Manfred Deistler and Gerhard Tintner, Identifiability (1981)

Manfred Deistler and Gerhard Tintner’s edited-volume survey contribution examines when observable behavior uniquely determines a statistical model’s unobservable parameters. Its two substantive sections connect identification in errors-in-the-variables models with the algebraic identification of multivariate autoregressive moving-average (ARMA) systems. The organizing distinction is between estimating parameters from observations and establishing whether those parameters can be uniquely recovered even with complete observational knowledge.

In this case a problem preceding estimation is, whether it is possible to determine uniquely the "internal" parameters from the complete knowledge of the "external" behavior of the system.

Identification therefore concerns the information supplied by a model and its restrictions, not merely the quantity of available data. The first substantive section surveys distributional assumptions, known measurement-error variances, and temporal structure as sources of identification. The second develops a more systematic account of observational equivalence, distinguishing restrictions grounded in substantive knowledge from conventions that select a unique mathematical representation.

The errors-in-the-variables discussion begins with characteristic functions and linear combinations of independent latent variables. Symmetric stable distributions provide an important example: linear conditional expectations need not imply normality. Deistler and Tintner emphasize the economic relevance of nonnormal, heavy-tailed distributions and suggest that econometric practice has neglected results available from probability theory. Distributional form matters directly for identification. In the elementary structure (U=X+Z_1), (V=\beta X+Z_2), with the latent variable independent of the errors, the authors give a characteristic-function condition for recovering the slope and then state:

Also, if the error variables $Z_1, Z_2$ are normally distributed, the parameter $\beta$ is identifiable if and only if $X$ is not normally distributed.

Nonnormality here supplies identifying information rather than simply violating a convenient estimation assumption. The discussion of functional relationships extends this point to relations between error-free systematic components when every observed variable contains measurement error. The authors recall the importance of such relationships in early econometrics and criticize the later tendency to neglect measurement errors while concentrating on equation disturbances.

For the basic two-variable model, observable means, variances, and covariance do not separately determine the slope, latent variance, and error variances. Identification requires additional information, such as a known error variance or the ratio of error variances; knowing both can produce overidentification. The survey also considers cumulants under nonnormality, instrumental variables, and Bayesian prior assumptions. Its treatment of likelihood warns that mechanical maximization does not resolve identification or incidental-parameter difficulties. One local statement requires caution: the text calls an error-variance estimator consistent while displaying a probability limit equal to half the target variance. The displayed result does not support that description.

Temporal information offers another route. A latent first-order autoregression allows lagged covariances of the measured series to separate latent persistence and variance from measurement noise. Dynamics thus supply information absent from contemporaneous moments alone. The weighted-regression discussion considers smooth deterministic components plus stochastic errors: finite differences reduce the smooth components and permit estimation of error covariances. Generalized eigenvalue equations then support tests of relations among the systematic components and estimation of their coefficients.

The ARMA section shifts from separating latent variation and measurement error to distinguishing alternative polynomial descriptions of the same process. Under the stated stability, invertibility, full-rank, and normalization conditions, the spectral density determines the transfer function and innovation covariance. Yet this does not uniquely determine the autoregressive and moving-average polynomial matrices.

However, for given $T(z)$, $A(z)$, and $B(z)$ are not unique and this is what causes the main part of the identifiability problem.

The authors formalize this ambiguity through left matrix fraction descriptions. After excluding common polynomial factors that introduce redundant dynamics, they characterize observational equivalence by unimodular left multiplication. Identifiability then means making the mapping from admissible polynomial matrices to observable behavior injective.

Two solutions follow. Structural identification uses prior knowledge about coefficients and lag lengths. The authors derive a sufficient rank condition under which affine parameter restrictions force the observational-equivalence transformation to be the identity. Overidentification can additionally exclude entire equivalence classes. Canonical identification instead selects one representative within each class without claiming substantive knowledge of its coefficients.

The Echelon form implements this second strategy. Row reduction minimizes the sum of row degrees; further normalization and degree restrictions establish uniqueness. The resulting dynamical indices organize the system’s lag structure. An alternative construction uses linear dependencies among rows of the transfer function’s Hankel matrix, connecting polynomial identification with minimal system realization.

The closing subsection explains why uniqueness alone is insufficient for estimation.

Of special interest are the structure of the parameter space (i.e., the set of free parameters in the triplets (A(z), B(z), Σ)) and the properties of the mapping relating (T(z), Σ) to the free parameters in (A(z), B(z), Σ).

With suitable strengthened assumptions, parameter spaces are open and the recovery mapping is continuous. Consistency for the transfer function and innovation covariance can then yield strong consistency for structural parameters or Echelon parameters with fixed dynamical indices. The contribution’s central achievement is this connection between informational uniqueness, algebraic representation, and statistical recovery: successful estimation depends not only on identifying a model, but also on choosing a parameterization whose behavior supports inference.