| 1 |
// --+ options: json=compute +-- |
| 2 |
// https://www3.nd.edu/~rwilliam/stats1/OLS-Stata9.pdf |
| 3 |
|
| 4 |
var income educ jobexp race; |
| 5 |
|
| 6 |
varexo e; |
| 7 |
|
| 8 |
parameters b1, b2, b3, b4; |
| 9 |
|
| 10 |
model; |
| 11 |
|
| 12 |
// The following three lines are required because the variables are declared as endogenous, but the equations |
| 13 |
// have no consequences on the estimation of the last equation in the block. |
| 14 |
educ = educ; |
| 15 |
jobexp = jobexp; |
| 16 |
race = race; |
| 17 |
|
| 18 |
[name='eqols'] |
| 19 |
income = b1 + b2*educ + b3*jobexp + b4*race + e; |
| 20 |
|
| 21 |
end; |
| 22 |
|
| 23 |
// Load data. |
| 24 |
ds = dseries('data.csv'); |
| 25 |
|
| 26 |
// Estimate last equatioon by OLS (ds is updated with the fitted values). |
| 27 |
ds = dyn_ols(ds, {}, {'eqols'}); |
| 28 |
|
| 29 |
/* |
| 30 |
** BAYESIAN ESTIMATION OF 'eqols' |
| 31 |
*/ |
| 32 |
|
| 33 |
// Set the number of estimated parameters. |
| 34 |
n = length(oo_.ols.eqols.beta); |
| 35 |
|
| 36 |
// We consider a kind of non informative prior by assuming that the inverse of the prior variance for β is zero. |
| 37 |
V0 = diag(Inf(n, 1)); |
| 38 |
|
| 39 |
// We choose the prior mean for β randomly (around the OLS estimate). |
| 40 |
beta0 = oo_.ols.eqols.beta + randn(n, 1); |
| 41 |
|
| 42 |
// Set prior for the variance of the error |
| 43 |
s2priormean = 17; |
| 44 |
s2df = 1; |
| 45 |
|
| 46 |
// Run bayesian estimation with normal-gamma prior. Because there is no closed |
| 47 |
// form expression for the joint posterior marginal distribution of β, a Gibbs |
| 48 |
// sampling algorithm is used (the prior for β and the inverse of σ² are independent). |
| 49 |
|
| 50 |
gibbslength = 300000; // Set the number of iterations in Gibbs |
| 51 |
burnin = 10000; // Set the number of iterations to be discarded (try to remove the effects of the initial condition). |
| 52 |
steps = 10; // Do not record all iterations (try to remove the dependence between the draws). |
| 53 |
|
| 54 |
ds = olsgibbs(ds, 'eqols', beta0, V0, s2priormean, s2df, gibbslength, burnin, steps, {'eqols', 'eqols_olsgibbs_fit'}); |
| 55 |
|
| 56 |
// Since we use a diffuse prior for β, the posterior mean of β should be close to the OLS estimate. |
| 57 |
if max(abs(oo_.ols.eqols.beta-oo_.olsgibbs.eqols.posterior.mean.beta))>.1 |
| 58 |
error('Something is wrong in the Gibbs sampling routine (univariate model)') |
| 59 |
end |
| 60 |
|
| 61 |
// We can plot the posterior distribution of the parameters: |
| 62 |
post = oo_.olsgibbs.eqols.posterior.density.b2; |
| 63 |
plot(post(:,1), post(:,2), '-k', 'linewidth', 2); |
| 64 |
axis tight |
| 65 |
box on |
| 66 |
hold on |
| 67 |
plot([oo_.olsgibbs.eqols.posterior.mean.beta(2), oo_.olsgibbs.eqols.posterior.mean.beta(2)], [0 max(post(:,2))], '-r') |
| 68 |
hold off |