Astrophysics (Index) | About |

**Bayesian statistics** is a statistical approach that deals with
**Bayesian probabilities**, probabilities representing your degree
of belief that something is fact.
This approach to statistics is not new,
but through much of the 20th century was not the typical method in use.
One reason for the revival is the world's
ever-growing number-crunching capacity,
as Bayesian methods can require a lot of computation.
With its renaissance, the term **frequentist statistics**
has been used to refer to the more common 20th-century approach.

An example of the *frequentist* approach to find out how many of
a city's voters plan to vote for *candidate A* by choosing a random
sample of voters and finding out their plans, then adopting the result
as an estimate of the entire city's voter population's plans, along
with an indication of your confidence based upon your sample size.
This confidence indication typically indicates how likely the actual
number is to be within some interval encompassing your calculated
estimate, e.g., "99% chance that it is within a percent of the actual
number."

The Bayesian approach applies the probability rule,
**Bayes' theorem** (aka **Bayes' law** or **Bayes' rule**)
to a Bayesian probability.
Bayes' theorem is a general mathematical probability theorem
not confined to Bayesian statistics, valid in any statistics method,
but in the case of Bayesian statistics,
it is exactly what is needed to
update a Bayesian probability with information provided by sample data.

Using the Bayesian approach, if you have an idea of how many plan
to vote for *candidate A*, and check with some voters, the theorem
can be used to factor the new evidence into your opinion.
For example, if you earlier believed there is a 90% chance *candidate A*
would receive a majority and polling three random people reveals all
three will vote for *candidate B*,
this naturally affects your opinion, and the theorem provides the means
to calculate an exact, rational adjustment.
*Three* is a very small sample, yet applying Bayes' theorem
to the Bayesian probability yields a sound result.
This adjustment-by-calculation is termed **Bayesian inference**.
In many straight forward cases with sufficient data, this can yield
a conclusion virtually matching that of the frequentist approach,
but it can also be applied in situations where the frequentist
method cannot, and it produces appropriate conclusions even when
based on very small amounts of data. It also has the benefit of
drawing correct conclusions from some kinds of disparate evidence,
mixing apples and oranges, so to speak.
Like any statistics method, its conclusions
depend upon careful regard for what aspects of the data are
truly independent.

Some reluctance to use the Bayesian approach stems from its unfamiliarity (statistics textbooks for many years didn't even mention it), the non-intuitive conclusions the theorem at times produces (though correct ones, assuming correct interpretation of the input data), the fact that it begins with an opinion (the initial Bayesian probability), which clearly affects the result and would seem to preclude objectivity in the conclusions, and that the calculations of "probabilities of probabilities" (either of which could be a sizable probability distribution) often lead to math equations that cannot be solved analytically, and very often not even through the usual numerical methods. Its revival, in addition to stemming from the world's increasing raw computational capacity, is due both to the additional conclusions that Bayesian statistics can draw (e.g., given the evidence, exactly how likely some particular distribution is to be correct), and the availability of new computational techniques that can handle the difficult math problems, in particular, Markov chain Monte Carlo (MCMC).

Bayesian statistics has its own jargon: the Bayesian probabilities
of a given application are spoken of as **prior** or **posterior**,
for that which represents your opinion *before* your "study", and
that which represents your opinion *after* factoring in the study's
data. Thus, phrases like:

**prior probability**,**prior distribution**,**prior mass**,**prior density****posterior probability**,**posterior distribution**,**posterior density**,**posterior mass**

correspond to
the probability,
the probability distribution,
the probability mass function,
and/or the probability density function
that are the input to and output of the calculation (the Bayesian inference).
(**Prior** and **posterior** may be seen "alone", referring to a
*distribution* or a probability *mass*, according to the context.)
A **Bayes factor** is the relative probability of two competing
models (perhaps one being the null hypothesis). If the models
produce distributions, it is the ratio of their respective integrals.

When embarking upon analysis, the necessity for a *prior* presents
a challenge: you need to figure out what to use, which can influence
the result. You may have to devise the prior, perhaps
characterizing the answer to the question "what do we know now".
I'm guessing priors are often the results of a preliminary or past
frequentist study. To establish that a Bayesian result is robust,
often multiple priors are tried specifically to see if the result
is dependent upon the selection from among reasonable priors.

http://en.wikipedia.org/wiki/Bayesian_statistics

http://www.scholarpedia.org/article/Bayesian_statistics

http://faculty.washington.edu/kenrice/BayesIntroClassEpi2018.pdf

Approximate Bayesian Computation (ABC)

Bernstein polynomial

Markov chain Monte Carlo (MCMC)

pulsar timing array (PTA)

templates