Bayesian estimation—thousands of built-in models, by combining
Over 60 likelihood models, including univariate and multivariate normal, logit, probit, ordered logit, ordered probit, Poisson ...
Many prior distributions, including normal, lognormal, multivariate normal, gamma, beta, Wishart ...
Continuous, binary, ordinal, count, and survival outcomes
Univariate, multivariate, and multiple-equation models
Linear models, nonlinear models, and generalized nonlinear models
Continuous univariate, multivariate, and discrete priors
Simply prefix your estimation command with bayes: to fit Bayesian regressions
Fit general Bayesian models using bayesmh
Save your MCMC and estimation results for future use
MCMC methods
Adaptive Metropolis–Hastings (MH)
Adaptive MH with Gibbs updates—hybrid
Full Gibbs sampling for some models
Graphical tools to check MCMC convergence
Posterior summaries: means, medians, SDs, MCSEs, CrIs
Hypothesis testing: interval, model posterior probabilities
Model comparison: DIC, Bayes factors
Bayesian estimation
Continuous, binary, ordinal, and count outcomes
Univariate, multivariate, and multiple-equation models
Linear models, nonlinear models, and generalized nonlinear models
Over 60 likelihood models including univariate and multivariate normal, logit, probit, ordered logit, ordered probit, Poisson, ...
Continuous univariate, multivariate, and discrete priors
Many prior distributions including normal, lognormal, multivariate normal, gamma, beta, Wishart, ...
Specialized priors such as flat, Jeffreys, and Zellner's g
Simply prefix your estimation command with bayes: to fit Bayesian regressions
Fit general Bayesian models using bayesmh
Or write your own programs to calculate likelihood function and choose built-in priors
Or write your own programs to calculate posterior density directly
Whether you are using one of the built-in models or writing your own, all diagnostics and inference features are available
Save your MCMC and estimation results for future use
MCMC methods
Adaptive Metropolis-Hastings (MH)
Hybrid Metropolis-Hastings (adaptive MH with Gibbs updates)
Full Gibbs sampling for certain likelihood and prior combinations
Graphical tools to check MCMC convergence visually
Explore MCMC efficiency by computing effective sample sizes, autocorrelation times, and efficiencies
Bayesian summaries
Posterior means, medians, and SDs
Monte Carlo standard errors (MCSEs)
Credible Intervals (CrIs)
Compute any of above for parameters or functions of parameters
Hypothesis testing
Interval-hypothesis testing for parameters or functions of parameters
Model-based hypothesis testing by computing model posterior probabilities
Model comparison
Bayesian information criteria such as deviance information criterion
Bayes factors
;)
You may be interested in Bayesian analysis if
you have some prior information available from previous studies that you would like to incorporate in your analysis. For example, in a study of preterm birthweights, it would be sensible to incorporate the prior information that the probability of a mean birthweight above 15 pounds is negligible.
your research problem may require you to answer a question: What is the probability that my parameter of interest belongs to a specific range? For example, what is the probability that an odds ratio is between 0.2 and 0.5?
you want to assign a probability to your research hypothesis. For example, what is the probability that a person accused of a crime is guilty?
And more.
Overview of Bayesian analysis.
Stata provides a suite of features for performing Bayesian analysis. The main estimation commands are bayes: and bayesmh. The bayes prefix is a convenient command for fitting Bayesian regression models—simply prefix your estimation command with bayes:. The bayesmh command fits general Bayesian models—you can choose from a variety of built-in models or program your own. The main simulation method is an adaptive Metropolis–Hastings (MH) Markov chain Monte Carlo (MCMC) method. Gibbs sampling is also supported for selected likelihood and prior combinations. Commands for checking convergence and efficiency of MCMC, for obtaining posterior summaries for parameters and functions of parameters, for hypothesis testing, for model comparison, for prediction, and for model checks are also provided.
Your Bayesian analysis can be as simple or as complicated as your research problem. Here's an overview.
Suppose we want to estimate the mean car mileage mpg. Our standard frequentist analysis may fit a regression model to mpg and look at the constant _cons.
. sysuse auto (1978 automobile data) . regress mpg
| Source | SS df MS | Number of obs = 74 | |
| F(0, 73) = 0.00 | |||
| Model | 0 0 . | Prob > F = . | |
| Residual | 2443.45946 73 33.4720474 | R-squared = 0.0000 | |
| Adj R-squared = 0.0000 | |||
| Total | 2443.45946 73 33.4720474 | Root MSE = 5.7855 |
| mpg | Coefficient Std. err. t P>|t| [95% conf. interval] | |
| _cons | 21.2973 .6725511 31.67 0.000 19.9569 22.63769 | |
The simplest way to fit the corresponding Bayesian regression in Stata is to simply prefix the above regress command with bayes:.
