations.
- Divide by n for the moment-matching variance. The equation is μ̂2 − x̄², which equals Σ(x − x̄)²/n. Using the usual sample variance with n − 1 (see how to find sample variance) gives 1.0267 instead of 0.8556 in the example and changes both estimates. Either can be defended, but state which you used.
- Match the parameterisation. Gamma is written with a scale θ in some sources and a rate in others (rate = 1/scale). Mixing them inverts the result.
- Do not confuse it with GMM. The generalized method of moments is a separate econometric estimator.
- Do not read the fit as validation. Matching moments always “works” arithmetically.
How to report a method of moments estimate (APA)
Name the method, the model, the sample size, the estimates and the variance convention, rounded consistently. Example:
Gamma shape and scale parameters were estimated by the method of moments (n = 6, variance computed with divisor n): α̂ = 10.99, θ̂ = 0.28.
If the estimates feed a later maximum likelihood fit, say so: “Method-of-moments estimates were used as starting values.”
How to calculate it in Excel, R or Python
NIST observes that most general-purpose statistical software has no explicit method-of-moments command, but the formulas are simple enough to implement in most programs. Each snippet below computes the same gamma example.
Excel
With the six values in A1:A6: mean =AVERAGE(A1:A6) in B1, variance with divisor n =VAR.P(A1:A6) in B2, then scale =B2/B1 (0.2790) and shape =B1^2/B2 (10.99).
R
x <- c(2.1, 3.4, 1.8, 4.5, 2.9, 3.7)
m <- mean(x)
v <- mean((x - m)^2) # divisor n
c(shape = m^2 / v, scale = v / m)
Python
import statistics as st
x = [2.1, 3.4, 1.8, 4.5, 2.9, 3.7]
m = st.fmean(x)
v = st.pvariance(x) # divisor n
print(m**2 / v, v / m) # shape, scale
Both give shape 10.99 and scale 0.2790.
Where it appears in research practice
A well-known applied use is meta-analysis: the DerSimonian-Laird estimator of between-study variance (τ²) is a moment-based method. According to Chapter 10 of the Cochrane Handbook (section 10.10.4.4), RevMan used only that method until 2024, and since 2024 offers restricted maximum likelihood (REML) too, with REML as the default and DerSimonian-Laird still selectable. See inverse-variance weighting in meta-analysis for how the choice plays out.
FAQ
What is the method of moments in statistics?
It is a parameter-estimation technique that sets theoretical moments (such as the mean and variance) equal to sample moments and solves for the unknown parameters.
Method of moments vs maximum likelihood: which is better?
NIST states that moment methods are simple when available but lack the optimality properties of maximum likelihood and least squares estimators, and are mainly useful as starting values for those methods. Maximum likelihood needs the full density; the moment method needs only the first few moments as functions of the parameters.
How many moments do I need?
As many as there are unknown parameters: the mean for one parameter, the mean and variance for two.
Why divide by n and not n − 1?
The method matches the second sample moment, giving the variance as (1/n) Σ(x − x̄)². Using n − 1 is a different convention; report which you used.
Is the generalized method of moments the same thing?
No. GMM is an econometric extension built on moment conditions and a weighting matrix.
Sources
- NIST/SEMATECH e-Handbook of Statistical Methods, “Method of Moments”: itl.nist.gov/div898/handbook/eda/section3/eda3651.htm (read 2026-11-12).
- UC Berkeley Stat 135, Spring 2026, Lecture 5, “Parameter Estimation: Method of Moments”: stat135.berkeley.edu/spring-2026/lectures/lecture-5.html (read 2026-11-12).
- Cochrane Handbook for Systematic Reviews of Interventions, Chapter 10, section 10.10.4.4: cochrane.org handbook chapter 10 (checked in CASRAI’s fact cache).
Data: Method-of-moments estimators for common distributions
Moment equations and resulting method-of-moments estimators for four common distributions, compiled by applying the three-step procedure from the UC Berkeley Stat 135 lecture on parameter estimation; the gamma row uses the standard gamma mean and variance.
| Distribution | Unknown parameters | Moment equation(s) | Method-of-moments estimator |
|---|---|---|---|
| Bernoulli(p) | p | μ1 = p | p̂ = x̄ |
| Exponential(λ) | λ | μ1 = 1/λ | λ̂ = 1/x̄ |
| Poisson(λ) | λ | μ1 = λ | λ̂ = x̄ (scaled by interval length, e.g. 0.8392 × 10 = 8.392) |
| Gamma(α, θ), shape-scale | α, θ | μ1 = αθ; μ2 − μ1² = αθ² | θ̂ = s²/x̄; α̂ = x̄²/s² (s² with divisor n) |
Sources
- UC Berkeley Stat 135, Lecture 5: Parameter Estimation: Method of Moments (accessed 2026-11-12)
- NIST/SEMATECH e-Handbook of Statistical Methods: Method of Moments (accessed 2026-11-12)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







