So suppose we have a model $f(x|\theta)$ that is theoretically identifiable, so that $\theta_1 \neq \theta_2$ implies $f(x|\theta_1) \neq f(x|\theta_2)$.

However, suppose that data collection is very limited and that, for all samples $X$ taken in practice, for every $\theta$ there exists a $\hat{\theta}$ such that $f(X|\theta) \approx f(X|\hat{\theta})$.

And an example of this would be if $X=(x_1,\dots,x_n)$ is a time series model in equilibrium, i.e., it is basically constant, over the time period considered, such that $X_i \approx x_c$ for all $i=1,2,3,\dots,N.$

If this was the case, and if for example we had $f(x|\theta) \propto \exp \left(-\prod_i \left(\Delta x_i-\alpha x_i^\beta \right)^2 \right)$, then the model would, for practical considerations, be unidentifiable since we would have

$$\log f(x|\theta) \propto - \prod_i \left(\Delta x_i-\alpha x_i^{\beta_1}\right)^2 \approx - \prod_i \left(\Delta x_i-\alpha x_c^{\beta_1}\right)^2 = -\prod_i \left(\Delta x_i-\tilde{\alpha} x_c^{\beta_2} \right)^2\\ \approx -\prod_i \left(\Delta x_i-\tilde{\alpha} x_i^{\beta_2}\right)^2 $$

for any $\beta_1,\beta_2 > 0$, where $\hat{\alpha} = \alpha x_c^{\beta_1-\beta_2}$. If the above was the case, then it would be practically impossible to determine between $(\alpha,\beta_1)$ and $(\hat{\alpha},\beta_2)$, even though the model is theoretically identifiable. I am curious if there is a word or phrase to describe this situation? And whether my concerns make sense.

  • $\begingroup$ Your example doesn't really illustrate what I think you are asking about. The "parameter" $\alpha$ is just the scaling constant for the density, so it is a function of $\beta$. Moreover, although you have not specified the bounds, the example density is monotone in $\beta$. Can you rethink your example and use another that better illustrates your query? (A more realistic example would be when you have explanatory variables in a regression that are almost perfectly correlated.) $\endgroup$
    – Ben
    Commented Feb 18, 2021 at 6:02
  • $\begingroup$ Thanks @Ben for pointing that out, I chose the first density that came to mind without thinking that through. I updated it to be of Gaussian form, as is typically assumed in time-series models. Is the problem more sensical now? $\endgroup$
    – user281754
    Commented Feb 18, 2021 at 6:35

1 Answer 1


Surely this is the case for almost every statistical model with a continuous valued parameter $\theta$ and a likelihood that is reasonably smooth. Perhaps this is an issue of computational precision rather than a statistical issue?

For instance, take the likelihood to be that of a Normal with mean $\theta$ and known variance $1$. This is an identifiable model.

Thus, for a sample size of $1$, $$\log f(x \mid \theta) = C -(x - \theta)^2$$ where $C$ is a scaling constant which ensures the likelihood is valid.

Now $$\log f(x \mid \theta + \delta) = C -(x - \theta - \delta)^2$$.

Now if $\delta$ is sufficiently small relative to $\theta$, (say $\theta = 5$, $\delta = 10^{-20}$) a computer could never tell $\log f(x \mid \theta)$ apart from $\log f(x \mid \theta + \delta)$. There is a mathematical distinction, but if a computer can't tell the difference then, to a modern statistician, they are effectively the same.

On a similar theme, it is in practice almost impossible to tell the difference between a sech distribution and the logistic distribution for certain parameter values, despite them having different probability density functions.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.