An MCMC sampler for loci using precomputed dispersion matrices, various priors, and a pre-selected set of variables. For use with BC1 (backcross) designs and recombinant inbred lines.
swapbc1(varcov, invars, rparm, nreps, ana.obj, locs=NULL,
locs.prior=NULL, tol=1e-10 )A list with components:
A k by k by nreps array of the locations sampled in each iteration.
A vector of length k*nreps with the
posteriors of the models.
A k by k matrix of the regression coefficients.
The call to swapbc1
The k*nreps posterior probabilities of the k-1 gene
models.
The k*nreps marginal posteriors for all k gene
models that could be formed using the current k-1 gene model
A vector with length(locs) elements. At
each step, the posterior associated with each candidate locus is
added to an element of this vector. After all steps are finished,
the result is normalized to sum to one. This turns out to be an
exceedingly stable estimate of the marginal posterior.
A vector with length(locs) elements. At
each step, the product of each posterior times the coefficient
associated with a candidate locus is
added to an element of this vector. After all steps are finished,
the result is normalized by the total marginal posterior. This turns
out to be an exceedingly stable estimate of the marginal (over all
models) posterior mean of the regression coefficients.
The result of make.varcov
Scalar or vector with nrow(varcov$var.x) elements;
the 'ridge' parameters for the independent variables - larger values
imply more shrinkage or a more concentrated prior for the regresion
coefficients.
How many cycles of MCMC to perform
A object produced by make.analysis.obj
Which variables to start in the model. The first of these is
immediately removed, so it is merely a placeholder. The number of
genes in the model is therefore k <- length(invars)
The columns of varcov\$var.x to use. The default uses all of
them.
The prior mass to associate with each variable. Typically, these sum to one, but sometimes they might each be set to one (as in computing lod scores).
Used in forming QR decomposition. Let it be.
Charles C. Berry cberry@ucsd.edu
An MCMC sampler for loci using the object of make.varcov is
executed. This sampler uses the exact posterior probability under the
assumed correctness of the regression model using expected genotypes
given marker values. This amounts to linearizing the likelihood with
respect to the (possibly unknown) locus states. For models where the
loci are fully informative markers this is the true posterior.
The chain is implemented as follows: given a set of regressor
variables to start, one variable is removed, all regressor
variables not in the model are examined to determine the effect of each
on the posterior. One variable is sampled. The process is repeated until
each variable has been removed and a new one sampled in its place
(possibly the same variable that was removed is sampled). And this whole
cycle is repeated nreps times.
Berry C.C. (1998) Computationally Efficient Bayesian QTL Mapping in Experimental Crosses. ASA Proceedings of the Biometrics Section, 164-169.
swapf2