7
$\begingroup$

I have been working with beta-bernoulli posteriors recently. Is it true that if $X,Y$ are independent rvs with $X \sim Beta(a_1+1,b_1+1)$ and $Y \sim Beta(a_2+1,b_2+1)$ then $\mathbb{P}(X>Y)>0.5$ iff $\frac{a_1}{a_1+b_1}>\frac{a_2}{a_2+b_2}$? $a_1,a_2,b_1,b_2$ are assumed to be whole numbers.

This makes intuitive sense as the estimated bernoulli parameter is greater in $X$ than in $Y$. I have observed this for all parameters within the range $0 \leq a_1,b_1,a_2,b_2\leq 6$. However, it does not seem to immediately follow from the definitions.

I have tried writing $\mathbb{P}(X>Y)=\frac{\int_{0}^{1}\int_{0}^{x}y^{a_2}(1-y)^{b_2}x^{a_1}(1-x)^{b_1}dy dx}{B(a_1+1,b_1+1)B(a_2+1,b_2+1)}$, but the numerator doesn't seem that easy to work with. I don't see a similar question on stackexchange. Can this be proved or can a counter example be produced for this?

Update1: It seems the result is false if $a_1,b_1,a_2,b_2$ are just positive real numbers as initially posed. See Henrys answer below. Can a counterexample be produced for whole numbers?

Update2: Henry has produced a whole number counterexample. This answers the question in the negative conclusively.

Improve this question
$\endgroup$
2

1 Answer 1

Reset to default
8
$\begingroup$

It does not seem to be a correct conjecture.

It seems your condition is that the mode for $X$ is greater than the mode for $Y$. Since in non-symmetric Beta distributions, the mode is not equal to the mean, it should be possible to find a counterexample.

One way is to take a non-symmetric case where $\frac{a_1}{a_1+b_1}=\frac{a_2}{a_2+b_2}$, find which tends to be greater more often, and then adjust the parameters slightly to get a counter-example which works.

So, using R, we might look at

set.seed(2023)
XgreaterY <- function(a1,b1,a2,b2,cases){ 
   mean(rbeta(cases, a1+1, b1+1) > rbeta(cases, a2+1, b2+1))
   }
XgreaterY(3, 1, 30, 10, 10^7)
# 0.390946  
XgreaterY(3, 1, 29, 11, 10^7) 
# 0.4385228

suggesting that $a_1=3, b_1=1, a_2=29, b_2=11$ provides a counterexample, since $\frac{a_1}{a_1+b_1} = 0.75 > 0.725 = \frac{a_2}{a_2+b_2}$ but it seems $\mathbb{P}(X>Y) < 0.44 < 0.5$

Improve this answer
$\endgroup$
5
  • $\begingroup$ Could this hold if $a_1,b_1,a_2,b_2$ are all whole numbers? I had tried it only for whole numbers. $\endgroup$
    – Sushant Vijayan
    Commented Jan 12, 2023 at 14:57
  • $\begingroup$ @SushantVijayan No: try $a_1=3,b_1=1,a_2=29,b_2=11$ or XgreaterY(3, 1, 29, 11, 10^7) to get a probability less than $0.44$ $\endgroup$
    – Henry
    Commented Jan 12, 2023 at 15:02
  • $\begingroup$ Thanks. can you update your answer with this example. I will accept your answer. $\endgroup$
    – Sushant Vijayan
    Commented Jan 12, 2023 at 15:06
  • $\begingroup$ would this be true if say I had used the means instead of the mode? $\endgroup$
    – Sushant Vijayan
    Commented Jan 12, 2023 at 15:13
  • 1
    $\begingroup$ @SushantVijayan I doubt it, and a similar approach should work, perhaps reversing the adjustment. The (mode) example in the answer compares $\text{Beta}(4,2)$ and $\text{Beta}(30,12)$. A (mean) example might compare $\text{Beta}(301,149)$ and $\text{Beta}(4,2)$ $\endgroup$
    – Henry
    Commented Jan 12, 2023 at 15:22

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.