. bayes: regress mpg
For teaching purposes, we will first discuss the bayesmh command for fitting general Bayesian models. We will return to the bayes prefix later.
To fit a Bayesian model, in addition to specifying a distribution or a likelihood model for the outcome of interest, we must also specify prior distributions for all model parameters.
For simplicity, let's model mpg using a normal distribution with a known variance of, say, 35 and use a noninformative flat prior (with a density of 1) for the mean parameter {mpg:_cons}.
. bayesmh mpg, likelihood(normal(35)) prior({mpg:_cons}, flat) rseed(12345)
Burn-in ...
Simulation ...
Model summary
| Likelihood: mpg ~ normal({mpg:_cons},35) Prior: {mpg:_cons} ~ 1 (flat) |
| Equal-tailed | ||
| mpg | Mean Std. dev. MCSE Median [95% cred. interval] | |
| _cons | 21.30713 .6933058 .015564 21.32559 19.91967 22.64948 | |
bayesmh discarded the first 2,500 burn-in iterations and used the subsequent 10,000 MCMC iterations to produce the results. The estimated posterior mean, the mean of the posterior distribution, of parameter {mpg:_cons} is close to the OLS estimate obtained earlier, as is expected with the noninformative prior. The estimated posterior standard deviation is close to the standard error of the OLS estimate.
The MCSE of the posterior mean estimate is 0.016. The MCSE is about the accuracy of our simulation results. We would like it to be zero, but that would take an infinite number of MCMC iterations. We used 10,000 iterations and have results accurate to about 1 decimal place. That's good enough, but if we wanted more accuracy, we could increase the MCMC sample size.
According to the credible interval, the probability that the mean of mpg is between 19.92 and 22.65 is about 0.95. Although the confidence interval reported in our earlier regression has similar values, it does not have the same probabilistic interpretation.
Because bayesmh uses MCMC, a simulation-based method, the results will be different every time we run the command. (Inferential conclusions should stay the same provided MCMC converged.) You may want to specify a random-number seed for reproducibility in your analysis.
The interpretation of our results is valid only if MCMC converged. Let's explore convergence visually.
. bayesgraph diagnostics {mpg:_cons}, histopts(normal)
The trace plot of {mpg:_cons} demonstrates good mixing. The autocorrelation dies off quickly. The posterior distribution of {mpg:_cons} resembles the normal distribution, as is expected for the specified likelihood and prior distributions. We have no reason to suspect nonconvergence.
Also see Multiple chains and Gelman-Rubin convergence diagnostic for investigating MCMC convergence using multiple chains.
We can now proceed with further analysis.
We can test an interval hypothesis that the mean mileage is greater than 21.
. bayestest interval {mpg:_cons}, lower(21)
Interval tests MCMC sample size = 10,000
prob1 : {mpg:_cons} > 21
| Mean Std. dev. MCSE | ||
| prob1 | .6735 0.46896 .0099939 | |
The estimated probability of this interval hypothesis is 0.67. This is in contrast with the classical hypothesis testing that provides a deterministic decision of whether to reject the null hypothesis that the mean is greater than 21 based on some prespecified level of significance. Frequentist hypothesis testing does not assign probabilistic statements to the tested hypotheses.
Suppose that based on previous studies, we have prior information that the mean mileage is normally distributed with mean 30 and variance 5. We can easily incorporate this prior information in our Bayesian model. We will also store our MCMC and estimation results for future comparison.
. bayesmh mpg, likelihood(normal(35)) prior({mpg:_cons}, normal(30,5))
saving(prior1_sim) rseed(12345)
Burn-in ...
Simulation ...
Model summary
| Likelihood: mpg ~ normal({mpg:_cons},35) Prior: {mpg:_cons} ~ normal(30,5) |
| Equal-tailed | ||
| mpg | Mean Std. dev. MCSE Median [95% cred. interval] | |
| _cons | 22.0617 .6683529 .014619 22.05628 20.75121 23.39481 | |
This prior resulted in a slight increase of the posterior mean estimate—the prior shifted the estimate toward the specified prior mean of 30.
Suppose that another competing prior is that the mean mileage is normally distributed with mean 20 and variance 4.
