3
$\begingroup$

Suppose I have 1 heads and 4 tails from 5 coin tosses. To find out the probability of 1 heads and 4 tails in my coin toss experiments, I decided to use Binomial Probability Mass Function for the calculation of the probability on the current observations.

I used Maximum a Posterior with the Beta prior $ (\alpha=5, \beta=5) $, instead of Maximum Likelihood Estimation, to estimate the parameter value of $ \theta $ and I got $ 0.3846 $ from Maximum a Posterior. Now, I have the parameter value of $\theta$ and I want to find out the probability that I would observe 1 heads and 4 tails, which one of the following equations should I plug in the estimate parameter value, $ 0.3846 $:

  1. $ {n}\choose{k} $ $ \Pi^n_{i=1} \theta^{x_i} (1-\theta)^{(1-x_i)} \theta^{(\alpha-1)} (1-\theta)^{(\beta-1)}$

The above equation is considering the prior probability for the calculation of the probability for my experiment. Or should I just plug the value of $ \theta $ into the Binomial Probability Mass Function?

  1. $ {n}\choose{k} $ $ \Pi^n_{i=1} \theta^{x_i} (1-\theta)^{(1-x_i)}$

I know this sounds very naive but I just want to make sure I am thinking correctly.

$\endgroup$
0

1 Answer 1

2
$\begingroup$

You have data $X \sim \mathcal{B}(\theta)$, and you observe $(X_1,\dots,X_n)$, $X_o = \sum X_i$, which means that for $k\in \{0,\dots,n\}$,

$$ \mathbb P(X_o = k \mid \theta) = \binom{n}{k} \theta^k(1-\theta)^{n-k} $$

From a bayesian perspective you also have a prior distribution on $\theta$ which is a Beta distribution $p(\alpha,\beta)$.
Then you get the posterior distribution of $\theta$, \begin{align*} p(\theta \mid X_o) &\propto p(\alpha,\beta) \mathbb P(X_o \mid \theta) \\ &\propto p(\alpha,\beta) \theta^{X_o} (1-\theta)^{n-X_o}, \end{align*} and you can derive the maximum a posteriori,

$$ \hat \theta= \arg\max_{\theta} p( \theta \mid X_o) $$

Now if you want to compute the probability of getting 1 head out of 5 (new and independent) coin tosses, given that the probability of head is $\hat \theta$, you can simply put $\hat \theta$ in the likelihood $\mathbb P(X \mid \theta)$:

$$ \mathbb P(X=1 \mid \hat \theta ) = \binom{5}{1} \hat \theta(1-\hat \theta)^4 $$


But from a Bayesian point of view it may be more relevant to consider all possible values of $\theta$, and not just a point estimate like $\hat \theta$, and rather compute

$$ \mathbb P(X= 1 \mid X_o) = \int_\Theta P(X = 1 \mid \theta) p(\theta \mid X_o)d\theta $$

$\endgroup$
3
  • $\begingroup$ Thank you for the answer and I am trying to understand the last equation but it does not come straight. Do you mean a prior probability distribution by $ p(\theta | X_0) $ and do you mean a likelihood by $ P(X=1 | \theta) $? I don't know If you can stretch the last equation a bit further, I would really appreciate. $\endgroup$ Jan 23, 2020 at 13:21
  • $\begingroup$ The last equation is the sum of $\mathbb P (X \mid \theta)$ over all values of $\theta$ weighted by $p(\theta \mid X_o)$ which is your current knowledge about the distribution of $\theta$. This weigthed sum is called the posterior predictive distribution. See stats.stackexchange.com/…. $\endgroup$
    – periwinkle
    Jan 23, 2020 at 13:25
  • $\begingroup$ Thank you very much!! $\endgroup$ Jan 23, 2020 at 13:29

Your Answer

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

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