AdMit: Adaptive Mixture of Student-t Distributions

Description

Function which performs the fitting of an adaptive mixture of Student-t distributions to approximate a target density through its kernel function

Usage

AdMit(KERNEL, mu0, Sigma0 = NULL, control = list(), ...)

Arguments

KERNEL

kernel function of the target density on which the adaptive mixture is fitted. This function should be vectorized for speed purposes (i.e., its first argument should be a matrix and its output a vector). Moreover, the function must contain the logical argument log. If log = TRUE, the function returns (natural) logarithm values of the kernel function. NA and NaN values are not allowed. (See *Details* for examples of KERNEL implementation.)

mu0

initial value in the first stage optimization (for the location of the first Student-t component) in the adaptive mixture, or location of the first Student-t component if Sigma0 is not NULL.

Sigma0

scale matrix of the first Student-t component (square, symmetric and positive definite). Default: Sigma0 = NULL, i.e., the scale matrix of the first Student-t component is estimated by the function AdMit.

control

control parameters (see *Details*).

…

further arguments to be passed to KERNEL.

Value

A list with the following components:

CV: vector (of length $H$) of coefficients of variation of the importance sampling weights.

mit: list (of length 4) containing information on the fitted mixture of Student-t distributions, with the following components:

p: vector (of length $H$) of mixing probabilities. mu: matrix (of size $H \times d$) containing the vectors of modes (in row) of the mixture components. Sigma: matrix (of size $H \times d^2$) containing the scale matrices (in row) of the mixture components. df: degrees of freedom parameter of the Student-t components.

where $H (\geq 1)$ is the number of components in the adaptive mixture of Student-t distributions and $d (\geq 1)$ is the dimension of the first argument in KERNEL.

summary: data frame containing information on the optimization procedures. It returns for each component of the adaptive mixture of Student-t distribution: 1. the method used to estimate the mode and the scale matrix of the Student-t component (`USER' if Sigma0 is provided by the user; numerical optimization: `BFGS', `Nelder-Mead'; importance sampling: `IS', with percentage(s) of importance weights used and scaling factor(s)); 2. the time required for this optimization; 3. the method used to estimate the mixing probabilities (`NLMINB', `BFGS', `Nelder-Mead', `NONE'); 4. the time required for this optimization; 5. the coefficient of variation of the importance sampling weights.

Details

The argument KERNEL is the kernel function of the target density, and it should be vectorized for speed purposes.

As a first example, consider the kernel function proposed by Gelman-Meng (1991): $$ k(x_1,x_2) = \exp\left( -\frac{1}{2} [A x_1^2 x_2^2 + x_1^2 + x_2^2 - 2 B x_1 x_2 - 2 C_1 x_1 - 2 C_2 x_2] \right) $$ where commonly used values are $A=1$, $B=0$, $C_1=3$ and $C_2=3$.

A vectorized implementation of this function might be:

    GelmanMeng <- function(x, A = 1, B = 0, C1 = 3, C2 = 3, log = TRUE)
    {
      if (is.vector(x))
        x <- matrix(x, nrow = 1)
      r <- -.5 * (A * x[,1]^2 * x[,2]^2 + x[,1]^2 + x[,2]^2
                - 2 * B * x[,1] * x[,2] - 2 * C1 * x[,1] - 2 * C2 * x[,2])
      if (!log)
        r <- exp(r)
      as.vector(r)
    }

This way, we may supply a point $(x_1,x_2)$ for x and the function will output a single value (i.e., the kernel estimated at this point). But the function is vectorized, in the sense that we may supply a $(N \times 2)$ matrix of values for x, where rows of x are points $(x_1,x_2)$ and the output will be a vector of length $N$, containing the kernel values for these points. Since the AdMit procedure evaluates KERNEL for a large number of points, a vectorized implementation is important. Note also the additional argument log = TRUE which is used for numerical stability.

As a second example, consider the following (simple) econometric model: $$ y_t \sim \, i.i.d. \, N(\mu,\sigma^2) \quad t=1,\ldots,T $$ where $\mu$ is the mean value and $\sigma$ is the standard deviation. Our purpose is to estimate $\theta = (\mu,\sigma)$ within a Bayesian framework, based on a vector $y$ of $T$ observations; the kernel thus consists of the product of the prior and the likelihood function. As previously mentioned, the kernel function should be vectorized, i.e., treat a $(N \times 2)$ matrix of points $\theta$ for which the kernel will be evaluated. Using the common (Jeffreys) prior $p(\theta) = \frac{1}{\sigma}$ for $\sigma > 0$, a vectorized implementation of the kernel function might be:

     KERNEL <- function(theta, y, log = TRUE)
     {
       if (is.vector(theta))
         theta <- matrix(theta, nrow = 1)
## sub function which returns the log-kernel for a given
       ## thetai value (i.e., a given row of theta)
       KERNEL_sub <- function(thetai)
       {
         if (thetai[2] > 0) ## check if sigma>0
	 { ## if yes, compute the log-kernel at thetai
           r <- - log(thetai[2])
	         + sum(dnorm(y, thetai[1], thetai[2], TRUE))
	 }
	 else
	 { ## if no, returns -Infinity
	   r <- -Inf
	 }
	 as.numeric(r)
       }
## 'apply' on the rows of theta (faster than a for loop)
       r <- apply(theta, 1, KERNEL_sub)
       if (!log)
         r <- exp(r)
       as.numeric(r)
     }

Since this kernel function also depends on the vector $y$, it must be passed to KERNEL in the AdMit function. This is achieved via the argument $\ldots$, i.e., AdMit(KERNEL, mu = c(0, 1), y = y).