. bayesmh mpg, likelihood(normal(35)) prior({mpg:_cons}, normal(20,4))
saving(prior2_sim) rseed(12345)
Burn-in ...
Simulation ...
Model summary
| Likelihood: mpg ~ normal({mpg:_cons},35) Prior: {mpg:_cons} ~ normal(20,4) |
| Equal-tailed | ||
| mpg | Mean Std. dev. MCSE Median [95% cred. interval] | |
| _cons | 21.17991 .6658923 .014733 21.17731 19.87617 22.47459 | |
The results using this prior are more similar to the earlier results with the noninformative prior.
We can compare our two models that used different informative priors. Estimation results of the models were stored under prior1 and prior2. To compare the models, we type
. bayesstats ic prior1 prior2 Bayesian information criteria
| DIC log(ML) log(BF) | ||
| prior1 | 472.0359 -242.5827 . | |
| prior2 | 470.8157 -235.7438 6.838942 | |
The second model has a lower DIC value and is thus preferable.
Bayes factors—log(BF)—are discussed in [BAYES] bayesstats ic. All we will say here is that the value of 6.84 provides very strong evidence in favor of our second model, prior2.
We can also compute posterior probabilities for each model.
. bayestest model prior1 prior2 Bayesian model tests
| log(ML) P(M) P(M|y) | ||
| prior1 | -242.5827 0.5000 0.0011 | |
| prior2 | -235.7438 0.5000 0.9989 | |
The posterior probability of the first model is very low compared with that of the second model. In fact, the posterior probability of the first model is near 0, whereas the posterior probability of the second model is near 1.
Continuing our car-mileage example, we now relax the assumption of a known variance of the normal distribution and model it as a parameter {var}. We specify a noninformative Jeffreys prior for the variance parameter.
. bayesmh mpg, likelihood(normal({var})) prior({mpg:_cons}, flat)
prior({var}, jeffreys) rseed(12345)
Burn-in ...
Simulation ...
Model summary
| Likelihood: mpg ~ normal({mpg:_cons},{var}) Priors: {mpg:_cons} ~ 1 (flat) {var} ~ jeffreys |
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| mpg | ||
| _cons | 21.30678 .7018585 .023602 21.30086 19.91435 22.72222 | |
| var | 34.38441 5.787753 .149506 33.73722 24.71946 47.7112 | |
Note that the MCSE for parameter {mpg:_cons} is larger in this model than it was in the model with a fixed variance. As the number of model parameters increases, the efficiency of the MH algorithm decreases, and the task of constructing an efficient algorithm becomes more and more important. In the above, for example, we could have improved the efficiency of MH by specifying the variance parameter in a separate block, block({var}), to be sampled independently of the mean parameter.
Even without adding the blocking, convergence diagnostics for both mean and variance look good.
. bayesgraph diagnostics _all
;)
;)
We can compute summaries for linear and nonlinear expressions of our parameters. Let's compute summaries for a standardized mean, which is a function of both the mean parameter and the variance parameter.
. bayesstats summary (mean_std: {mpg:_cons}/sqrt({var}))
Posterior summary statistics MCMC sample size = 10,000
mean_std : {mpg:_cons}/sqrt({var})
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| mean_std | 3.670299 .3183261 .008546 3.661119 3.060899 4.308195 | |
See Linear regression for how to fit linear regression models using the bayes prefix. Continuing with bayesmh, the command makes it easy to include explanatory variables in our Bayesian models. The syntax for regressions looks just as it does in other Stata estimation commands. For example, we can include an indicator of whether the car is foreign or domestic when modeling the mean car mileage.
. bayesmh mpg foreign, likelihood(normal({var}))
prior({mpg:_cons foreign}, flat) prior({var}, jeffreys) rseed(12345)
Burn-in ...
Simulation ...
Model summary
| Likelihood: mpg ~ normal(xb_mpg,{var}) Priors: {mpg:_cons foreign} ~ 1 (flat) (1) {var} ~ jeffreys |
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| mpg | ||
| foreign | 4.987455 1.43297 .054471 5.014443 2.034135 7.775843 | |
| _cons | 19.81477 .7796195 .030952 19.79542 18.27116 21.35802 | |
| var | 29.52163 5.304377 .194809 28.82301 20.82704 41.50129 | |
We specified a flat prior for both the constant and the coefficient of foreign.
