Skip to contents

Computes the log pointwise predictive density and the posterior-variance form of WAIC without posterior simulation. Each row of the response matrix is one multivariate observation.

Usage

compute_analytical_waic(Y, X, Mn, mn, Psi_n, nu_n)

Arguments

Y

Finite numeric response matrix with dimensions n by q.

X

Finite numeric design matrix with dimensions n by p. Include an intercept column explicitly when required.

Mn

Symmetric positive-definite p by p posterior row covariance of the regression coefficients, conditional on the response covariance.

mn

Finite numeric p by q posterior coefficient mean. Supply the actual mean, not its precision-weighted parameterization.

Psi_n

Symmetric positive-definite q by q inverse-Wishart posterior scale matrix.

nu_n

Finite scalar inverse-Wishart posterior degrees of freedom, strictly greater than q - 1.

Value

A list containing the original scalar summaries WAIC, lppd, and p_waic; their difference elpd_waic; and an n by 4 pointwise matrix with columns lppd, p_waic, elpd_waic, and WAIC.

Details

The likelihood and posterior are $$Y_i \mid B, \Sigma \sim N_q(x_i B, \Sigma),$$ independently across rows, and $$B \mid \Sigma, D \sim MN(m_n, M_n, \Sigma), \quad \Sigma \mid D \sim IW(\nu_n, \Psi_n).$$ The inverse-Wishart density is proportional to \(|\Sigma|^{-(\nu_n+q+1)/2} \exp\{-\mathrm{tr}(\Psi_n\Sigma^{-1})/2\}\). In particular, mn is the coefficient mean itself. When using MNIW_R() with H = diag(n), pass Vbnew, Cnew, Snew, and vnew as Mn, mn, Psi_n, and nu_n, respectively.

Let \(h_i=x_i M_n x_i^T\) and \(d_i=(y_i-x_i m_n)\Psi_n^{-1}(y_i-x_i m_n)^T\). The row predictive density is multivariate Student t with \(\nu_n-q+1\) degrees of freedom, mean \(x_i m_n\), and scale \((1+h_i)\Psi_n/(\nu_n-q+1)\). The WAIC penalty for row i is $$p_i = \tfrac14\left\{ \sum_{j=1}^q\psi_1((\nu_n-j+1)/2) + 2q h_i^2 + 4h_i\nu_n d_i + 2\nu_n d_i^2 - 4d_i\right\}.$$ Here \(\psi_1\) is the trigamma function. WAIC is \(-2\sum_i(\mathrm{lppd}_i-p_i)\).

Use the observations that fitted the posterior to compute training-data WAIC. Scoring other observations evaluates the supplied posterior at those observations; it does not refit the model. These functions assume independent likelihood rows and do not directly compute WAIC for correlated spatial GP rows or a full state-space emulator. Compare models using the same response data and the same definition of an observation. Lower WAIC indicates higher estimated predictive accuracy among the compared models.

References

Watanabe, S. (2010). Asymptotic equivalence of Bayes cross validation and widely applicable information criterion in singular learning theory. Journal of Machine Learning Research, 11, 3571–3594.

Author

Sudipto Banerjee (original derivation and development implementation).

Examples

X <- cbind(1, c(-1, 0, 1, 2))
Y <- cbind(c(0.1, 0.4, 1.2, 1.7), c(1.2, 0.8, 0.5, 0.1))
post <- MNIW_R(X, Y, H = diag(nrow(Y)), v = 4, S = diag(2),
               C = matrix(0, 2, 2), Vb = diag(2))
score <- compute_analytical_waic(Y, X, post$Vbnew, post$Cnew,
                                 post$Snew, post$vnew)
score$WAIC
#> [1] 10.1313
score$pointwise
#>            lppd    p_waic  elpd_waic     WAIC
#> [1,] -0.9815339 0.6702952 -1.6518291 3.303658
#> [2,] -0.6549797 0.2292402 -0.8842199 1.768440
#> [3,] -0.7079442 0.2629221 -0.9708664 1.941733
#> [4,] -0.9120960 0.6466405 -1.5587365 3.117473