Why Moment Equations Differ Across Models
The method of moments uses features of a sample, such as its average and spread, to estimate the parameters of a probability model. Its basic instruction is simple: set sample moments equal to the corresponding population moments and solve. The important work is choosing moments that actually identify the parameters, solving the resulting equations correctly, and checking that the estimates belong to the model’s parameter space.
The same sample statistic can have different meanings across distribution families. A sample mean directly estimates the parameter in a Bernoulli or Poisson model, but in an exponential model it estimates a mean that is the reciprocal of the rate. A two-parameter uniform model requires information about both location and spread. This comparison is aligned with the SOA VEE Mathematical Statistics syllabus.
The population moments used below are defined under the stated distribution and parameter restrictions. For a model with \(k\) parameters, we generally need \(k\) moment equations. Having the same number of equations and parameters is not by itself enough: the equations must determine the parameters uniquely, and the solution must be feasible. In particular, matching a moment does not guarantee that an estimate lies in the allowed parameter space.
One Moment Identifies One Parameter
For a one-parameter model, the first population moment often gives a direct equation. The exact equation depends on the family: the mean equals the Bernoulli probability or Poisson rate, but it equals the reciprocal of the exponential rate.
| Model and parameter space | Population mean | Method-of-moments estimate | Feasibility check |
|---|---|---|---|
| Bernoulli(\(p\)), \(0\le p\le1\) | \(E[X]=p\) | \(\hat p_{\mathrm{MM}}=\bar X\) | \(0\le\bar X\le1\) for a binary sample |
| Poisson(\(\lambda\)), \(\lambda\ge0\) | \(E[X]=\lambda\) | \(\hat\lambda_{\mathrm{MM}}=\bar X\) | Always feasible; if the model instead requires \(\lambda>0\), an all-zero sample gives a boundary value, not an admissible estimate |
| Exponential with rate \(\lambda>0\) | \(E[X]=1/\lambda\) | \(\hat\lambda_{\mathrm{MM}}=1/\bar X\) | Requires \(\bar X>0\); this holds for samples of strictly positive waiting times |
For the exponential model in the table, the assumed density is \(f_X(x)=\lambda e^{-\lambda x}\) for \(x\ge0\). Here \(\lambda\) is a rate, so its units are the inverse of the units of \(X\). If the parameter is instead a scale \(\theta=1/\lambda\), the mean is \(\theta\) and the moment estimate is \(\hat\theta_{\mathrm{MM}}=\bar X\). Rate and scale are different parameterizations of the same family, not interchangeable numerical estimates.
Worked Example: Bernoulli Probability From a Binary Sample
A hypothetical quality audit records whether each of six independently selected components passes a stress screen. Let \(X_i=1\) for a pass and \(X_i=0\) otherwise, and assume \(X_i\sim\operatorname{Bernoulli}(p)\), with \(0\le p\le1\). The recorded outcomes are \(1,0,1,1,0,1\). Estimate \(p\) by moments.
Method. For a Bernoulli variable, \(E[X]=1\cdot p+0\cdot(1-p)=p\). Match this population mean to \(m_1=\bar X\).
Calculate the sample moment. The sum of the outcomes is \(1+0+1+1+0+1=4\), so \(\bar X=4/6=2/3\). The moment equation is \(\hat p_{\mathrm{MM}}=\bar X\), giving \(\hat p_{\mathrm{MM}}=2/3\).
Check. The estimate is between zero and one, as required. It also equals the observed pass proportion, \(4/6\). The common mistake avoided here is treating the Bernoulli mean as a function that needs inversion beyond \(p=E[X]\).
Worked Example: Poisson Rate With an Observed Zero
Suppose the daily number of minor service calls at a hypothetical facility is modeled as independent observations from a Poisson distribution with rate \(\lambda\ge0\) calls per day. The counts over five days are \(0,1,2,0,3\). Find the moment estimate and assess what would happen if every count had been zero.
Method. The Poisson mean is \(E[X]=\lambda\). Match it to the first sample moment.
Calculate. The total is \(0+1+2+0+3=6\), and there are five observations. Thus \(\bar X=6/5=1.2\), so \(\hat\lambda_{\mathrm{MM}}=1.2\) calls per day.
