In the case of Poisson models I would also say that the application often dictates whether your covariates would act additively (which would then imply an identity link) or multiplicatively on a linear scale (which would then imply a log link). But Poisson models with an identity link also normally only make sense and can only stably be fit if one imposes nonnegativity constraints on the fitted coefficients - this can be done using the nnpois function in the R addreg package or using the nnlm function in the NNLM package. So I don't agree that one should fit Poisson models with both an identity and log link and see which one ends up having the best AIC and infer the best model based on purely statistical grounds - rather, in most cases it is dictated by the underlying structure of the problem that one tries to solve or the data at hand.
For example, in chromatography (GC/MS analysis) one would often measure the superimposed signal of several approx Gaussian shaped peaks and this superimposed signal is measured with an electron multiplier, which means that the measured signal are ion counts and therefore Poisson distributed. Since each of the peaks have by definition a positive height and act additively and the noise is Poisson, a nonnegative Poisson model with identity link would be appropriate here, and a log link Poisson model would be plain wrong. In engineering Kullback-Leibler loss is often used as a loss function for such models, and minimizing this loss is equivalent to optimizing the likelihood of a nonnegative identity-link Poisson model (there is also other divergence/loss measures like alpha or beta divergence that have Poisson as a special case).
Below is a numerical example, including a demonstration that a regular unconstrained identity link Poisson GLM does not fit (because of the lack of nonnegativity constraints) and some details on how to fit nonnegative identity-link Poisson models using nnpois, here in the context of deconvoluting a measured superposition of chromatographic peaks with Poisson noise on them using a banded covariate matrix that contains shifted copies of the measured shape of a single peak. Nonnegativity here is important for several reasons: (1) it is the only realistic model for the data at hand (peaks here cannot have negative heights), (2) it is the only way to stably fit a Poisson model with identity link (as otherwise predictions could for some covariate values go negative, which wouldn't make sense and would give numerical problems when one would try to evaluate the likelihood), (3) nonnegativity acts to regularize the regression problem and greatly helps to obtain stable estimates (e.g. you typically don't get the overfitting problems as with ordinary unconstrained regression, where overfitting typically results in coefficients with very large absolute magnitude and opposite signs - you see this effect in the OLS estimate below) and (4) nonnegativity constraints result in sparser estimates that are frequently closer to the ground truth; for the deconvolution problem below e.g. performance is about as good as LASSO regularization, but without requiring one to tune any regularization parameter. (L0-pseudonorm penalized regression still performs slightly better but at a greater computational cost)
# we first simulate some data
require(Matrix)
n = 200
x = 1:n
npeaks = 20
set.seed(123)
u = sample(x, npeaks, replace=FALSE) # unkown peak locations
peakhrange = c(10,1E3) # peak height range
h = 10^runif(npeaks, min=log10(min(peakhrange)), max=log10(max(peakhrange))) # unknown peak heights
a = rep(0, n) # locations of spikes of simulated spike train, which are assumed to be unknown here, and which needs to be estimated from the measured total signal
a[u] = h
gauspeak = function(x, u, w, h=1) h*exp(((x-u)^2)/(-2*(w^2))) # peak shape function
bM = do.call(cbind, lapply(1:n, function (u) gauspeak(x, u=u, w=5, h=1) )) # banded matrix with peak shape measured beforehand
y_nonoise = as.vector(bM %*% a) # noiseless simulated signal = linear convolution of spike train with peak shape function
y = rpois(n, y_nonoise) # simulated signal with random poisson noise on it - this is the actual signal as it is recorded
par(mfrow=c(1,1))
plot(y, type="l", ylab="Signal", xlab="x", main="Simulated spike train (red) to be estimated given known blur kernel & with Poisson noise")
lines(a, type="h", col="red")
# let's now deconvolute the measured signal y with the banded covariate matrix containing shifted copied of the known blur kernel/peak shape bM
# first observe that regular OLS regression without nonnegativity constraints would return very bad nonsensical estimates
weights <- 1/(y+1) # let's use 1/variance = 1/(y+eps) observation weights to take into heteroscedasticity caused by Poisson noise
a_ols <- lm.fit(x=bM*sqrt(weights), y=y*sqrt(weights))$coefficients # weighted OLS
plot(x, y, type="l", main="Ground truth (red), unconstrained OLS estimate (blue)", ylab="Peak shape", xlab="x", ylim=c(-max(y),max(y)))
lines(x,-y)
lines(a, type="h", col="red", lwd=2)
lines(-a_ols, type="h", col="blue", lwd=2)
# now we use weighted nonnegative least squares with 1/variance obs weights as an approximation of nonnegative Poisson regression
# this gives very good estimates & is very fast
library(nnls)
library(microbenchmark)
microbenchmark(a_wnnls <- nnls(A=bM*sqrt(weights),b=y*sqrt(weights))$x) # 7 ms
plot(x, y, type="l", main="Ground truth (red), weighted nnls estimate (blue)", ylab="Signal (black) & peaks (red & blue)", xlab="Time", ylim=c(-max(y),max(y)))
lines(x,-y)
lines(a, type="h", col="red", lwd=2)
lines(-a_wnnls, type="h", col="blue", lwd=2)
# note that this weighted least square estimate in almost identical to the nonnegative Poisson estimate below and that it fits way faster!!!
