This article$^1$ on p.16 specifies Haldane's prior as:

$$p(\theta) = \frac{1}{θ(1−θ)}$$.

However, other$^2$ source on p.6 specifies Haldane's prior as proportional to $\frac{1}{θ(1−θ)}$, i.e. $$p(\theta) \propto \frac{1}{θ(1−θ)}$$.

Could anyone clarify which expression is the accurate one.

1. Approximation of improper priors

2. Bayesian Analysis of Some Common Distributions

  • 6
    $\begingroup$ This is an improper prior because its integral diverges. Since it is improper, constants of proportionality have no relevance because no normalizing constant exists. $\endgroup$ – whuber Jan 31 '18 at 21:02

Haldane prior is beta distribution with parameters $\alpha = \beta = 0$. So it is

$$ f(p) = \frac{p^{\alpha-1} (1-p)^{\beta-1}}{B(\alpha, \beta)} = \frac{p^{-1}(1-p)^{-1}}{B(0, 0)} $$

where $B(0, 0)$ is the normalizing constant that is infinite as described in Wikipedia:

The function $p^{-1}(1-p)^{-1}$ can be viewed as the limit of the numerator of the beta distribution as both shape parameters approach zero: $\alpha, \beta \to 0$. The Beta function (in the denominator of the beta distribution) approaches infinity, for both parameters approaching zero, $\alpha, \beta \to 0$. Therefore, $p^{-1}(1-p)^{-1}$ divided by the Beta function approaches a 2-point Bernoulli distribution with equal probability $1/2$ at each Dirac delta function end, at $0$ and $1$, and nothing in between, as $\alpha, \beta \to 0$.

So Haldane prior is not a proper distribution. It is an abstract idea of what would be the beta distribution be if it had $\alpha = \beta = 0$ parameters. As a distribution, it is rather not usable, yet it can be used as an "uninformative" prior for binomial distribution. It is often described in it's approximate form $f(p) \propto p^{-1}(1-p)^{-1}$, since the normalizing constant is meaningless.

| cite | improve this answer | |
  • $\begingroup$ Thanks @Tim; can it be shown how the Beta function approaches $\infty$ as $\alpha, \beta→0$. $\endgroup$ – AlexMe Jan 31 '18 at 22:01
  • $\begingroup$ @AlexStat it follows from $\Gamma(0) = \infty$ $\endgroup$ – Tim Jan 31 '18 at 22:30
  • 1
    $\begingroup$ @Tim: in which sense $f(p) \propto p^{-1}(1-p)^{-1}$ is approximate? There is no imprecision in defining the prior without a constant since all choices of a constant lead to the same posterior (if defined) or none( if undefined). $\endgroup$ – Xi'an Mar 2 '18 at 18:35
  • $\begingroup$ @Xi'an as a distribution it lacks the constant. $\endgroup$ – Tim Mar 2 '18 at 18:41
  • 1
    $\begingroup$ Imho, it is a measure (or the density of a measure against the Lebesgue measure onm $(0,1)$) rather than a distribution. As the notion of distribution is not set outside "probability distributions". $\endgroup$ – Xi'an Mar 2 '18 at 18:48

The second expression is correct as this is an improper distribution, i.e. it doesn't integrate to $1$. Thus it doesn't have a density and you can only specify it up to proportionality.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.