Why Identifiability Matters
The method of moments turns a statistical model into equations: theoretical moments are matched to sample moments, and the resulting equations are solved for unknown parameters. Solving the equations is not enough, however. The equations might have several parameter solutions, or they might not distinguish some parameters at all. Identifiability through moments asks whether the particular moments being used determine the model parameters uniquely.
This is a property of the model and the selected moments, not a guarantee that every sample will produce sensible estimates. The distinction is useful when choosing moment equations and interpreting their solutions. It is aligned with the SOA VEE Mathematical Statistics syllabus.
For a random variable \(X\), a common choice is the first \(k\) raw moments \(m_j(\theta)=E_\theta[X^j]\), provided those expectations exist. Given independent observations \(X_1,\ldots,X_n\) from the model, the corresponding sample moments are \(\hat m_j=n^{-1}\sum_{i=1}^n X_i^j\). The method-of-moments equations set \(m_j(\theta)=\hat m_j\). Identifiability concerns whether a moment vector can come from more than one parameter value; it does not assert that every possible sample moment vector is attainable.
Solving the Equations Is the Global Test
The most direct way to assess global identification is to compare all parameter values that produce the same selected moments. In a concrete problem, this usually means solving the moment equations and checking whether the solution is unique over the stated parameter space. A unique solution establishes identification for that moment vector; to establish global identification of the model, one must show that no two distinct parameter values anywhere in \(\Theta\) have the same moment vector.
The number of equations alone does not settle the question. Having as many selected moments as parameters makes a square system, but a nonlinear square system can still have multiple solutions. Conversely, a special structure of the model may allow parameters to be recovered even when a chosen set of equations looks different from a simple one-equation-per-parameter system.
Worked Example: Recovering Both Endpoints of a Uniform Model
Suppose \(X\sim\operatorname{Uniform}(a,b)\), with parameter space \(a<b\), and use its first two raw moments. The observed sample moments are \(\hat m_1=6\) and \(\hat m_2=39\). Determine whether the moment equations identify the endpoints and use the Jacobian as a local diagnostic.
Method and equations. For a uniform distribution on \((a,b)\), direct integration gives \(m_1=(a+b)/2\) and \(m_2=(a^2+ab+b^2)/3\). Equivalently, its variance is \((b-a)^2/12=m_2-m_1^2\). Matching the first moment gives
The second equation gives \(m_2-m_1^2=39-36=3\), so
Because \(a<b\), the width is \(b-a=6\), not \(-6\). Combining the sum and difference gives
Global check. The mean fixes the sum \(a+b\), and the variance fixes the positive width \(b-a\). Both together determine exactly one ordered pair under \(a<b\). Thus the moment map is globally one-to-one on this parameter space.
Jacobian check. For \(M(a,b)=(m_1,m_2)\), its Jacobian matrix is
At \((a,b)=(3,9)\), the determinant is \((9-3)/6=1\), which is nonzero. The calculation supports local invertibility there, while the sum-and-width argument establishes the stronger global conclusion. Common mistake avoided: the determinant is \((b-a)/6\), not \((b-a)/12\); the latter misses part of the determinant calculation.
Check. The recovered endpoints give mean \((3+9)/2=6\) and second raw moment \((9+27+81)/3=39\), matching both specified moments.
The Jacobian as a Local Diagnostic
For differentiable moment functions, the \(k\)-by-\(p\) Jacobian has entry \(\partial m_j/\partial\theta_r\) in row \(j\), column \(r\). It describes how the selected moments change under small changes in the parameters. When \(k=p\), a nonzero determinant at an interior parameter value is a useful local test, subject to the regularity conditions below.
The hypotheses matter. Differentiability at just one point is not enough for this inverse-function conclusion; continuous differentiability on a neighborhood is a standard sufficient regularity condition. Also, a zero determinant does not prove nonidentifiability. It means this particular nonsingular-Jacobian test has not established local invertibility. The moment equations must still be examined directly or with another appropriate argument.
