Baynsian Statistics?!

**Winzer** · Jun3-09, 03:47 AM

Suppose I have a regular quarter and I had to guess heads or tails. I have a 50% chance of getting heads or tails. After I flip it say I get the result: heads. If it is to be flipped a second time, classically I would say I still have a 50% chance pf getting heads or tails. However, from Bayesian statistics I was told that I should lean more towards tails.
Why?

Apparently Bayesian statistics accounts for subjective probability. Being born and raised on classical stats "subjective" and "probability" should not be together. Could someone give me some clear reasoning on Bayesian stats?

**MathematicalPhysicist** · Jun3-09, 04:27 AM

Are you sure that's what it says?
then how do you calculate this probability that after getting head what are the probabilities to get again heads immediatly afterwards.

**SW VandeCarr** · Jun3-09, 02:43 PM

Originally Posted by Winzer

Suppose I have a regular quarter and I had to guess heads or tails. I have a 50% chance of getting heads or tails. After I flip it say I get the result: heads. If it is to be flipped a second time, classically I would say I still have a 50% chance pf getting heads or tails. However, from Bayesian statistics I was told that I should lean more towards tails.
Why?

No. Bayesian statistics does not say that prior flips of a coin influence the outcome of next coin flip. These are assumed to be independent events under both frequentist and Bayesian inference. There's a lot of misunderstanding about this.

First, Bayes Theorem is a statement about conditional probability, not about what is called Bayesian statistics.

So called Bayesian statistics is really about the concept of statistical likelihood. A likelihood (L) is derived from probabilities but is not itself a probability as it ranges over all positive real numbers whereas probabilities range over the closed interval 0,1. In practice lnL and likelihood ratios are used.

The important difference between frequentist inference and Bayesian inference is that in the former, the distribution is assumed and the probability of the data is estimated under this assumption. In Bayesian inference the likelihood of a distribution is estimated given the data. This means that maximum likelihood estimation (MLE) is robust for any underlying distribution whereas frequentist inference is not.

**mgb_phys** · Jun3-09, 02:52 PM

No. Bayesian statistics does not say that prior flips of a coin influence the outcome of next coin flip. These are assumed to be independent events under both frequentist and Bayesian inference. There's a lot of misunderstanding about this.

An unfair coin flip is often used as an example of Bayesian statistics.

The joke is that after 50 heads a frequentist still believes that the next flip has a 50:50 chance of being tails.
While a Bayesian at least starts to suspect he is dealing with a rigged coin!

**SW VandeCarr** · Jun3-09, 03:07 PM

Originally Posted by mgb_phys

An unfair coin flip is often used as an example of Bayesian statistics.

The joke is that after 50 heads a frequentist still believes that the next flip has a 50:50 chance of being tails.
While a Bayesian at least starts to suspect he is dealing with a rigged coin!

That is a joke. Unfortunately many people believe it. With either type of inference, assumptions need to be made regarding independent events. However if you don't make this assumption, the data is the basis for inference under MLE, not a presumed underlying distribution.

**mgb_phys** · Jun3-09, 03:18 PM

Originally Posted by SW VandeCarr

That is a joke. Unfortunately many people believe it. With either type of inference, assumptions need to be made regarding independent events.

But the whole point of Bayesian is that it's for when you don't know the underlying assumptions - like most of science!

**SW VandeCarr** · Jun3-09, 03:36 PM

Originally Posted by mgb_phys

But the whole point of Bayesian is that it's for when you don't know the underlying assumptions - like most of science!

I edited the post you're responding to. I agree, but is some cases you can get into trouble. In the coin flip example the presumed underlying distribution is a uniform p=0.5 H or T. This is widely accepted. Deviations from this are to be expected. How much deviation is acceptable? There is no particular advantage using Bayesian inference here as the frequentist will notice when the results deviate significantly from expectations. The real advantage of Bayesian inference is where the assumption of a particular underlying distribution is weak.

Edit: What I was particularly objecting to was the suggestion to the OP that if the first toss was heads, the Bayesian would say that the probability slightly favored tails in the second toss. This is simply wrong. I you were to make a Bayesian inference re n coin tosses, the inference would be based on the outcome of n tosses. If you were dumb enough to make an inference based on one toss that came out heads, Bayesian inference would favor heads again.