2
$\begingroup$

The following question is motivated by material from pages 358-364 of Blitzstein and Hwang's Introduction to Probability. (NB: I have modified the book's notation in various places to make my question clearer.)


First, suppose that $\{N_t\}_{t \in \mathbb{R}^+}$ is a family of r.v.s indexed by the set $\mathbb{R}^+$ of positive reals, and that for each $t \in \mathbb{R}^+$, we have $N_t \sim \mathrm{Poisson}(\lambda t)$, where $\lambda$ is an unknown parameter, independent of $t$.

Or, to put it in Bayesian terms, we have

$$ N_t\mid\lambda \sim \mathrm{Poisson}(\lambda t) $$

Now, take1 $\mathrm{Gamma}(n_0, t_0)$ as the (conjugate) prior distribution for $\lambda$. I.e.

$$ \lambda \sim \mathrm{Gamma}(n_0, t_0) $$

Starting from these assumptions, the authors derive (p. 362-364) the posterior distribution for $\lambda$ as

$$ \lambda \mid (N_t = n) \sim \mathrm{Gamma}(n_0 + n, t_0 + t) $$

Furthermore, authors derive (p. 364) the posterior expectation for $\lambda$ as

$$ \operatorname{E}(\lambda \mid N_t = n) = \frac{n_0 + n}{t_0 + t} $$

These results suggest that the first parameter of a $\mathrm{Gamma}$ distribution is akin to a count (of occurrences), and the second one is akin to an length of time.

Under this interpretation, the prior distribution $\mathrm{Gamma}(n_0, t_0)$ would be obtained from the total number of occurrences ($n_0$) observed over one or more intervals of time adding up to a total of $t_0$ (time units). To get the posterior distribution, we update the parameters of the prior's $\mathrm{Gamma}$ by adding a number of occurrences ($n$) observed during a new interval of time of length $t$, so that now we have a total of $n_0 + n$ occurrences observed in a cumulative interval of length $t_0 + t$.


On the other hand, on p. 358 the authors note that a $\mathrm{Gamma}(1, \nu)$ distribution is equivalent to an $\mathrm{Exponential}(\nu)$ distribution, defined as the distribution whose PDF is $f(t) = \nu e^{-\nu t}$. In this case, the second parameter of the $\mathrm{Gamma}$ distribution seems to be behaving like a rate (occurrences per unit time), rather than a length of time.


I'm puzzled by these two radically different interpretations of the $\mathrm{Gamma}$ distribution's second parameter (first as a length of time, and later as a rate).

Is my reasoning above wrong? If not, is there some way to unify or rationalize such divergent interpretations?


1 The book uses the convention that $\mathrm{Gamma}(a, b)$ is the distrubtion whose PDF is $f(x) = \frac{(b x)^a e^{-b x}}{x \Gamma(a)},\; x > 0$.

$\endgroup$
2
  • $\begingroup$ If we have a total of $n_0+n$ occurrences observed in a cumulative interval of length $t_0+t$, then $\frac{n_0 + n}{t_0 + t}$ looks like a "rate" to me. Am I missing something? $\endgroup$ – Just_to_Answer Aug 9 '17 at 1:30
  • $\begingroup$ @Just_to_Answer: Yes, but that expression refers to the posterior expectation of $\lambda$, a Poisson parameter. In this post I'm discussing the interpretation of the second parameter for the Gamma. $\endgroup$ – kjo Aug 9 '17 at 9:01
1
$\begingroup$

You are, more or less, looking at different coordinate systems.

The conjugate prior distribution is for a rate parameter, while the other is the distribution for a time random variable.

$$\bbox[2ex]{\underbrace{\quad \lambda}_{\text{a rate}} \sim \mathcal{Gamma}(n_0,\underbrace{t_0\qquad}_{\text{time interval}})} \\ \text{versus}\\\bbox[2ex]{\underbrace{\qquad\quad X}_{\text{a time interval}} \sim \mathcal{Gamma}(1 ,\underbrace{\nu\quad}_{\text{a rate}})}$$

That is all.

$\endgroup$
1
$\begingroup$

Two conventional ways of parametrizing the family of Gamma distributions are these: \begin{align} & \frac 1 {\Gamma(\alpha)} \left( \frac t \mu \right)^{\alpha-1} e^{-t/\mu} \, \frac{dt} \mu \qquad \text{for } t\ge0 \\[12pt] & \frac 1 {\Gamma(\alpha)} (\lambda t)^{\alpha-1} e^{-\lambda t} \, (\lambda\, dt) \qquad \text{for } t\ge 0 \end{align}

If $\lambda = \dfrac 1 \mu$ then these are both the same distribution, with expected value $\alpha\mu = \dfrac\alpha \lambda.$ If $t$ is measured in units of time, then so is $\mu,$ and $\lambda$ is then in units of frequency (i.e. $\lambda$ is a rate).

$\endgroup$
1
  • $\begingroup$ That's not the issue here. They're using the same distribution (shape-rate parameters); just for different coordinate systems. $\endgroup$ – Graham Kemp Aug 9 '17 at 3:54

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.