1
$\begingroup$

I want to quickly create probability distributions for the two parameters to the gamma distribution (shape k and scale θ) after each sample taken. How would you go about this?

My goal is to estimate the likelihood that a future sample will fall within a given range or, equivalently, what proportion of the population fall between two values. Once I have an estimate for my two parameters, I will need to determine what area of the parameter space meets my criterion, then double integrate over their joint distribution to produce my likelihood. How would you go about this?

If you see a more straightforward or efficient way to achieve this estimate please tell me as well. Keep in mind that two of the goals are minimization of samples taken and efficiency of calculation.

$\endgroup$
1
  • 1
    $\begingroup$ You can reasonably approximate the MLE for a Gamma distribution online, by using the formulas on wikipedia. $\endgroup$
    – Batman
    Commented May 21, 2017 at 21:37

1 Answer 1

1
$\begingroup$

Supposing that efficiency is a serious concern, you can use the method of moments estimators, which can be easily updated. (See the engineering statistics handbook here. See also here.) Periodically, though, you can use the maximum likelihood estimates (when you have computational time available).

Suppose we are working with the PDF $$ f(x\,|\,k,\theta) = \frac{x^{k - 1}}{\Gamma(k)\theta^k}\exp\left( \frac{-x}{\theta} \right) $$ and have IID $x_i\sim \Gamma[k,\theta]$.

The method of moments estimators are then: $$ \hat{\theta} = \frac{1}{\bar{x}n}\sum_i (x_i - \bar{x})^2 \;\;\;\;\&\;\;\;\; \hat{k} = \frac{\bar{x}}{\hat{\theta}} $$ Now to do online updates. At update $j$, let $\bar{x}_j$ be the sample mean and define $\gamma_j$ as the sum of squared differences: $$ \gamma_j = \sum_i (x_i - \bar{x}_j)^2 $$ Suppose we have the estimates for data points $x_1,\ldots,x_{n-1}$ (so that we have $n$ points in total now). Then, we can do the update as: \begin{align} \bar{x}_{n} &= \bar{x}_{n-1} + \frac{x_n - \bar{x}_{n-1}}{n} \\[2mm] \gamma_{n} &= \gamma_{n-1} + (x_n - \bar{x}_{n-1})(x_n - \bar{x}_{n}) \\[2mm] \hat{\theta}_n &= \frac{\gamma_n}{\bar{x}_n\,n} \\[2mm] \hat{k}_n &= \frac{\bar{x}_n}{\hat{\theta}_n} \end{align} You can then integrate $f(x|\hat{k}_n,\hat{\theta}_n)$ to estimate the probability of a sample falling in a certain interval, for example.

$\endgroup$
2
  • $\begingroup$ Thanks so much! I've been working on this for years and still haven't gotten a satisfactory solution. You should update your user profile with a name more memorable than user3658307. $\endgroup$
    – bukzor
    Commented Jul 26, 2017 at 16:46
  • $\begingroup$ @bukzor You're welcome :) I guess I need to think of a better name. Btw, if it is helpful, feel free to use the accept button. :3 $\endgroup$ Commented Jul 26, 2017 at 17:08

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .