Say I'm trying to set a bayesian prior for a Bernoulli trial of coin flips.

The equation I'm interested in is the $p(x|I)$ from the Bayesian numerator: $$P(x|data)\propto x^{N_H}(1-x)^{N_t}p(x|I)$$ where $I$ is the background information.

NOTE: The Bayes denominator takes the form (for a uniform prior $p(x|I)$): $$\int_0^1{x^{N_H}(1-x)^{N_t}p(x|I)} = \frac{\Gamma(N_H+1)\Gamma(N-N_H-1)}{\Gamma(N+2)}$$ Alternative, if I choose the conjugate prior $p(x|i)=x^{\alpha}(1-x)^\beta$, the form of the Bayes denominator stays the same: $$\int_0^1{x^{N_H}(1-x)^{N_t}p(x|I)} = \frac{\Gamma(N_H+\beta+1)\Gamma(N-N_H+\alpha+1)}{\Gamma(N+\alpha+\beta+2)}$$

Now, say I'm interested in using this conjugate prior with the following data:


I feel like the way I choose the $\alpha$ and $\beta$ parameters for $p(x|I)=x^{\alpha}(1-x)^\beta$ is to calculate the mean and stdev for the given data and then somehow "fit" those values to the prior. But I can't seem to figure out the mechanics of doing so.

Hoping someone can put me on the right track


According to a usual Bayesian analysis, the data ($n=35$ tosses with $x = 15$ Heads and $n-x = 20$ Tails, from a binomial distribution with unknown success probability $\theta$) are expressed in terms of the likelihood $$p(x|\theta) \propto \theta^x(1-\theta)^{n-x} = \theta^{15}(1-\theta)^{20}.$$

The parameters $\alpha$ and $\beta$ of the conjugate prior distribution $p(\theta) \propto \theta^{\alpha -1}(1-\theta)^{\beta = 1}$ are chosen based on prior information. This may come from past experience with coins (some people say actual physical coins seldom have heads probabilities much different from 1/2), faith in the honesty of the person who supplied the coin, and so on. Accordingly, you might pick a relatively non-informative prior: with $\alpha = \beta = 1$ (uniform) or $\alpha = \beta = 1/2$ (Jeffrey's); see section 2 of this Wikipedia article.

By contrast, if you have a vague hunch the coin might be slightly biased in favor of heads, then you might pick an informative beta prior $\mathsf{Beta}(\alpha = 22, \beta = 16),$ which has mean $\mu = 22/38 \approx 0.58$ and puts roughly 95% of its probability in the interval $(0.42, 0.73),$ as confirmed by the computation below in R statistical software.

qbeta(c(.025,.975), 22, 16)
## 0.4209972 0.7290206

Then you would obtain the posterior probability as the product of the prior and likelihood distributions. Using the informative prior above, this would give

$$p(\theta|x) \propto p(\theta) \times p(x|\theta) \propto \theta^{22 - 1}(1-\theta)^{16-1} \times \theta^{15}(1-\theta)^{20} = \theta^{37-1}(1-\theta)^{36-1},$$

where we notice that the right hand expression is the kernel (density without normalization constant) of the distribution $\mathsf{Beta}(\alpha=37,\,\beta=36).$ In that case a 95% Bayesian credible (probability) interval for $\theta$ is $(0.39,.0.62).$

 qbeta(c(.025, .975), 37, 36)
 ## 0.3931000 0.6202429

If you had used the non-informative uniform prior, the posterior would be $\mathsf{Beta}(\alpha=15,\,\beta=20)$ and the 95% Bayesian interval estimate of $\theta$ would be $(0.27, 0.59).$

 qbeta(c(.025, .975), 15, 20)
 ## 0.2718502 0.5930306

Both Bayesian interval estimates include $\theta = 0.5,$ but the informative prior information has 'melded' with the binomial data to give a slightly different result than with the noninformative or 'flat' prior. The interval estimate from the flat prior is almost entirely due to the data. [The frequentist Agresti 95% confidence based just on the data is $(0.28, 0.59)$.]

In practice, there is usually no precise way to turn a pre-experiment 'hunch' into a prior distribution that reflects that hunch. Considerations with beta priors are that the distribution mean is $\mu = \frac{\alpha}{\alpha + \beta},$ there is also a (somewhat messier) formula for $\sigma$ in terms of $\alpha$ and $\beta$, and with software such as R, it is easy to find priors that put a 'reasonable' amount of probability into a 'reasonable' interval. In continuing, sequential investigations, one might use the posterior distribution for one phase of experimentation as the prior distribution for the next.

Reasons for using a conjugate beta prior (to go with the the binomial likelihood) are that the mathematics is simple, it is not necessary to deal with the 'denominator' of Bayes' Theorem, and the posterior distribution is easily recognized.

  • $\begingroup$ Hi BruceET, Thanks for your excellent answer. I guess I was confusing myself more than I needed to. By just choosing $\alpha$ and $\beta$ to be the occurrences of a "heads" or "tails", respectively, you take into account all of the information that the prior data has to offer. This, for some reason, still lacks some formality to it, but I think I understand the logic quite will now (e.g. the distribution will become taller and thinner the more data you add). $\endgroup$ – lstbl Mar 12 '17 at 4:01
  • $\begingroup$ The choice of priors for Bayesian data analysis has been a topic of controversy and discussion. As long as the prior doesn't take such an exaggerated position as to overwhelm the data, the exact choice of the prior is usually not an issue in practice. If there is doubt about which prior to use, one can try several to see how sensitive the posterior is to the choice of the prior. As above, interval estimates from noninformative priors are usually numerically about the same as frequentist confidence intervals (even if philosophical interpretations differ). $\endgroup$ – BruceET Mar 12 '17 at 6:00

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.