Worked Example: A Nonsingular Jacobian but Global Ambiguity
Consider a model in which \(X\) takes the values \(\theta\) and \(-\theta\), each with probability \(1/2\), where the parameter space is \(\Theta=\mathbb{R}\). Use the first two raw moments. Suppose the matched sample moments are \(\hat m_1=0\) and \(\hat m_2=9\). Determine the solutions and interpret the Jacobian.
Method. Compute the moments from the two possible values:
The first equation \(m_1(\theta)=0\) holds for every \(\theta\). The second equation requires \(\theta^2=9\), so
There are two parameter solutions, so the selected moments do not identify \(\theta\) globally on \(\mathbb{R}\). The Jacobian is the two-by-one matrix
At either solution it has rank \(1\), equal to the number of parameters. The scalar moment \(m_2(\theta)=\theta^2\) has nonzero derivative at both \(3\) and \(-3\), so it is locally invertible around each solution. Nevertheless, the two separate local neighborhoods do not remove the global ambiguity between the two solutions. In fact, the full distribution is also the same at \(\theta=3\) and \(\theta=-3\), since both assign probability \(1/2\) to each of \(-3\) and \(3\).
Check. At either solution, the first moment is \(0\) and the second moment is \(9\). Common mistake avoided: a local Jacobian check cannot be used to claim global uniqueness.
A Two-Parameter Moment Map with a Unique Solution
The next example shows how solving the equations can establish uniqueness, while the Jacobian explains why small changes in the moments can locally be translated into small changes in the parameters.
Worked Example: Identifying Shift and Scale in an Exponential Model
Let \(X=\mu+Z\), where \(Z\) has an exponential distribution with rate \(\lambda>0\), and let \(\mu\in\mathbb{R}\). Thus \(E[Z]=1/\lambda\) and \(\operatorname{Var}(Z)=1/\lambda^2\). The matched raw moments are \(\hat m_1=4\) and \(\hat m_2=20\). Find the parameter solution and assess identification.
Method and moment equations. Write \(t=1/\lambda\), which is positive. Since \(E[X]=\mu+t\) and \(\operatorname{Var}(X)=t^2\), the second raw moment is \(E[X^2]=\mu^2+2\mu t+2t^2\). The equations are therefore
For any solution, subtract the square of the first moment from the second raw moment. This gives the variance:
Since \(t>0\), \(t=2\). The first equation then gives \(\mu=4-2=2\), and \(\lambda=1/t=1/2\). The positivity restriction on \(t\) is essential: it selects one scale from the two algebraic roots of \(t^2=4\).
Global check. For every parameter pair in this model, \(t\) is the positive square root of \(m_2-m_1^2\); then \(\mu=m_1-t\) and \(\lambda=1/t\). These recovery equations show that no other pair in the stated parameter space produces the same two moments.
Jacobian check. In coordinates \((\mu,t)\), the moment map has Jacobian
At \((\mu,t)=(2,2)\), the determinant is \(4\), which is nonzero. This agrees with local identification. The equation-based recovery argument, rather than the Jacobian alone, establishes global identification. Common mistake avoided: treating \(t\) as an unrestricted real number would incorrectly retain the negative root, even though \(t=1/\lambda>0\).
Check. With \(\mu=2\) and \(t=2\), \(m_1=2+2=4\) and \(m_2=2^2+2(2)(2)+2(2^2)=4+8+8=20\), as required.
How to Use the Diagnostic Carefully
A practical analysis begins with the model’s parameter space and the exact moments being used. Then solve the moment equations, retaining the parameter restrictions as part of the problem. When the equations are nonlinear, different algebraic roots may fall inside or outside the stated parameter space; uniqueness must be judged among allowed parameter values.
Specify each selected theoretical moment as a function of the parameter vector, and state where the moments exist.
