Statistics Glossary → Model identification (std.lv, unit loading)
Model identification (std.lv, unit loading)concept
The condition that a model's parameters can be estimated from the observed data as a single set of values; in latent variable models this requires constraints such as fixing the scale.
A model is identified if different parameter values cannot produce the same distribution of the data. The simplest necessary condition is that the number of free parameters does not exceed the amount of independent information in the data (e.g. the p(p + 1)/2 variances and covariances in structural equation models); the difference is the model's degrees of freedom. A latent factor, moreover, has no natural scale: doubling the loadings and quartering the factor variance yields the same covariances. This indeterminacy is resolved in two common ways: fixing one item's loading on each factor to 1 (unit loading, marker item), or fixing the factor variance to 1 (std.lv). Both give the same fit; only the scale of the parameters changes.
In instrumental variables regression, identification requires at least one excluded instrument for each endogenous regressor (the order condition); if there are more instruments than endogenous variables the model is overidentified, and the surplus can be tested with the Sargan or Hansen J test. In the Heckman selection model, a variable that appears in the selection equation but not in the outcome equation (an exclusion restriction) strengthens identification; without it, identification rests only on the functional form of the normality assumption and is fragile. A model that is identified in theory can still be empirically underidentified if the data contain near-collinear structures, which shows up as very large standard errors or convergence problems.
An everyday example
A tailor cannot measure someone's 'body size' directly; he only takes chest, waist and shoulder measurements. Without choosing a unit for 'body size' (e.g. treating the chest measurement as 1, or expressing size as a standard score), the numerical value of this abstract quantity stays undetermined. Whichever unit he picks, the ranking of people and the relationships between measurements do not change; only the scale of the numbers does.
In MerQur
In confirmatory factor analysis the 'Additional output: Factor variance = 1 identification (std.lv)' fixes factor variances at 1 and reports all loadings, including the first item's, with standard errors. Instrumental variables regression gives overidentification diagnostics (Sargan, Hansen J); in sample selection and truncated regression the identification of the Heckman model rests on the exclusion restriction in the selection equation, and a warning is given when it is missing.
Analyses where this concept appears
Instrumental Variables Regression (2SLS)Confirmatory Factor Analysis (CFA)Sample Selection and Truncated Regression (Heckman)