Академический Документы
Профессиональный Документы
Культура Документы
The following diagram summarizes conjugate prior relationships for a number of common
sampling distributions. Arrows point from a sampling distribution to its conjugate prior
distribution. The symbol near the arrow indicates which parameter the prior is unknown.
These relationships depends critically on choice of parameterization, some of which are
uncommon. This page uses the parameterizations that make the relationships simplest to state,
not necessarily the most common parameterizations. See footnotes below. Other diagrams on this
site:
Parameterizations
Let C(n, k) denote the binomial coefficient(n, k).
The geometric distribution has only one parameter, p, and has PMF f(x) = p (1-p)x.
The binomial distribution with parameters n and p has PMF f(x) = C(n, x) px(1-p)n-x.
The negative binomial distribution with parameters r and p has PMF f(x) = C(r + x - 1, x) pr(1p)x.
The Bernoulli distribution has probability of success p.
The beta distribution has PDF f(p) = ( + ) p-1(1-p)-1 / (() ()).
The exponential distribution parameterized in terms of the rate has PDF f(x) = exp(- x).
The gamma distribution parameterized in terms of the rate has PDF f(x) = x-1exp(- x) / ().
The Poisson distribution has one parameter and PMF f(x) = exp(-) x/ x!.
The normal distribution parameterized in terms of precision ( = 1/2) has PDF f(x) = (/2)1/2
exp( -(x - )2/2 ).
The lognormal distribution parameterized in terms of precision has PDF f(x) = (/2)1/2 exp(
-(log(x) - )2/2 ) / x.
Posterior parameters
For each sampling distribution, assume we have data x1, x2, ..., xn.
If the sampling distribution for x is binomial(m, p) with m known, and the prior distribution is
beta(, ), the posterior distribution for p is beta( + xi, + mn - xi). The Bernoulli is the
special case of the binomial with m = 1.
If the sampling distribution for x is negative binomial(r, p) with r known, and the prior
distribution is beta(, ), the posterior distribution for p is beta( + nr, + xi). The geometric
is the special case of the negative binomial with r = 1.
If the sampling distribution for x is gamma(, ) with known, and the prior distribution on is
gamma(0, 0), the posterior distribution for is gamma(0 + n, 0 + xi). The exponential is a
special case of the gamma with = 1.
If the sampling distribution for x is Poisson(), and the prior distribution on is gamma(0, 0),
the posterior on is gamma(0 + xi, 0 + n).
If the sampling distribution for x is normal(, ) with known, and the prior distribution on is
normal(0, 0), the posterior distribution on is normal((0 0 + xi)/(0 + n), 0 + n0).
If the sampling distribution for x is normal(, ) with known, and the prior distribution on is
gamma(, ), the posterior distribution on is gamma( + n/2, (n-1)S2) where S2 is the sample
variance.
If the sampling distribution for x is lognormal(, ) with known, and the prior distribution on
is normal(0, 0), the posterior distribution on is normal((0 0 + xi)/(0 + n), 0 + n0).
If the sampling distribution for x is lognormal(, ) with known, and the prior distribution on
is gamma(, ), the posterior distribution on is gamma( + n/2, (n-1)S2) where S2 is the
sample variance.
http://www.johndcook.com/conjugate_prior_diagram.html
(From D. Zwillinger and S. Kokoska, Standard Probability and Statistics Tables and Formulae,
Chapman & Hall/CRC, Boca Raton, Florida, 2000. With permission.)
The chart above is adapted from the chart originally published by Lawrence Leemis in 1986
(Relationships Among Common Univariate Distributions, American Statistician 40:143-146.)
Leemis published a larger chart in 2008 which is available online.
Parameterizations
Student t: f(x) is proportional to (1 + (x2/))-( + 1)/2 for positive x. The parameter is called the
degrees of freedom.
Weibull: f(x) = (/) x-1 exp(- x/) for positive x. The parameter is the shape and is the
scale.
Double exponential : f(x) = exp(-|x-|/) / 2 for all x. The parameter is the location and
mean; is the scale.
For comparison, see distribution parameterizations in R/S-PLUS and Mathematica.
Relationships
In all statements about two random variables, the random variables are implicitly independent.
Geometric / negative binomial: If each Xi is geometric random variable with probability of
success p then the sum of n Xi's is a negative binomial random variable with parameters n and p.
Negative binomial / geometric: A negative binomial distribution with r = 1 is a geometric
distribution.
Negative binomial / Poisson: If X has a negative binomial random variable with r large, p near
1, and r(1-p) = , then FX FY where Y is a Poisson random variable with mean .
Beta-binomial / discrete uniform: A beta-binomial (n, 1, 1) random variable is a discrete
uniform random variable over the values 0 ... n.
Beta-binomial / binomial: Let X be a beta-binomial random variable with parameters (n, , ).
Let p = /( + ) and suppose + is large. If Y is a binomial(n, p) random variable then FX
FY.
Hypergeometric / binomial: The difference between a hypergeometric distribution and a
binomial distribution is the difference between sampling without replacement and sampling with
replacement. As the population size increases relative to the sample size, the difference becomes
negligible.
Geometric / geometric: If X1 and X2 are geometric random variables with probability of success
p1 and p2 respectively, then min(X1, X2) is a geometric random variable with probability of
success p = p1 + p2 - p1 p2. The relationship is simpler in terms of failure probabilities: q = q1 q2.
Poisson / Poisson: If X1 and X2 are Poisson random variables with means 1 and 2 respectively,
then X1 + X2 is a Poisson random variable with mean 1 + 2.
Binomial / Poisson: If X is a binomial(n, p) random variable and Y is a Poisson(np) distribution
then P(X = n) P(Y = n) if n is large and n$ is small. For more information, see Poisson
approximation to binomial.
Cauchy sum: If X1 is a Cauchy (1, 1) random variable and X2 is a Cauchy (2, 2), then X1 +
X2 is a Cauchy (1 + 2, 1 + 2) random variable.
Student t / Cauchy: A random variable with a t distribution with one degree of freedom is a
Cauchy(0,1) random variable.
Student t / Snedecor F: If X is a t random variable with degree of freedom, then X2 is an
F(1,) random variable.
Snedecor F / Snedecor F: If X is an F(1, 2) random variable then 1/X is an F(2, 1) random
variable.
Exponential / Exponential: If X1 and X2 are exponential random variables with mean 1 and 2
respectively, then min(X1, X2) is an exponential random variable with mean 1 2/(1 + 2).
Exponential / Weibull: If X is an exponential random variable with mean , then X1/ is a
Weibull(, ) random variable.
Weibull / Exponential: If X is a Weibull(1, ) random variable, X is an exponential random
variable with mean .
Exponential / Double exponential: If X and Y are exponential random variables with mean ,
then X-Y is a double exponential random variable with mean 0 and scale
Double exponential / exponential: If X is a double exponential random variable with mean 0
and scale , then |X| is an exponential random variable with mean .
http://www.johndcook.com/distribution_chart.html