A model is identifiable if a single set of parameters can be found that will yield the best fit.

A model is identifiable if a single set of parameters can be found that will yield the best fit.

For example, consider the classic ANOVA model:
$$ y_{ij} = \mu + \alpha_j + \varepsilon_i $$ where $y_{ij}$ represents observed scores decomposed into a population mean, $\mu$, a mean shift, $\alpha_j$, associated with condition $j$, and each unit's individual divergence from their condition's mean, $\varepsilon_i$. When there are $J$ conditions, there are $J+1$ parameters to fit in this model. Without additional constraints, this model is unidentifiable; for instance if three conditions had means $3$, $4$, and $5$, they could be fit equally well with:
\begin{array}[llll] \mu \mu = 1 \qquad &\alpha_1 = \; \; \ 2 \qquad &\alpha_2 = \; \; \ 3 \qquad &\alpha_3 = \; \; \ 4 \\ \mu = 4 \qquad &\alpha_1 = -1 \qquad &\alpha_2 = \; \; \ 0 \qquad &\alpha_3 = \; \; \ 1 \\ \mu = 9 \qquad &\alpha_1 = -6 \qquad &\alpha_2 = -5 \qquad &\alpha_3 = -4 \\ ... \end{array} (Therefore, in practice the ANOVA is given additional constraints such as $\frac{1}{N}\sum_j n_j\alpha_j = 0$.)

Although issues with identifiability are trivial in the above example, they can be more subtle in other contexts. Identifiability concerns can arise in fitting structural equations models (), for example.

history | excerpt history