FULLmLL {nem} | R Documentation |
The function the full marginal likelihood of a phenotypic hierarchy.
The full marginal likelihood equals the marginal likelihood mLL
averaged over the error probabilities $α$ and $β$.
FULLmLL(Phi, D1, D0, a0, b0, a1, b1, Pe, Pm=NULL, lambda=0)
Phi |
an adjacency matrix with unit main diagonal |
D1 |
count matrix: phenotypes x genes. How often did we see an effect after interventions? |
D0 |
count matrix: phenotypes x genes. How often did we NOT see an effect after intervention? |
a0, b0, a1, b1 |
Hyperparameters |
Pe |
prior of effect positions in the hierarchy. A matrix of size phenotypes x genes, where each row contains positive numbers summing to 1. |
Pm |
prior on model graph (n x n matrix) with entries 0 <= priorPhi[i,j] <= 1 describing the probability of an edge between gene i and gene j. |
lambda |
regularization parameter to incorporate prior assumptions. |
Additionally to the marginal likelihood introduced in Markowetz et al (2005),
we can average over the error probabilities $α$ and $β$ assuming Beta priors.
The parameters of the two Beta priors are hyperparameters of the full marginal likelihood score.
The four hyperparameters fall into two categories:
a1
and b0
are weights for observing the predicted state, while a0
and b1
are weights for observing errors.
We suggest setting a1=b0
and a0=b1
.
The ratio between the two values should correspond to our assessment of the noise level.
See the example section for an application.
The function FULLmLL
is usually called from within function score
.
mLL |
full marginal likelihood of a model |
pos |
posterior distribution of effect positions in the hierarchy |
mappos |
maximum aposteriori estimate of effect positions |
Florian Markowetz <URL: http://genomics.princeton.edu/~florian>
Markowetz F, Probabilistic Models for Gene Silencing Data. PhD thesis, Free University Berlin, 2006.
data("BoutrosRNAi2002") res <- nem(BoutrosRNAiDiscrete[,9:16],type="FULLmLL",hyperpara=c(1,9,9,1))