# Is p-value similar or equal to its reversed version?

I'm going to check if: $P(T(x)\geq t \mid H_0) = P(H_0\mid T(x)>t)$.

As a sidenote, it's a self-study post. I did my math courses a little while ago, so I do not remember exactly how it goes with probability calculus and kindly ask for help.

Actually I know there is no equation between these two conditional probabilities, so that's why we have bayesian branch of statistics, but I'd like to find out on my own.

Suppose I had $10$ coin flips and got $7$ heads.

Then, according to NHST, I state the null hypothesis that $H_0: \theta = 0.5$, where $\theta$ is a probability of obtaining heads.

The alternative hypothesis is $H_0: \theta > 0.5$ (for ease of calculations it's one-sided test). $T(x)$ is my test statistics, is a ratio of heads in 10 coin flips, I got $7$ heads so $T(x)=\frac{7}{10}$.

The p-value is $P(T(x)\geq \frac{7}{10}\mid H_0) = P(T(x)\geq \frac{7}{10}\mid \theta=0.5) = 0.17$

A one-line code in R is:

pbinom(6,10,0.5, lower.tail = FALSE)


But I also want to check if reverse probability ($P(\theta=0.5\mid T(x)\geq\frac{7}{10})$) is similar or equal to $0.17$.

This time I need to use the Bayes' Theorem and define prior probability for $\theta$.

I need to evaluate $P(\theta=0.5\mid T(x)\geq\frac{7}{10})$. Previously, I tried to break down the condition into summation, but as @Ian pointed in his answer - the conditional probability is not additive over the condition.

Doing this step-by-step I get: $P(\theta=0.5\mid T(X)\geq 0.7)=\frac{P(\theta=0.5, T(x)>0.7)}{P(T(x)>0.7)}=\frac{P(T(x)>0.7\mid\theta=0.5)\cdot P(\theta=0.5)}{P(T(x)>0.7}$

Here, I have to assume some prior distribution for $\theta$, so I chose discrete uniform on set $\{0.1, 0.2, \ldots, 1\}$ (just to have some prior).

The numerator $P(T(x)\geq 0.7, \theta=0.5)= P(T(x)=0.7\mid\theta=0.5)\cdot P(\theta=0.5) + P(T(x)=0.8\mid\theta=0.5)\cdot P(\theta=0.5)+P(T(x)=0.9\mid\theta=0.5)\cdot P(\theta=0.5)+P(T(x)=1.0\mid\theta=0.5)\cdot P(\theta=0.5) = {10\choose 3}\cdot 0.5^7\cdot 0.5^3\cdot \frac{1}{10}+{10\choose 2}\cdot 0.5^8\cdot 0.5^2\cdot \frac{1}{10}+ {10\choose 9}\cdot 0.5^9\cdot 0.5^1\cdot \frac{1}{10}+ {10\choose 10}\cdot 0.5^{10}\cdot 0.5^0\cdot \frac{1}{10}=0.017$

The denominator.
Probably, it's correct to rewrite the denominator: $P(T(x)\geq0.7)=P(T(x)=0.7)+P(T(x)=0.8)+P(T(x)=0.9)+P(T(x)=1.0)$.
Now I have to calculate those terms, I do it by: $P(T(x)=0.7)=\sum_{\theta\in\{0.1,0.2,\ldots 1\}} P(T(x)=0.7, \theta)\cdot f(\theta)= \sum_{\theta\in\{0.1,0.2,\ldots 1\}}{10\choose 3}\cdot\theta^7\cdot(1-\theta)^3\cdot\frac{1}{10} = \frac{1}{10}{10\choose 3}\sum_{\theta\in\{0.1,0.2,\ldots 1\}}\theta^7\cdot(1-\theta)^3$
It's far faster to use R in the calculations.

sum(0.1*choose(10,7)*\theta^7*(1-\theta)^3)


The result is $P(T(x)=0.7)=0.091$ Doing this for $P(T(x)=0.8)$, $P(T(x)=0.9)$, and $P(T(x)=1.0)$, I get (respectively) $0.0906, 0.0828$ and $0.149$. Then these four sums are again add up to get $P(T(x)\geq 0.7)=0.414$

Finally, the result is $P(\theta=0.5\mid T(x)\geq 0.7)= \frac{P(T(x)\geq 0.7|\theta=0.5)\cdot P(\theta=0.5)}{P(T(x)\geq 0.7)} = \frac{0.017}{0.414}=0.0.41\not=0.17$.

This proves that 'reversed' conditional probabilities, p-value and its reverse, are not equal, I'm really proud of the result.. but the thing that bothers me...

[Q] Is this all correct? Especially in terms of probability calculus and all conditional probabilities.

1. I don't think you need all these calculations to see what's going on in this case. Whatever you choose for a discrete prior for $\theta$, the result of $P(\theta=0.5 \mid T \geq 0.7)$ will depend on what that prior actually was. In particular, if $\theta=0.5$ was impossible under your prior for whatever reason (maybe you're paranoid and assume every coin you find is biased), this conditional probability would be zero regardless of what data you got. That said, no, the conditional probability is not additive over the condition, only over the event whose probability is being measured. $P(A \cap B)$ is additive as a function of $B$, but you have to divide by $P(B)$ which breaks the additivity.
2. You're correct. In general you really want to look at the posterior distribution for $\theta$, which is again a continuous distribution, so you would want to measure the density of this posterior distribution at $0.5$.