Check. This estimate is in the specified parameter space \(\lambda\ge0\). If the counts were all zero, then \(\bar X=0\) and the same equation would give \(\hat\lambda_{\mathrm{MM}}=0\), which is feasible under this parameter space. If a problem insists on \(\lambda>0\), zero is not an admissible estimate; one must report that the moment solution is on the excluded boundary rather than claim it is a valid positive estimate. The common mistake is to state \(\lambda>0\) and then overlook the all-zero sample case.
Worked Example: Exponential Rate From Waiting Times
A hypothetical monitoring system records four strictly positive waiting times, in hours: \(0.5,1,1.5,3\). Assume they are independent and identically distributed with exponential rate \(\lambda>0\) per hour. Estimate the rate by moments and interpret its units.
Method. Under this rate parameterization, \(E[X]=1/\lambda\). The first-moment equation is \(\bar X=1/\hat\lambda_{\mathrm{MM}}\), so the solution is the reciprocal of the sample mean.
Calculate. The sum is \(0.5+1+1.5+3=6\) hours, and \(\bar X=6/4=1.5\) hours. Therefore, \(\hat\lambda_{\mathrm{MM}}=1/1.5=2/3\) per hour.
Check. The estimated mean waiting time implied by this rate is \(1/(2/3)=1.5\) hours, equal to the observed mean. The reciprocal converts hours to inverse hours. The common mistake avoided is reporting \(1.5\) as the rate: \(1.5\) is the estimated mean waiting time, while \(2/3\) per hour is the estimated rate.
Two Moments Identify Two Parameters
When the model has two parameters, a single sample mean generally cannot determine both. For the uniform and normal families considered here, match the first two raw moments, or equivalently match the mean and variance. Use \(v_n\), the variance with divisor \(n\), because it is exactly the centered sample moment implied by this matching procedure.
For the uniform model, adding and subtracting the two parameter estimates is a useful check: their midpoint is \(\bar X\), and their half-width is \(\sqrt{3v_n}\). If \(v_n=0\), the equations give \(\hat a=\hat b=\bar X\), which is not an admissible pair for a nondegenerate uniform distribution with \(a<b\). For the normal model, \(v_n=0\) likewise gives a boundary value that is not an admissible positive variance. In either case the moment equations have produced a boundary solution, not a feasible estimate under the strict model.
Use the mean and variance of the specified family, with the model’s parameterization made explicit.
Calculate \(\bar X\) and, when a variance is needed, \(v_n=\frac{1}{n}\sum(X_i-\bar X)^2\).
Substitute the sample moments, solve for each parameter, and verify strict or non-strict parameter restrictions.
Why the uniform endpoint equations hold. Write the matched mean and variance as \(M=(a+b)/2\) and \(V=(b-a)^2/12\). Since \(a<b\), \(b-a\) is positive, so solving the variance equation gives \(b-a=\sqrt{12V}=2\sqrt{3V}\). The mean equation gives \(a+b=2M\). Adding and subtracting these equations yields \(b=M+\sqrt{3V}\) and \(a=M-\sqrt{3V}\). Replacing \(M\) by \(\bar X\) and \(V\) by \(v_n\) gives the stated estimates when \(v_n>0\). The positive square root is required by \(b-a>0\).
Worked Example: Estimating Both Uniform Endpoints
A hypothetical measurement is modeled as a uniform distribution on unknown endpoints \(a<b\). Five independent recorded values are \(2,4,6,8,10\). Estimate both endpoints by matching the mean and variance.
Method. Use \(\hat a_{\mathrm{MM}}=\bar X-\sqrt{3v_n}\) and \(\hat b_{\mathrm{MM}}=\bar X+\sqrt{3v_n}\). First calculate the required sample moments.
Calculate the mean. The sum is \(2+4+6+8+10=30\), so \(\bar X=30/5=6\).
Calculate the variance with divisor \(n\). The deviations from six are \(-4,-2,0,2,4\); their squares are \(16,4,0,4,16\), with total \(40\). Thus \(v_n=40/5=8\).
Solve. The half-width is \(\sqrt{3v_n}=\sqrt{24}\approx4.8990\). Therefore \(\hat a_{\mathrm{MM}}=6-\sqrt{24}\approx1.1010\) and \(\hat b_{\mathrm{MM}}=6+\sqrt{24}\approx10.8990\). Values are rounded to four decimal places.