Ask whether \(M(\theta)=M(\eta)\) can hold for distinct allowed values \(\theta\) and \(\eta\). For a particular data vector, solve the matched equations and inspect all solutions.
For a square, continuously differentiable moment map, a nonzero determinant at an interior point gives local invertibility. It does not rule out a second solution elsewhere.
Say whether the evidence gives uniqueness for the observed moment equations, global identification on the parameter space, or only local identification near a point.
For a rectangular Jacobian, its rank describes the number of locally independent first-order directions represented by the selected moments. Full column rank can give a local identification conclusion when a suitable set of \(p\) moment components forms a square, continuously differentiable map with nonsingular Jacobian. Rank deficiency is not, by itself, a proof that the map is non-injective: higher-order behavior can still distinguish parameter values. For example, the one-dimensional map \(g(\theta)=\theta^3\) is one-to-one even though \(g'(0)=0\).
Common Exam Traps
- Equating a square system with uniqueness. The equation count does not rule out multiple roots. Solve the system and check the full parameter space.
- Calling local identification global. A nonzero determinant supports a conclusion only in a neighborhood under the inverse function theorem’s conditions.
- Overinterpreting a zero determinant. It shows that the nonsingular-Jacobian test does not apply at that point; it does not establish nonidentification.
- Ignoring restrictions. A rate, scale, probability, or ordered pair may be restricted. Apply those restrictions when deciding which equation solutions are valid.
- Making a determinant arithmetic error. For the uniform moment map above, differentiating the second moment gives the row \(((2a+b)/3,(a+2b)/3)\). The determinant simplifies to \((b-a)/6\), not \((b-a)/12\).
Check Your Understanding
For each problem, distinguish conclusions about the particular moment equations from conclusions about local or global identification.
- A uniform model has \(a<b\) and matched moments \(m_1=10\), \(m_2=109\). Solve for \(a,b\), and give the Jacobian determinant at the solution. (Answer: \(a=10-3\sqrt{3}\), \(b=10+3\sqrt{3}\); determinant \(=\sqrt{3}\), so the map is locally invertible there and the ordered endpoints are globally identified.)
- Let \(X=\mu+Z\), where \(Z\) is exponential with rate \(\lambda>0\), \(\mu\in\mathbb{R}\), and the matched moments are \(m_1=7\), \(m_2=57\). Recover \(\mu,\lambda\). (Answer: \(t=1/\lambda=\sqrt{8}=2\sqrt{2}\), \(\mu=7-2\sqrt{2}\), and \(\lambda=1/(2\sqrt{2})\); the positive scale restriction gives a unique pair.)
- For the symmetric two-point model \(P(X=\theta)=P(X=-\theta)=1/2\), where \(\theta\in\mathbb{R}\), the matched moments are \(m_1=0\), \(m_2=16\). List the solutions and state whether the second moment is locally invertible at each. (Answer: \(\theta=4,-4\); \(dm_2/d\theta=2\theta\) is nonzero at both, so the map is locally invertible around each solution but not globally one-to-one.)
- Suppose \(m_1=\alpha+\beta\) and \(m_2=\alpha^2+\beta^2\), with \((\alpha,\beta)\in\mathbb{R}^2\), and matched moments \(m_1=5,m_2=13\). Find the ordered parameter solutions and calculate the Jacobian determinant at either solution. (Answer: the solutions are \((2,3)\) and \((3,2)\); \(\det J=2(\beta-\alpha)\), equal to \(2\) at \((2,3)\) and \(-2\) at \((3,2)\), so each is locally identifiable but the pair is not globally identified.)
- A one-parameter moment map is \(g(\theta)=\theta^3\) on \(\mathbb{R}\). At \(\theta=0\), its derivative is zero. Does this establish nonidentification? (Answer: No. \(g\) is globally one-to-one on \(\mathbb{R}\); a zero derivative means the nonsingular-derivative test does not establish local invertibility there, not that identification fails.)