As we mentioned earlier, the easiest way to fit Bayesian regression models in Stata is by using the bayes prefix. For example, we can fit the above regression model simply by typing
. bayes: regress mpg foreign Burn-in ... Simulation ... Model summary
| Likelihood: mpg ~ regress(xb_mpg,{sigma2}) Priors: {mpg:foreign _cons} ~ normal(0,10000) (1) {sigma2} ~ igamma(.01,.01) |
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| mpg | ||
| foreign | 4.955081 1.4276 .037613 4.968955 2.098257 7.709329 | |
| _cons | 19.82446 .7541294 .02255 19.79524 18.34954 21.33642 | |
| sigma2 | 29.49413 5.083196 .115685 29.04474 20.96174 40.56927 | |
The main difference between the bayes prefix and the bayesmh command is that bayes: builds all model parameters automatically and assigns default priors for them. Depending on a regression model, bayes: may also use different sampling settings than bayesmh, such as blocking of model parameters to improve the efficiency of the sampling algorithm.
In the above, bayes: used the default normal priors with 0 mean and variance of 10,000 for the regression coefficients and the default inverse-gamma prior with scale and shape parameters of 0.01 for the error variance. Also, because the regression coefficients and the error variance are a priori independent, bayes: samples them separately in two different blocks. The following bayesmh's specification produces identical results, provided that the same random-number seed is specified.
. bayesmh mpg foreign, likelihood(normal({sigma2}))
prior({mpg:foreign _cons}, normal(0,10000))
prior({sigma2}, igamma(0.01,0.01))
block({mpg:foreign _cons}) block({sigma2})
We can fit a multivariate normal regression to model two size characteristics of automobiles—trunk space, trunk, and turn circle, turn—as a function of where the car is manufactured, foreign, foreign or domestic. The syntax for the regression part of the model is just like the syntax for Stata's mvreg (multivariate regression) command.
We model the covariance matrix of trunk and turn as the matrix parameter {Sigma,matrix}. We specify noninformative normal priors with large variances for all regression coefficients and use Jeffreys prior for the covariance. The MH algorithm has very low efficiencies for sampling covariance matrices, so we use Gibbs sampling instead. The regression coefficients are sampled by using the MH method.
. bayesmh trunk turn = foreign, likelihood(mvnormal({Sigma, matrix}))
prior({trunk:} {turn:}, normal(0,1000))
prior({Sigma, matrix}, jeffreys(2))
block({Sigma, matrix}, gibbs) rseed(12345)
Burn-in ...
Simulation ...
Model summary
| Likelihood: trunk turn ~ mvnormal(2,xb_trunk,xb_turn,{Sigma,m}) Priors: {trunk:foreign _cons} ~ normal(0,1000) (1) {turn:foreign _cons} ~ normal(0,1000) (2) {Sigma,m} ~ jeffreys(2) |
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| trunk | ||
| foreign | -3.348432 1.056294 .046491 -3.337957 -5.417821 -1.375893 | |
| _cons | 14.75301 .5450302 .019877 14.73202 13.73661 15.89116 | |
| turn | ||
| foreign | -6.004471 .8822641 .038093 -5.991697 -7.751608 -4.267331 | |
| _cons | 41.42375 .4817673 .017979 41.42091 40.48585 42.42505 | |
| Sigma_1_1 | 17.11048 3.028132 .036471 16.77325 12.21387 23.93075 | |
| Sigma_2_1 | 7.583515 2.026102 .024179 7.39855 4.153189 12.07798 | |
| Sigma_2_2 | 12.537 2.175963 .024705 12.29787 9.014795 17.45602 | |
Watch how to fit this model using bayesmh's GUI.
As with linear regression, we can more easily use the bayes prefix to fit a Bayesian multivariate linear regression by simply prefixing the corresponding mvreg command with bayes:.
. bayes: mvreg trunk turn = foreign Burn-in ... Simulation ... Model summary
| Likelihood: trunk turn ~ mvnormal(2,xb_trunk,xb_turn,{Sigma,m}) Priors: {trunk:foreign _cons} ~ normal(0,10000) (1) {turn:foreign _cons} ~ normal(0,10000) (2) {Sigma,m} ~ jeffreys(2) |
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| trunk | ||
| foreign | -3.346145 1.060465 .04055 -3.335271 -5.492078 -1.385659 | |
| _cons | 14.72701 .5892327 .023431 14.70941 13.56203 15.90882 | |
| turn | ||
| foreign | -6.038268 .8997933 .035909 -6.031058 -7.830854 -4.371916 | |
| _cons | 41.43276 .5044536 .01985 41.41877 40.41507 42.47317 | |
| Sigma_1_1 | 16.79679 2.895355 .134145 16.44924 12.1547 23.59289 | |
| Sigma_2_1 | 7.385689 1.904771 .083678 7.1838 4.208478 11.61884 | |
| Sigma_2_2 | 12.1799 2.069804 .101891 11.97961 8.749997 17.08906 | |
As an example of a nonlinear model, we consider a change-point analysis of the British coal-mining disaster dataset for the period of 1851 to 1962. This example is adapted from Carlin, Gelfand, and Smith (1992). In these data, the count variable records the number of disasters involving 10 or more deaths.
