The CMLMT Procedure#

First Input Argument: Pointer to Procedure#

The first input argument is the pointer to the procedure computing the log-likelihood function and, optionally, the gradient and/or Hessian. You can create a pointer to a procedure by prepending the name of the procedure with an ampersand (&). For example, if your log-likelihood procedure is named lpr, then your call to cmlmt() would look like this (at least with respect to the first input):

// The & makes the first input a pointer to a procedure
out = cmlmt(&lpr, p);

Second Input Argument: Model Parameters#

CMLMT allows you to pass the model parameters as either a \(Px1\) matrix where \(P\) is the number of parameters in the model, or inside a PV structure containing the model parameters.

PV Parameter Instance#

The GAUSS Run-Time Library contains special functions that work with the PV structure. They are prefixed by “pv” and defined in pv.src. These functions store matrices and arrays with parameters in the structure and retrieve the original matrices and arrays along with various kinds of information about the parameters and parameter vector from it.

The advantage of the PV structure is that it permits you to retrieve the parameters in the form of matrices and/or arrays ready for use in calculating your log-likelihood. The matrices and arrays are defined in your command file when the start values are set up. It isn’t necessary that a matrix or array be completely free parameters to be estimated. There are pvPack functions that take mask arguments defining what is a parameter versus what is a fixed value. There are also functions for handling symmetric matrices where the parameters below the diagonal are duplicated above the diagonal.

For example, a PV structure is created in your command file:

// Declare 'p' to be a PV struct
struct PV p;

// Set PV defaults (required)
p = pvCreate();

// Create 3x1 vector
garch = { .1, .1, .1 };

// Add 3x1 vector, 'garch' to 'p' with name "garch"
p = pvPack(p, garch, "garch");

Fixed Parameters#

A matrix or array in the model may contain a mixture of fixed values along with parameters to be estimated. In most cases, it is simplest to assign some parameters as fixed by using the active member of the cmlmtControl structure.

You can also specify some parameters to be fixed to their start value by ‘packing’ them in the PV struct with pvPackM(), which has an additional argument called a “mask,” strictly conformable to the input matrix or array, indicating which elements are fixed (the corresponding element in the mask is zero) or being estimated (the corresponding element in the mask is nonzero). For example:

// Declare 'p' to be a 'PV' structure
struct PV p;

// Initialize PV struct
p = pvCreate();

// Create parameter matrix
b = { 1.0, 0.0, 0.0;
      0.5, 1.0, 0.2;
      0.3, 0.0, 1.0 };

// Create mask with 1 for each
// free parameter to be estimated
// or a 0 for fixed parameters
b_mask = { 0, 0, 0;
           1, 0, 1;
           1, 0, 1 };

p = pvPackM(p, b, "beta", b_mask);

In this case, there are four free parameters to be estimated, \(b_{21}\), \(b_{23}\), \(b_{31}\), and \(b_{33}\). The parameters \(b_{11}\) and \(b_{22}\) are fixed to 1.0 and \(b_{12}\), \(b_{13}\), and \(b_{32}\) are fixed to 0.0.

Symmetric Matrices#

The pvPackS() procedure “packs” a symmetric matrix into the PV structure in which only the lower left portion of the matrix contains independent parameters while the upper left is duplicated from the lower left. The following packed matrix contains three nonredundant parameters. When this matrix is unpacked, it will contain the upper nonredundant portion of the matrix equal to the lower portion.

// Create symmetric matrix
vc = { 1.2 0.4,
        0.4 2.1 };

// Pack symmetric matrix, using 'pvPackS'
p = pvPackS(p, vc, "phi");

Suppose that you wish to specify a correlation matrix in which only the correlations are free parameters. You would then use pvPackSM().

// Create starting correlation matrix
cor = { 1.0 0.2,
        0.2 1.0 };

// Fix the diagonal elements at their starting value
msk = { 0 1,
        1 0 };

// Pack symmetric matrix
p = pvPackSM(p, cor, "R", msk);

Optional Input Argument: Instance of a cmlmtControl Structure#

The cmlmtControl structure is an optional input. If used, it must be the final argument passed into cmlmt(). The members of the cmlmtControl structure instance set the options for the optimization. For example, suppose you want cmlmt() to stop after 100 iterations:

// Declare 'c0' to be a cmlmtControl structure
struct cmlmtControl c0;

// Fill 'c0' with default values
c0 = cmlmtControlCreate();

// Set the 'maxIters' member to 100
c0.maxIters = 100;

The cmlmtControlCreate() procedure sets all of the defaults. The default values for all the members of a cmlmtControl instance can be found in that procedure located at the top of comtutil.src in the GAUSS src subdirectory.

Optional Extra Input Arguments#

Any data that your objective procedure needs other than the model parameters can be passed in as optional dynamic arguments to cmlmt(). These optional input arguments can be any GAUSS type such as matrices strings, arrays, structures, etc. You will pass these arguments to cmlmt(), between the parameter vector and the control structure. cmlmt() will pass them, untouched, to your log-likelihood procedure.

For a simple example, suppose that you have a least squares problem for which you need to supply the X matrix and y vector.

// Log-likelihood procedure with extra data arguments 'y' and 'X'
proc (1) = myLoglikelihood(b_hat, y, X, ind);
    local res;

    struct modelResults mm;

    if ind[1];
        res = y - X * b_hat;
        mm.function = res'res;
    endif;

    retp(mm);
endp;

X = //code to load or create ‘X’
y = //code to load or create ‘y’

//Starting parameter values
b_start = { 1, 1, 1 };

// Call cmlmt
struct cmlmtResults out;
out = cmlmt(&myLoglikelihood, b_start, y, X);

Since this example does not pass in a control structure, the extra data arguments, y and X are the final inputs to cmlmt().

On this page