Note: This page has been translated by MathWorks. Please click here

To view all translated materals including this page, select Japan from the country navigator on the bottom of this page.

To view all translated materals including this page, select Japan from the country navigator on the bottom of this page.

Bayesian linear regression model with samples from prior or posterior distributions

The Bayesian linear regression
model object `empiricalblm`

contains samples from the prior
distributions of *β* and
*σ*^{2}, which MATLAB^{®} uses to characterize the prior or posterior distributions.

The data likelihood is $$\prod _{t=1}^{T}\varphi \left({y}_{t};{x}_{t}\beta ,{\sigma}^{2}\right)},$$ where
*ϕ*(*y _{t}*;

You can create a Bayesian linear regression model with an empirical prior directly
using `bayeslm`

or `empiricalblm`

.
However, for empirical priors, estimating the posterior distribution requires that the
prior closely resemble the posterior. Hence, empirical models are better suited for
updating posterior distributions estimated using Monte Carlo sampling (e.g.,
semiconjugate and custom prior models) given new data.

`estimate`

For semiconjugate, empirical, or custom prior models, `estimate`

estimates the posterior distribution using Monte Carlo
sampling. That is, `estimate`

characterizes the posterior
distribution by a large number of draws from that distribution.
`estimate`

stores the draws in the `BetaDraws`

and `Sigma2Draws`

properties of the returned Bayesian linear
regression model object. Hence, when you estimate `semiconjugateblm`

, `empiricalblm`

, or `customblm`

model objects, `estimate`

returns an
`empiricalblm`

model object.

If you want to update an estimated posterior distribution using new data, and you
have draws from the posterior distribution of *β* and
*σ*^{2}, then you can create an
empirical model using `empiricalblm`

.

creates a Bayesian linear
regression model object (`PriorMdl`

= empiricalblm(`NumPredictors`

,'`BetaDraws`

',BetaDraws,'`Sigma2Draws`

',Sigma2Draws)`PriorMdl`

)
composed of `NumPredictors`

predictors and an
intercept. The random samples from the prior distributions of
*β* and
*σ*^{2},
`BetaDraws`

and `Sigma2Draws`

,
respectively, characterize the prior distributions.
`PriorMdl`

is a template defining the prior
distributions and dimensionality of *β*.

uses additional options specified by one or more
`PriorMdl`

= empiricalblm(`NumPredictors`

,'`BetaDraws`

',BetaDraws,'`Sigma2Draws`

',Sigma2Draws,`Name,Value`

)`Name,Value`

pair arguments.
`Name`

is a property name, except
`NumPredictors`

, and `Value`

is
the corresponding value. `Name`

must appear inside
single quotes (`''`

). You can specify several
`Name,Value`

pair arguments in any order as
`Name1,Value1,...,NameN,ValueN`

.

`estimate` | Fit parameters of Bayesian linear regression model to data |

`simulate` | Simulate regression coefficients and disturbance variance of Bayesian linear regression model |

`forecast` | Forecast responses of Bayesian linear regression model |

`plot` | Visualize prior and posterior densities of Bayesian linear regression model parameters |

`summarize` | Distribution summary statistics of Bayesian linear regression model |

After implementing sampling importance resampling to sample from the posterior distribution,

`estimate`

,`simulate`

, and`forecast`

compute the*effective sample size*(*ESS*), which is the number of samples required to yield reasonable posterior statistics and inferences. Its formula isIf$$ESS=\frac{1}{{\displaystyle {\sum}_{j}{w}_{j}^{2}}}.$$

*ESS*<`0.01*NumDraws`

, then MATLAB throws a warning. The warning implies that, given the sample from the prior distribution, the sample from the proposal distribution is too small to yield good quality posterior statistics and inferences.If the effective sample size is too small, then:

Increase the sample size of the draws from the prior distributions.

Adjust the prior distribution hyperparameters, and then resample from them.

Specify

`BetaDraws`

and`Sigma2Draws`

as samples from*informative*prior distributions. That is, if the proposal draws come from nearly flat distributions, then the algorithm can be inefficient.

Was this topic helpful?