To gain even more speed, implementation of KERNEL might rely on C or Fortran code using the functions .C and .Fortran. An example is provided in the file AdMitJSS.R in the package's folder.

The argument control is a list that can supply any of the following components:

Ns: number of draws used in the evaluation of the importance sampling weights (integer number not smaller than 100). Default: Ns = 1e5.
Np: number of draws used in the optimization of the mixing probabilities (integer number not smaller than 100 and not larger than Ns). Default: Np = 1e3.
Hmax: maximum number of Student-t components in the adaptive mixture (integer number not smaller than one). Default: Hmax = 10.
df: degrees of freedom parameter of the Student-t components (real number not smaller than one). Default: df = 1.
CVtol: tolerance for the relative change of the coefficient of variation (real number in [0,1]). The adaptive algorithm stops if the new component leads to a relative change in the coefficient of variation that is smaller or equal than CVtol. Default: CVtol = 0.1, i.e., 10%.
weightNC: weight assigned to the new Student-t component of the adaptive mixture as a starting value in the optimization of the mixing probabilities (real number in [0,1]). Default: weightNC = 0.1, i.e., 10%.
trace: tracing information on the adaptive fitting procedure (logical). Default: trace = FALSE, i.e., no tracing information.
IS: should importance sampling be used to estimate the mode and the scale matrix of the Student-t components (logical). Default: IS = FALSE, i.e., use numerical optimization instead.
ISpercent: vector of percentage(s) of largest weights used to estimate the mode and the scale matrix of the Student-t components of the adaptive mixture by importance sampling (real number(s) in [0,1]). Default: ISpercent = c(0.05, 0.15, 0.3), i.e., 5%, 15% and 30%.
ISscale: vector of scaling factor(s) used to rescale the scale matrix of the mixture components (real positive numbers). Default: ISscale = c(1, 0.25, 4).
trace.mu: Tracing information on the progress in the optimization of the mode of the mixture components (non-negative integer number). Higher values may produce more tracing information (see the source code of the function optim for further details). Default: trace.mu = 0, i.e., no tracing information.
maxit.mu: maximum number of iterations used in the optimization of the modes of the mixture components (positive integer). Default: maxit.mu = 500.
reltol.mu: relative convergence tolerance used in the optimization of the modes of the mixture components (real number in [0,1]). Default: reltol.mu = 1e-8.
trace.p, maxit.p, reltol.p: the same as for the arguments above, but for the optimization of the mixing probabilities of the mixture components.

References

Ardia, D., Hoogerheide, L.F., van Dijk, H.K. (2009a). AdMit: Adaptive Mixture of Student-t Distributions. R Journal 1(1), pp.25-30. 10.32614/RJ-2009-003

Ardia, D., Hoogerheide, L.F., van Dijk, H.K. (2009b). Adaptive Mixture of Student-t Distributions as a Flexible Candidate Distribution for Efficient Simulation: The R Package AdMit. Journal of Statistical Software 29(3), pp.1-32. 10.18637/jss.v029.i03

Gelman, A., Meng, X.-L. (1991). A Note on Bivariate Distributions That Are Conditionally Normal. The American Statistician 45(2), pp.125-126.

Hoogerheide, L.F. (2006). Essays on Neural Network Sampling Methods and Instrumental Variables. PhD thesis, Tinbergen Institute, Erasmus University Rotterdam (NL). ISBN: 9051708261. (Book nr. 379 of the Tinbergen Institute Research Series.)

Hoogerheide, L.F., Kaashoek, J.F., van Dijk, H.K. (2007). On the Shape of Posterior Densities and Credible Sets in Instrumental Variable Regression Models with Reduced Rank: An Application of Flexible Sampling Methods using Neural Networks. Journal of Econometrics 139(1), pp.154-180.

Hoogerheide, L.F., van Dijk, H.K. (2008). Possibly Ill-Behaved Posteriors in Econometric Models: On the Connection between Model Structures, Non-elliptical Credible Sets and Neural Network Simulation Techniques. Tinbergen Institute discussion paper 2008-036/4.

Examples

Run this code

# NOT RUN {
<!-- % -->
# }
# NOT RUN {
  ## NB : Low number of draws for speedup. Consider using more draws!
  ## Gelman and Meng (1991) kernel function
  GelmanMeng <- function(x, A = 1, B = 0, C1 = 3, C2 = 3, log = TRUE)
  {
    if (is.vector(x))
      x <- matrix(x, nrow = 1)
    r <- -.5 * (A * x[,1]^2 * x[,2]^2 + x[,1]^2 + x[,2]^2
              - 2 * B * x[,1] * x[,2] - 2 * C1 * x[,1] - 2 * C2 * x[,2])
    if (!log)
      r <- exp(r)
    as.vector(r)
  }

  ## Run AdMit (with default values)
  set.seed(1234)
  outAdMit <- AdMit(GelmanMeng, mu0 = c(0.0, 0.1), control = list(Ns = 1e4))
  print(outAdMit)

  ## Run AdMit (using importance sampling to estimate
  ## the modes and the scale matrices)
  set.seed(1234)
  outAdMit <- AdMit(KERNEL = GelmanMeng, 
                    mu0 = c(0.0, 0.1), 
                    control = list(IS = TRUE, Ns = 1e4))
  print(outAdMit)
# }

Run the code above in your browser using DataLab