# an unconstrained identity-link Poisson GLM will not fit:
glmfit = glm.fit(x=as.matrix(bM), y=y, family=poisson(link=identity), intercept=FALSE)
# returns Error: no valid set of coefficients has been found: please supply starting values
# so let's try a nonnegativity constrained identity-link Poisson GLM, fit using bbmle (using port algo, ie Quasi Newton BFGS):
library(bbmle)
XM=as.matrix(bM)
colnames(XM)=paste0("v",as.character(1:n))
yv=as.vector(y)
LL_poisidlink <- function(beta, X=XM, y=yv){ # neg log-likelihood function
-sum(stats::dpois(y, lambda = X %*% beta, log = TRUE)) # PS regular log-link Poisson would have exp(X %*% beta)
}
parnames(LL_poisidlink) <- colnames(XM)
system.time(fit <- mle2(
minuslogl = LL_poisidlink ,
start = setNames(a_wnnls+1E-10, colnames(XM)), # we initialise with weighted nnls estimates, with approx 1/variance obs weights
lower = rep(0,n),
vecpar = TRUE,
optimizer = "nlminb"
)) # very slow though - takes 145s
summary(fit)
a_nnpoisbbmle = coef(fit)
plot(x, y, type="l", main="Ground truth (red), nonnegative Poisson bbmle ML estimate (blue)", ylab="Signal (black) & peaks (red & blue)", xlab="Time", ylim=c(-max(y),max(y)))
lines(x,-y)
lines(a, type="h", col="red", lwd=2)
lines(-a_nnpoisbbmle, type="h", col="blue", lwd=2)
# much faster is to fit nonnegative Poisson regression using nnpois using an accelerated EM algorithm:
library(addreg)
microbenchmark(a_nnpois <- nnpois(y=y,
x=as.matrix(bM),
standard=rep(1,n),
offset=0,
start=a_wnnls+1.1E-4, # we start from weighted nnls estimates
control = addreg.control(bound.tol = 1e-04, epsilon = 1e-5),
accelerate="squarem")$coefficients) # 100 ms
plot(x, y, type="l", main="Ground truth (red), nonnegative Poisson nnpois estimate (blue)", ylab="Signal (black) & peaks (red & blue)", xlab="Time", ylim=c(-max(y),max(y)))
lines(x,-y)
lines(a, type="h", col="red", lwd=2)
lines(-a_nnpois, type="h", col="blue", lwd=2)
# or to fit nonnegative Poisson regression using nnlm with Kullback-Leibler loss using a coordinate descent algorithm:
library(NNLM)
system.time(a_nnpoisnnlm <- nnlm(x=as.matrix(rbind(bM)),
y=as.matrix(y, ncol=1),
loss="mkl", method="scd",
init=as.matrix(a_wnnls, ncol=1),
check.x=FALSE, rel.tol=1E-4)$coefficients) # 3s
plot(x, y, type="l", main="Ground truth (red), nonnegative Poisson nnlm estimate (blue)", ylab="Signal (black) & peaks (red & blue)", xlab="Time", ylim=c(-max(y),max(y)))
lines(x,-y)
lines(a, type="h", col="red", lwd=2)
lines(-a_nnpoisnnlm, type="h", col="blue", lwd=2)
