Statistical Identification

Statistical identification is the determination of whether a quantity of interest can be uniquely recovered from the probability distribution of observable data under specified assumptions. In statistics and econometrics, it distinguishes what can in principle be learned from what can be estimated accurately using a particular sample. A quantity is identifiable when different values necessarily have different observable implications; it is nonidentifiable when distinct values can produce exactly the same observable distribution. Identification may concern an entire model, individual parameters, or a particular feature such as a mean or causal effect. (stat.cmu.edu)

Formal definition

Consider a statistical model $\{P_\theta:\theta\in\Theta\}$ , where $\theta$ denotes the parameters and $P_\theta$ is the distribution of the observations. The model is globally identifiable if

P_{\theta_1}=P_{\theta_2} \quad\Longrightarrow\quad \theta_1=\theta_2.

Thus, the mapping $\theta\mapsto P_\theta$ must be an injective function. Two parameter values producing the same observable distribution are called observationally equivalent. (pages.stern.nyu.edu)

Frequently the target is not the complete parameter vector but a function $\psi(\theta)$ . This target is identifiable if

P_{\theta_1}=P_{\theta_2} \quad\Longrightarrow\quad \psi(\theta_1)=\psi(\theta_2).

An unidentified model can therefore contain identifiable quantities. For example, observations may determine the sum of two parameters without determining either parameter separately. More generally, identification asks whether the target is constant across all structures compatible with the observables and maintained assumptions. (arxiv.org)

Identification versus estimation

Identification treats the observable distribution as known. Estimation uses finite data to approximate an identified quantity or set, while statistical inference assesses uncertainty caused by sampling. These are distinct problems: an identifiable parameter may be estimated imprecisely, whereas an unidentified parameter cannot become uniquely recoverable merely by collecting more observations of the same kind under unchanged assumptions. (stat.cmu.edu)

For example, consider

Y_i\sim N(a+b,\sigma^2).

The observable distribution identifies $\mu=a+b$ , but replacing $(a,b)$ with $(a+c,b-c)$ leaves it unchanged for every $c$ . No sample size can distinguish these alternatives. Estimating $\mu$ precisely does not identify its decomposition into $a$ and $b$ . (mc-stan.org)

Identification is consequently necessary for an estimator to consistently recover a target over all observationally equivalent possibilities. It is not, by itself, a guarantee of useful finite-sample performance or of the regularity conditions required by familiar asymptotic methods. (arxiv.org)

Forms of identification

Global, local, and generic identification

Global identification rules out observationally equivalent alternatives throughout the parameter space. Local identification rules them out within a neighborhood of a particular parameter value; equivalent alternatives may still exist elsewhere. Generic identification establishes identification except on a specified exceptional subset, commonly a lower-dimensional or measure-zero set. Its meaning depends on how that exception is defined. (pages.stern.nyu.edu)

Under appropriate regularity conditions, local identification in a finite-dimensional parametric model can be characterized by nonsingularity of the Fisher information matrix. Such conditions matter: an information-matrix calculation is not an unconditional test of global identification. (pages.stern.nyu.edu)

Point and partial identification

A target is point identified when the observable distribution and assumptions determine one value. Under partial identification, they determine a set of admissible values instead. For a known observable distribution $P$ , the identified set is

\Psi_I(P)= \{\psi(\theta):P_\theta=P,\ \theta\in\Theta\}.

A singleton gives point identification. A larger set can still be informative if it excludes possible values. Sharp bounds describe exactly the admissible range: each included value must be attainable under the assumptions. An interval that merely contains that range is an outer bound, not necessarily a sharp characterization. (arxiv.org)

An identified set is not a confidence interval. The former describes ambiguity remaining even with perfect knowledge of the observable distribution; the latter addresses finite-sample uncertainty. In partially identified models, confidence procedures may target the unknown parameter or the entire identified set, with different coverage requirements. (arxiv.org)

Weak identification

Weak identification occurs when changes in parameters have only small observable consequences, making alternatives difficult to distinguish. Unlike exact nonidentification, it need not involve identical distributions. In instrumental-variable models, a familiar source is a weak relationship between instruments and endogenous explanatory variables. Conventional normal approximations and standard confidence procedures can then be misleading, motivating inference methods robust to weak identification. (tandfonline.com)

Representative examples

Linear regression

In a fixed-design linear regression model,

Y=X\beta+\varepsilon,\qquad \varepsilon\sim N(0,\sigma^2I),

the coefficient vector is identified when the design matrix $X$ has full column rank. If $Xv=0$ for a nonzero vector $v$ , then

X(\beta+v)=X\beta,

so $\beta$ and $\beta+v$ imply the same observable distribution. Exact collinearity therefore leaves individual coefficients unidentified, although some combinations may remain identifiable. This concerns exact redundancy, not merely strong correlation between predictors. (mc-stan.org)

Mixture models and label symmetry

In a two-component Gaussian mixture model,

f(y)=\pi\phi(y;\mu_1,\sigma_1^2) +(1-\pi)\phi(y;\mu_2,\sigma_2^2),

exchanging the component labels gives the same density. The labeled parameters are consequently not globally identifiable without a labeling convention or distinguishing restrictions. The mixture density and predictions that are invariant to labels can nevertheless remain well-defined. This symmetry also creates difficulties for interpreting component-specific posterior summaries and diagnosing computational convergence. (mc-stan.org)

Missing outcomes and bounds

Suppose $Y$ is binary, $R=1$ indicates that it is observed, and the available data reveal

p=P(R=1),\qquad q=P(Y=1\mid R=1).