The graph below suggests a fairly abrupt decrease in the rate of disasters around the 1887–1895 period.
;)
Let's estimate the date when the rate of disasters changed.
We will fit the model
count ~ Poisson(mu1), if year < cp
count ~ Poisson(mu2), if year >= cp
cp—the change point—is the main parameter of interest.
We will use noninformative priors for the parameters: flat priors for the means and a uniform on [1851,1962] for the change point.
We will model the mean of the Poisson distribution as a mixture of mu1 and mu2.
. webuse coal
(British coal-mining disaster data, 1851-1962)
. bayesmh count, likelihood(dpoisson({mu1}*sign(year<{cp})+{mu2}*sign(year>={cp})))
prior({mu1 mu2}, flat)
prior({cp}, uniform(1851,1962))
initial({mu1 mu2} 1 {cp} 1906) rseed(12345)
Burn-in ...
Simulation ...
Model summary
| Likelihood: count ~ poisson({mu1}*sign(year<{cp})+{mu2}*sign(year>={cp})) Priors: {mu1 mu2} ~ 1 (flat) {cp} ~ uniform(1851,1962) |
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| cp | 1890.269 2.382976 .079318 1890.505 1886.102 1896.136 | |
| mu1 | 3.135416 .2912121 .011915 3.120647 2.602249 3.728935 | |
| mu2 | .9376916 .1150646 .00399 .9352093 .721913 1.17637 | |
The change point is estimated to have occurred in 1890 with the corresponding 95% CrI of [1886,1896].
We may also be interested in estimating the ratio between the two means.
. bayesstats summary (ratio: {mu1}/{mu2})
Posterior summary statistics MCMC sample size = 10,000
ratio : {mu1}/{mu2}
| Equal-tailed | ||
| Mean Std. dev. MCSE Median [95% cred. interval] | ||
| ratio | 3.39185 .5080558 .018972 3.357478 2.544546 4.476566 | |
After 1890, the mean number of disasters decreased by a factor of about 3.4 with a 95% credible range of [2.5, 4.5].
The interpretation of our change-point results is valid only if MCMC converged. We can explore convergence visually.
. bayesgraph diagnostics {cp} (ratio: {mu1}/{mu2})
;)
;)
The graphical diagnostics for {cp} and the ratio look reasonable. The marginal posterior distribution of the change point has the main peak at about 1890 and two smaller bumps around the years 1886 and 1896, which correspond to local peaks in the number of disasters.
In Multivariate linear
regression, we showed you how to use the command line to fit a
Bayesian multivariate regression. Watch
Graphical user interface for Bayesian analysis to see
how to fit this model and more using the GUI. The video demonstrates
bayesmh's GUI for fitting the model. Watch A prefix for fitting Bayesian regressions
for a brief overview of the GUI for the bayes prefix.
Carlin, B. P., A. E. Gelfand, and A. F. M. Smith. 1992. Hierarchical Bayesian analysis of changepoint problems. Journal of the Royal Statistical Society, Series C 41: 389–405.
Stata's Bayesian analysis features are documented in their own manual. You can read more about Bayesian analysis, more about Stata's Bayesian features, and see many worked examples in Bayesian Analysis Reference Manual.
Also see Bayesian model averaging.
The Stata Blog: Bayesian binary item response theory models using bayesmh
The Stata Blog: Bayesian modeling: Beyond Stata's built-in models
The Stata Blog: Bayesian logistic regression with Cauchy priors using the bayes prefix
Also see Bayesian lasso and Running multiple Markov chains in parallel.
Watch Bayesian analysis in Stata
Watch Introduction to Bayesian analysis, part 1: The basic concepts
Watch Introduction to Bayesian analysis, part 2: MCMC and the Metropolis-Hastings algorithm
Watch bayes: prefix for fitting Bayesian regressions
Watch Bayesian linear regression using the bayes prefix
Watch Bayesian linear regression using the bayes prefix: How to specify custom priors
