The output files produced by ADMIXMAP should be processed by running the R script AdmixmapOutput.R. This produces several text files and graphs (by default in postscript format). The R script is run automatically if you use the Perl script to invoke ADMIXMAP.
The adequacy of the burn-in period can be evaluated by the Geweke diagnostics in the R output. If the burn-in period is adequate, the numbers in this table should have approximately a standard normal distribution.
The mixing of the MCMC sampler can be evaluated by examining the autocorrelation plots. Autocorrelation extending beyond 20 iterations (2 thinned draws if every = 10) indicates slow mixing.
Acceptance rates for the Metropolis-Hastings samplers used by the program are printed to screen and logfile.
The adequacy of the total number of iterations can be evaluated by examining a plot of the statistic of interest calculated from all iterations since the end of the burn-in period, against the iteration number. Where inference is based on the mean of a parameter, this statistic is an ergodic (cumulative) average over all iterations to that point. Plots of ergodic averages of the population-level parameters are given in file ErgodicAveragePlots.ps.
The stratificationtest outputs results of a diagnostic test for residual population stratification that is not explained by the fitted model. For details of how this test is calculated, and a discussion of how to interpret it, see Hoggart (2003). The test is based on testing for allelic association between unlinked loci that is not explained by the model. The result is a "Bayesian p-value": p < 0.5 indicates lack of fit. The "Bayesian p-value" calculated by this test is more conservative than a classical p-value. Our experience has been that a test p-value of 0.3 or less is fairly strong evidence for residual stratification. Where this statistic yields evidence of lack of fit, the model should be specified with more subpopulations, unless there is some other reason for lack of fit such as mis-specified allele frequencies.
The dispersiontest outputs results of a diagnostic test for variation between the allele frequencies in the unadmixed populations that have been sampled to calculate the prior parameter values in priorallelefreqfile and the corresponding ancestry-specific allele frequencies in the admixed population under study. Again the results are "Bayesian p-values", for which the deviation of the test p-value from its expected value of 0.5 does not provide an absolute measure of the strength of evidence for lack of fit. For each subpopulation, the test statistic is calculated as a summary test over all loci and for each locus separately. Examination of the test statistic for each locus may reveal errors in coding, or errors in specifying the prior allele frequencies.
The option dispersiontest is valid only where option priorallelefreqfile has been specified. Where allele frequencies have been specified as fixed, option allelefreqtest should be specified and the output file should be examined.
No diagnostic test for lack of fit of the distribution of individual admixture proportions to the model is yet implemented. However the plot DistributionIndividualAdmixture can be examined to compare the estimated distribution of individual admixture proportions (based on the posterior means for individual admixture) with an estimate for the distribution of individual admixture values in the population (based on the posterior means for the Dirichlet parameters of this distribution).
The deviance and Deviance Information Criterion (DIC) are computed each time.
For an analysis of a single individual, with option chib, the log marginal likelihood, also known as the log evidence, is computed.
With option thermo=1, the marginal likelihood is approximated for any model. The greater the value of numannealedruns, the more accurate will be the approximation, but the longer the program will take to run.