Check. The estimated endpoints have midpoint \(6\), and their width is \(2\sqrt{24}\). The fitted uniform variance is \((2\sqrt{24})^2/12=96/12=8\), matching \(v_n\). The observations lie between the estimates, and \(\hat a<\hat b\). The common mistake avoided is using the observed minimum and maximum as if they were the two moment estimates.
Worked Example: Estimating Normal Mean and Variance
Suppose four independent observations of a hypothetical process are \(1,2,4,5\), modeled by \(N(\mu,\sigma^2)\) with \(\sigma^2>0\). Estimate both parameters by moments.
Method. Match \(E[X]=\mu\) to \(\bar X\), and \(\operatorname{Var}(X)=\sigma^2\) to \(v_n\).
Calculate the mean. The sum is \(1+2+4+5=12\), so \(\bar X=12/4=3\). Hence \(\hat\mu_{\mathrm{MM}}=3\).
Calculate the variance. Deviations from three are \(-2,-1,1,2\), whose squares sum to \(4+1+1+4=10\). Thus \(v_n=10/4=2.5\), and \(\hat\sigma^2_{\mathrm{MM}}=2.5\).
Check. The fitted normal model has mean \(3\) and variance \(2.5>0\), so both estimates are feasible. The common mistake avoided is replacing the moment-matching variance with the unbiased sample variance \(10/(4-1)=10/3\). That statistic uses divisor \(n-1\), whereas the second central sample moment used in these equations has divisor \(n\).
What the Comparison Reveals
The moment equations expose how much information each model parameter requires. The first moment alone identifies the single parameter in the Bernoulli, Poisson, and exponential models discussed here. In the exponential case, however, it identifies the rate through a reciprocal. The two parameters in the uniform and normal models require two equations; the mean determines location, while the variance determines spread.
A method-of-moments solution is not automatically an admissible estimate. Check the parameter space every time: Bernoulli probabilities must lie in \([0,1]\); a Poisson rate can include zero if that is the chosen model; an exponential rate must be positive; a nondegenerate uniform model requires \(a<b\); and a nondegenerate normal model requires \(\sigma^2>0\). If strict restrictions exclude the solution, report that fact rather than silently changing the model.
- Check the parameterization. An exponential rate and an exponential scale are reciprocals; their moment estimates are not the same.
- Check the divisor. Moment matching uses the sample moments as defined here, including divisor \(n\) in \(v_n\). Do not substitute a different variance estimator without justification.
- Check the boundary. A formula can return a boundary value even when the specified model excludes it, as for an all-zero Poisson sample under \(\lambda>0\).
- Check the number of equations. One moment cannot generally determine two parameters, and two equations must still yield a unique feasible solution.
Check Your Understanding
For each problem, use the stated moment-matching convention and check the parameter restrictions.
- A hypothetical inspection records binary outcomes \(1,0,1,1,0\) under an independent Bernoulli model. Find the moment estimate of \(p\). (Answer: \(\hat p_{\mathrm{MM}}=3/5=0.6\).)
- Five days of counts from a hypothetical Poisson process are \(0,2,1,4,1\). Using the parameter space \(\lambda\ge0\), find the rate estimate. (Answer: \(\hat\lambda_{\mathrm{MM}}=8/5=1.6\) counts per day.)
- Four positive waiting times, measured in hours, are \(0.25,0.5,0.75,1.5\) under an exponential model parameterized by rate \(\lambda\). Find the rate estimate and its units. (Answer: \(\bar X=0.75\) hours and \(\hat\lambda_{\mathrm{MM}}=4/3\) per hour.)
- Four observations from a uniform model with unknown endpoints are \(3,5,7,9\). Find both moment estimates, rounded to four decimal places. (Answer: \(\bar X=6,\ v_n=5,\ \hat a_{\mathrm{MM}}=2.1270,\ \hat b_{\mathrm{MM}}=9.8730\).)
- Five observations from a normal model with unknown mean and positive variance are \(2,2,4,6,6\). Find the moment estimates of \(\mu\) and \(\sigma^2\), and state whether the variance estimate is feasible. (Answer: \(\hat\mu_{\mathrm{MM}}=4,\ \hat\sigma^2_{\mathrm{MM}}=16/5=3.2\), which is positive and feasible.)