Why Moment Systems Need More Than Algebra
When a model has several unknown parameters, method of moments requires a system of equations: match a suitable collection of theoretical moments to the corresponding sample moments, then solve for the parameters. The algebra may produce more than one candidate, or candidates outside the model’s allowed parameter space. A solution is therefore not complete until it has been checked for both validity and uniqueness.
For a sample \(X_1,\ldots,X_n\), the \(r\)th raw sample moment is \(m'_r=n^{-1}\sum_{j=1}^{n}X_j^r\). If a model with parameter vector \(\boldsymbol{\theta}\) has theoretical raw moments \(\mu'_r(\boldsymbol{\theta})=E[X^r]\), the method-of-moments equations set selected \(m'_r\) equal to \(\mu'_r(\boldsymbol{\theta})\). With two parameters, two equations are typical; the equations need not be linear.
The sample moments are computed from the observed data; the theoretical moments come from the specified model. Keep these roles distinct, and state any restrictions on the parameters before solving. For example, a scale parameter may have to be strictly positive, while probabilities must lie between zero and one, inclusive.
A Reliable Procedure for Solving Moment Systems
A good solution shows where each equation comes from and how the candidate roots are checked. The following sequence is useful for linear and nonlinear systems alike.
Write down the distributional setup and restrictions such as \(b>0\), \(0\le p\le1\), or probabilities that must sum to one.
Use the moments implied by the model, including whether each is a raw moment or a central moment.
Equate theoretical moments to the corresponding observed sample moments. Substitute or eliminate variables, and retain every real candidate root until it has been checked.
Substitute every candidate into the original equations, apply the full parameter-space restrictions, and determine whether any other valid candidates exist.
The last step matters because eliminating a parameter can create a polynomial with several roots, and solving an equation after squaring can introduce extraneous roots. Conversely, selecting one convenient root without checking the others can conceal a genuine failure of uniqueness.
Worked Systems of Moment Equations
Worked Example: A Location-Scale Model with a Unique Solution
Suppose \(Z\) takes the values \(-1\) and \(1\), each with probability \(1/2\), and the model is \(X=a+bZ\), where \(a\) is any real number and \(b>0\). The observed first two raw sample moments are \(m'_1=4\) and \(m'_2=20\). Find the method-of-moments estimates and check uniqueness.
Method. First calculate the needed moments of \(Z\): \(E[Z]=(-1)(1/2)+(1)(1/2)=0\) and \(E[Z^2]=1\). Thus \(E[X]=a+bE[Z]=a\). Also,
Match these theoretical moments to the sample moments:
Substitute \(a=4\) into the second equation: \(16+b^2=20\), so \(b^2=4\), with algebraic roots \(b=2\) and \(b=-2\). The model requires \(b>0\), so only \(b=2\) is admissible. The estimates are \(\hat a=4\) and \(\hat b=2\).
Check. These values give \(E[X]=4\) and \(E[X^2]=4^2+2^2=20\), exactly the stated moments. The negative root satisfies the squared equation but violates the model restriction. Since \(a\) is fixed by the first equation and exactly one root for \(b\) satisfies \(b>0\), the valid solution is unique. Common mistake avoided: reporting both roots for \(b\) without applying the parameter space.
Worked Example: Solving for Two Probabilities
A hypothetical discrete loss model assigns probabilities \(p\) and \(q\) to losses of \(1\) and \(3\), respectively, and probability \(1-p-q\) to a loss of \(0\). The model restrictions are \(p\ge0\), \(q\ge0\), and \(p+q\le1\). Suppose \(m'_1=1\) and \(m'_2=2.2\). Estimate \(p\) and \(q\), and verify that the solution is unique and valid.
Method. The theoretical raw moments are \(E[X]=0(1-p-q)+1p+3q=p+3q\) and \(E[X^2]=0^2(1-p-q)+1^2p+3^2q=p+9q\). The moment equations are
Subtract the first equation from the second: \(6q=1.2\), so \(q=0.2\). Substituting into the first equation gives \(p+3(0.2)=1\), hence \(p=0.4\). The remaining probability is \(1-p-q=1-0.4-0.2=0.4\).
Check. The probabilities are nonnegative and sum to \(0.4+0.2+0.4=1\). Their moments are \(E[X]=0.4+3(0.2)=1\) and \(E[X^2]=0.4+9(0.2)=2.2\), as required. The two equations have a unique solution: subtracting them determines \(q\) uniquely, and then the first determines \(p\) uniquely. Common mistake avoided: using the second raw moment as though it were the variance; here \(E[X^2]=2.2\), whereas the variance would be \(2.2-1^2=1.2\).
Worked Example: Two Valid Solutions in a Nonlinear System
Consider the hypothetical model \(X=\theta+B\), where \(B\) is Bernoulli with \(P(B=1)=p\), \(P(B=0)=1-p\), and the parameter restrictions are \(\theta\ge0\) and \(0\le p\le1\). Suppose \(m'_1=2.4\) and \(m'_2=6\). Determine whether the moment estimates are unique.
