MTGSAM Utility Programs

There are several utilities now available to simplify some of the tasks necessary in using the Gibbs sampling programs. A brief description of each of the programs follows.

HERITCOR.FOR
This program generates a new sample file, MTGS64, that contains values of the heritabilities (fraction of phenotypic variance for other variances) and correlations. The program reads the output from MTGSPREP to determine the model so that the calculation of phenotypic variance is determined without user intervention. The program is linked with subroutines in the MTGSSUB.FOR, MTGSRSB.FOR, and timing subroutines. A batch file and make file for automated compiling and linking are included.

LAGCOR.FOR
This program calculates the lags of samples written to MTGS61 or MTGS64.

The user has the option of specifying additional burn-in rounds if the value given in MTGSRUN is considered insufficient.

The program requests the gap between lag values calculated and the number of correlations to calculate. For example, if a value of 5 is used to for the lag and 20 correlations are requested, the program would calculate the correlation of samples written 5, 10, 15, ... 100 records apart in the sample file.

It is important to remember that the samples in the sample files may already be thinned depending on the option chosen in MTDFRUN. This thinning rate should be considered when determing any additional burn-in and when deciding on the gap between values calculated.

The user specifies the original frequency of sample writing used in MTGSRUN so that correct lags can be calculated based on the original (unthinned) chain. The lags listed in the output file correspond to the original chain.

For example, assume that a thinning rate of 20 had been used in MTGSRUN (i.e., every 20th sample written to MTGS61). If the user requested that a gap of 5 records be used for calculating a total of 10 correlations, then the program would calculate correlations of samples written 5 apart corresponding to lags of 5, 10, 15, ..., 50 records apart in the THINNED chain; these correlations would correspond to lags of 100, 200, ..., 1000 samples apart in the ORIGINAL chain.

MNSAMP.FOR
This program calculates the mean of the sampled values in MTGS61 or MTGS64. This is especially useful to monitor a Gibbs chain for a run that has not completed. The program allows the user to specify additional burn-in rounds if the value given in MTGSRUN is considered insufficient.

PULLDAT.FOR
This program is used to read sampling data from MTGS61, 62, or 64 and copy that information to a user specified text file. The user specifies the number of variables to copy to the text file (with a shortcut for writing all of the variables). The program allows the user to specify additional burn-in rounds if the value given in MTGSRUN is considered insufficient. The program will write the sample information for a user specified interval between samples read from the sample file.

PULLMAT.FOR
This program is used very similar to PULLDAT.FOR described above, except that the output file is a user specified MATLAB file. The resulting data set can be loaded into MATLAB using the "load filename" command. The filename used must have .mat as the file extension. The matrix will be called gs (in lower case) and the rows will contain the string of values for each parameter specified (rather than columns). This is a result of an attempt to minimize memory usage. The only way to write the data file with columns containing the variable is to read the whole data set in - which will often be impossible. A result of this scheme is that the number of samples in the sample file (MTGS61 or 64) must be known in advance.

GIBBSIT.FOR
This is a FORTRAN program written by Raftery and Lewis. The description comes from the header of the program: This program calculates the number of iterations required in a Marcov chain Monte Carlo (e.g., Gibbs sampling) run. The user has to specify the precision required. The program returns the number of iterations required to estimate the posterior cdf of the q-quantile of the quantity of interest (a function of the parameters) to within +-r with probability s. It also gives the number of "burn-in" iterations required for the conditional distribution given any starting point (of the derived two-state process) to be within epsilon of the actual equilibrium distribution.

This program is not distributed with the utilities program, but is available directly from StatLib at http://lib.stat.cmu.edu/general/gibbsit.

• Get the readme.utl file for information about obtaining the utilities
• Get the MTGSAM utility programs (compressed)
• Get the Gibbsit program from StatLib

Contact

Please direct questions and comments to Curt Van Tassell.

Return to MTGSAM home page