Assuming $0<p<1$ , the population mean satisfies

E[Y]=pq+(1-p)P(Y=1\mid R=0).

Without restrictions on unobserved outcomes, their success probability can range from zero to one. Hence

pq\leq E[Y]\leq pq+1-p.

These are sharp bounds: every value within them can be generated by a suitable distribution of missing outcomes. Additional assumptions about the missingness process can narrow the range or produce point identification, but that extra precision depends on those assumptions. (doi.org)

Causal identification and research design

In causal inference, the target concerns what would happen under an intervention rather than merely an observed association. The observed relationship between treatment and outcome does not generally determine a causal effect because different mechanisms of treatment selection can produce the same observations. (hsph.harvard.edu)

Let $A$ be treatment, $Y(a)$ a potential outcome, and $X$ observed covariates. A common identification argument uses:

Consistency: the observed outcome equals $Y(a)$ when $A=a$ .
Conditional exchangeability: $Y(a)$ is independent of treatment given $X$ .
Positivity: treatment level $a$ has positive probability for relevant covariate values.

Together, these permit identification through

E[Y(a)] = E_X\!\left[E[Y\mid A=a,X]\right].

This expression connects an intervention-defined target to quantities in the observed distribution. The conditions are substantive assumptions, not consequences of fitting a regression. (hsph.harvard.edu)

An identification strategy specifies the design features or restrictions that establish such a connection. Randomization, covariate adjustment, and instrumental variables support different identification arguments. Instrumental-variable approaches require restrictions beyond an instrument’s association with treatment, and weak relevance can undermine conventional inference even when the assumed restrictions establish identification. (hsph.harvard.edu)

Restrictions, normalization, and Bayesian analysis

Identification is always relative to what is observed and what is assumed. Observing additional variables, changing an experimental design, or restricting the admissible model class can separate previously equivalent structures. Restrictions may be substantive, such as excluding a direct effect, or conventional, such as fixing a scale or choosing a component-label ordering. A normalization chooses a representation; it should not be confused with evidence supporting a substantive restriction. (stat.cmu.edu)

In Bayesian inference, a prior distribution can produce a proper and computationally tractable posterior where the likelihood alone does not distinguish parameters. This can be useful, but it does not make the observable-distribution mapping injective. Along observationally equivalent alternatives, relative posterior support reflects the prior rather than discrimination supplied by the likelihood. (mc-stan.org)

Likewise, regularization or an imposed constraint can select one solution from many observationally equivalent possibilities. A unique numerical answer is therefore not sufficient evidence that the underlying target has been identified by the data. (mc-stan.org)

Historical development and scope

Identification became central to twentieth-century econometrics through the study of structural economic models and simultaneous equations. Work by Tjalling Koopmans and colleagues in the late 1940s and around 1950 distinguished observational equivalence from uncertainty due to finite samples. Thomas Rothenberg’s 1971 paper developed general parametric identification criteria and related local identification to the information matrix under regularity conditions. (arxiv.org)

Partial-identification research, prominently developed by Charles Manski, broadened the emphasis from recovering a unique value to determining what ranges are supported by data and credible assumptions. The framework applies to missing data, treatment effects, and structural models; related identification problems also arise in latent-variable modeling and other statistical disciplines. (link.springer.com)

Identification does not establish that maintained assumptions are true. A model can identify a parameter internally while relying on restrictions that the observations cannot independently verify. Conversely, failure of point identification need not make analysis uninformative: bounds can exclude values, identify the sign of a target, or show explicitly how conclusions depend on stronger assumptions. (hsph.harvard.edu)

References

A general theory of identificationarxiv.org
Lecture 23, Estimating Causal Modelsstat.cmu.edu
The Identification Zoo: Meanings of Identification in Econometricsaeaweb.org
Microeconometrics with Partial Identificationarxiv.org
Partial identification with missing data: concepts and findingsdoi.org
Partial Identification of Probability Distributionslink.springer.com
Causal Inference: What Ifhsph.harvard.edu
Problematic Posteriorsmc-stan.org
2 Label switching in mixture modelsmc-stan.org
12 Priors for identifiabilitymc-stan.org