I have asked this question on Stackoverflow and was told it does not relate
to the sites' mission as it is statistical question, thus I brought it here.
I am fitting a gam mode in the mgcv package to study associations of
environmental pollutants and mortality. The aim is to choose a model with
lowest mgcv and also to reduce the PACF to less than < |0.1|.
library(gamair)
library(mgcv)
data(chicago)
y<-matrix(0,12,5)
for (i in 1:12) {
testdf<-gam(death ~ pm10median + o3median + s(time,k=i*14,fx=T) +
s(tmpd,bs="cr"), data=chicago,family=poisson)
y[i,1]<-i
y[i,2]<-summary(testdf)$p.coeff[2]
ll<-summary(testdf)$p.coeff[2]-1.96*summary(testdf)$se[2]
ul<-summary(testdf)$p.coeff[2]+1.96*summary(testdf)$se[2]
y[i,3]<-ll
y[i,4]<-ul
y[i,5]<-(testdf)$gcv.ubre
}
y
pacf(residuals(testdf),lag.max=90)
The partial autocorrelation of the model residuals at lag 0 and some
subsequent lags is large, thus I want to include autoregressive models to
remove this correlation.
In environmental epidemiology there are some suggestions on how to choose
the best fit of a model based (one example is - Model Choice in Time Series
Studies of Air Pollution and Mortality, Peng et al.
http://biostats.bepress.com/cgi/viewcontent.cgi?article=1055&context=jhubiostat;
)
1) use fixed degree of freedom based on literature
2) Chose df that reduces AIC, BIC
3) Choose df that best predicts the pollution series or outcome series
4) Minimum residual autocorrelation (PACF)- some suggest to reduce the
absolute value of PACF to < 1.
I will follow existing literature in choosing the df of my model, but would
also like to use data driven model choice as a sensitivity analysis. I
would appreciate any suggestion on how to improve this model, and above
all, how to fit a model that would reduce the absolute value of PACF to <
0.1.
Thanks
--
View this message in context:
http://r.789695.n4.nabble.com/GAM-model-to-reduce-PACF-of-a-model-tp4649397.html
Sent from the R help mailing list archive at Nabble.com.