Method. Since \(E[B]=p\) and \(B^2=B\), the theoretical moments are
The equations are \(\theta+p=2.4\) and \(\theta^2+(2\theta+1)p=6\). From the first equation, \(p=2.4-\theta\). Substitute this into the second:
The discriminant is \(3.8^2-4(1)(3.6)=14.44-14.4=0.04\), so the roots are \(\theta=(3.8\pm0.2)/2\), or \(\theta=2\) and \(\theta=1.8\). The corresponding probabilities are \(p=2.4-\theta\), giving \(p=0.4\) and \(p=0.6\), respectively. Both parameter pairs satisfy \(\theta\ge0\) and \(0\le p\le1\).
Check both candidates. For \((\theta,p)=(2,0.4)\), the mean is \(2+0.4=2.4\), and the second raw moment is \(2^2+(2(2)+1)(0.4)=4+2=6\). For \((\theta,p)=(1.8,0.6)\), the mean is \(1.8+0.6=2.4\), and the second raw moment is \(1.8^2+(2(1.8)+1)(0.6)=3.24+2.76=6\). Thus both are valid; the solution is not unique.
Common mistake avoided: choosing one root and calling it the estimate without testing the other. A moment system can be satisfied by multiple parameter values, even when every candidate lies in the permitted parameter space.
Why Elimination Works—and What It Does Not Prove
Substitution is valid because the first equation can express one parameter in terms of the other. In the last example, \(\theta+p=2.4\) is equivalent to \(p=2.4-\theta\). Any solution of the original system must satisfy that relation, so substituting it into the second equation preserves every solution. Solving the resulting quadratic and then recovering \(p\) therefore finds all candidates, provided every root is retained and checked in the original equations.
This also explains the distinction between solving and verifying. The polynomial roots are candidates; checking the original equations confirms they solve the moment system, and checking the parameter restrictions confirms they belong to the model. In the example, both roots pass both checks, so there is no basis for selecting just one.
For differentiable theoretical moment functions \(g_1(\theta_1,\theta_2)\) and \(g_2(\theta_1,\theta_2)\), the Jacobian matrix of the moment system is
A nonzero determinant at a solution can indicate that the equations locally distinguish nearby parameter values under the usual differentiability conditions. It is not, by itself, proof that no distant solution exists. In the shifted Bernoulli example, the determinant is \(1-2p\): it equals \(0.2\) at \(p=0.4\) and \(-0.2\) at \(p=0.6\). Both solutions are locally nonsingular, yet the system has two valid solutions. A global uniqueness claim requires checking all solutions in the permitted parameter space.
Common Checks Before Reporting Estimates
Before presenting a solution to a moment system, make the conclusion match what the equations establish. A useful final check is to substitute each candidate into the original moment equations, not merely the polynomial obtained after elimination.
- Check the moment type. Do not equate a raw sample moment to a theoretical variance. For example, \(E[X^2]\) and \(\operatorname{Var}(X)\) differ by \(E[X]^2\).
- Check the parameter space. A negative scale, a probability above one, or a negative implied category probability is not a valid estimate for a model that forbids it.
- Check every root. A quadratic may have two admissible roots, one admissible root, or none. Do not assume the first outcome is unique because one root was found.
- Check the recovered parameters. After solving for one parameter, use the original relation to calculate the others and verify their restrictions too.
- State the conclusion accurately. Report a unique valid solution only after excluding other admissible candidates. If multiple candidates remain, report them all and state that the system is nonunique.
Check Your Understanding
For each system, write the equations, solve them, and check the indicated parameter restrictions.
- A model is \(X=a+bZ\), where \(E[Z]=0\), \(E[Z^2]=1\), and \(b>0\). Given \(m'_1=2\) and \(m'_2=13\), find \(a\) and \(b\). (Answer: \(a=2,\ b=3\), the unique valid solution.)
- A hypothetical loss takes values \(0,2,5\) with probabilities \(1-p-q,p,q\), respectively. Given \(m'_1=1.5\) and \(m'_2=4.5\), solve for \(p,q\) and the probability at zero. (Answer: \(p=0.5,\ q=0.1\), and \(P(X=0)=0.4\), the unique valid solution.)
- Let \(X=\theta+B\), with \(B\) Bernoulli of success probability \(p\), \(\theta\ge0\), and \(0\le p\le1\). Given \(m'_1=3\), \(m'_2=10\), determine whether any real moment solution exists. (Answer: no real solution; substitution gives \(\theta^2-5\theta+7=0\), whose discriminant is \(25-28=-3\).)
- In the same shifted Bernoulli model, suppose \(m'_1=1.7\) and \(m'_2=3.14\). Solve the system and state whether it is unique. (Answer: \(\theta=1.2,\ p=0.5\), the unique valid solution; the quadratic has a repeated root.)
- For the system in Worked Example 3, why is it incorrect to report only \((\theta,p)=(2,0.4)\)? (Answer: \((1.8,0.6)\) also satisfies both moment equations and the parameter restrictions, so both valid solutions must be reported.)