hapassoc {hapassoc}R Documentation

EM algorithm to fit maximum likelihood estimates of trait associations with SNP haplotypes

Description

This function takes a dataset of haplotypes in which rows for individuals of uncertain phase have been augmented by “pseudo-individuals” who carry the possible multilocus genotypes consistent with the single-locus phenotypes. The EM algorithm is used to find MLE's for trait associations with covariates in generalized linear models.

Usage

hapassoc(form,haplos.list,baseline = "missing" ,family = binomial(),
freq = FALSE, maxit = 50, tol = 0.001, ...)

Arguments

form model equation in usual R format
haplos.list list of haplotype data from pre.hapassoc
baseline optional, haplotype to be used for baseline coding. Default is the most frequent haplotype.
family binomial, poisson, gaussian or freq are supported, default=binomial
freq initial estimates of haplotype frequencies, default values are calculated in pre.hapassoc using standard haplotype-counting (i.e. EM algorithm without adjustment for non-haplotype covariates)
maxit maximum number of iterations of the EM algorithm; default=50
tol convergence tolerance in terms of the maximum difference in parameter estimates between interations; default=0.001
... additional arguments to be passed to the glm function such as starting values for parameter estimates in the risk model

Value

it number of iterations of the EM algorithm
beta estimated regression coefficients
freq estimated haplotype frequencies
fits fitted values of the trait
wts final weights calculated in last iteration of the EM algorithm. These are estimates of the conditional probabilities of each multilocus genotype given the observed single-locus genotypes.
var joint variance-covariance matrix of the estimated regression coefficients and the estimated haplotype frequencies
dispersionML maximum likelihood estimate of dispersion parameter (to get the moment estimate, use summary.hapassoc)
family family of the generalized linear model (e.g. binomial, gaussian, etc.)
response trait value
converged TRUE/FALSE indicator of convergence. If the algorithm fails to converge, only the converged indicator is returned.

Note

When fitting logistic regression models (i.e. family=binomial()), you will see warning messages:

non-integer #successes in a binomial glm! in: eval(expr, envir, enclos)

even when the response variable includes only counts. These warnings result from fitting a weighted logistic regression at each iteration of the EM algorithm and can be safely ignored.

References

Burkett K, McNeney B, Graham J (2004). A note on inference of trait associations with SNP haplotypes and other attributes in generalized linear models. Human Heredity, 57:200-206

See Also

pre.hapassoc,summary.hapassoc,glm,family.

Examples

data(hypoDat)
example.pre.hapassoc<-pre.hapassoc(hypoDat, 3)

example.pre.hapassoc$initFreq # look at initial haplotype frequencies
#      h000       h001       h010       h011       h100       h101       h110 
#0.25179111 0.26050418 0.23606001 0.09164470 0.10133627 0.02636844 0.01081260 
#      h111 
#0.02148268 

names(example.pre.hapassoc$haploDM)
# "h000"   "h001"   "h010"   "h011"   "h100"   "pooled"

# Columns of the matrix haploDM score the number of copies of each haplotype 
# for each pseudo-individual.

# Logistic regression for a multiplicative odds model having as the baseline 
# group homozygotes '001/001' for the most common haplotype

example.regr <- hapassoc(affected ~ attr + h000+ h010 + h011 + h100 + pooled,
                  example.pre.hapassoc, family=binomial())

# Logistic regression with separate effects for 000 homozygotes, 001 homozygotes 
# and 000/001 heterozygotes

example2.regr <- hapassoc(affected ~ attr + I(h000==2) + I(h001==2) +
                   I(h000==1 & h001==1), example.pre.hapassoc, family=binomial())


[Package hapassoc version 0.7 Index]