transformedTrain: Performs Time Transformation of Spike Trains Fitted with glm...
In STAR: Spike Train Analysis with R

Description Usage Arguments Details Value Note Author(s) References See Also Examples

Transform spike times from a glm or gam fitted model as defined by Ogata (1988) and Brown et al (2002). If the model structure is "correct" and if the model parameters are properly estimated the result of the time transformation should be the realization of a Poisson process with rate 1.

1	transformedTrain(obj, target = obj$data$event, select)

`obj`	An object returned by `gam` or `glm`.
`target`	A binary (0,1) vector of integers with the same length as `dim(obj$data)[1]` or a vector of indexes giving the discretized times of events. All these indexes should then be included in `seq(dim(obj$data)[1])`.
`select`	A character string defining a condition to be fulfilled by the event in order to be selected, like: `time <= 6`. This is evaluated after parsing in the data frame of `obj`.

The fitted.values component of obj contains the (estimated) probability to observe a spike in each time bin where the covariates required by the fitted model were defined. It is then straightforward to show using the concept of product integral (Kalbfleisch and Prentice, 2002; Andersen et al, 1993),provided that the time bin width is small enough to have a very small probability in each bin, that the cumulated sum of these probabilities is the expected number of events observed up to a given time. This expected number of events which is returned by transformedTrain. It is also the result of the "time transformation" proposed by Ogata (1988) and brought to the spike train analysis field under the name "time rescaling (theorem)" by Brown et al (2002).

transformedTrain can also be used to transform the times of the spikes of neurons whose spike trains were simultaneously recorded and discretized in exactly the same way as the neuron used to generate obj. This is useful to explore the possibility of functional interactions between a putative pre-synaptic neuron (whose spike train would correspond to argument target) and a post-synaptic one used to generate obj.

transformedTrain returns an object of class transformedTrain inheriting from class spikeTrain. The object is fundamentally a numeric vector with strictly increasing elements containing the transformed times (or the expected number of events).

As mentioned only the spikes for which the covariates of the model are available have their times transformed. That practically means that the length of the transformedTrain object returned by function transformedTrain can be shorter than the length of the original spikeTrain object (or more precisely than the number of spikes defined in target). If one works with a model involving the elapsed times since the last three spikes then the fourth spike of the train will be the first to be transformed. You should therefore expect some left truncation of the data at the beginning of each acquisition epoch.

Christophe Pouzat christophe.pouzat@gmail.com

Ogata, Yosihiko (1988) Statistical Models for Earthquake Occurrences and Residual Analysis for Point Processes. Journal of the American Statistical Association 83: 9-27.

Brown, E. N., Barbieri, R., Ventura, V., Kass, R. E. and Frank, L. M. (2002) The time-rescaling theorem and its application to neural spike train data analysis. Neural Computation 14: 325-346.

Kalbfleisch, John D. and Prentice, Ross L. (2002) The Statistical Analysis of Failure Time Data. Wiley Interscience.

Andersen, Per Kragh, Borgan, Ornulf, Gill, Richard D. and Keiding, Niels (1993) Statistical Models Based on Counting Processes. Springer-Verlag.

plot.transformedTrain, summary.transformedTrain, mkGLMdf, data.frame, glm, mgcv

## Not run: 
## Let us consider neuron 1 of the CAL2S data set
data(CAL2S)
CAL2S <- lapply(CAL2S,as.spikeTrain)
CAL2S[["neuron 1"]]
renewalTestPlot(CAL2S[["neuron 1"]])
summary(CAL2S[["neuron 1"]])
## Make a data frame with a 4 ms time resolution
cal2Sdf <- mkGLMdf(CAL2S,0.004,0,60)
## keep the part relative to neuron 1, 2 and 3 separately
n1.cal2sDF <- cal2Sdf[cal2Sdf$neuron=="1",]
n2.cal2sDF <- cal2Sdf[cal2Sdf$neuron=="2",]
n3.cal2sDF <- cal2Sdf[cal2Sdf$neuron=="3",]
## remove unnecessary data
rm(cal2Sdf)
## Extract the elapsed time since the second to last and
## third to last for neuron 1. Normalise the result. 
n1.cal2sDF[c("rlN.1","rsN.1","rtN.1")] <- brt4df(n1.cal2sDF,"lN.1",2,c("rlN.1","rsN.1","rtN.1"))
## load mgcv library
library(mgcv)
## fit a model with a tensorial product involving the last
## three spikes and using a cubic spline basis for the last two
## To gain time use a fixed df regression spline
n1S.fitA <- gam(event ~ te(rlN.1,rsN.1,bs="cr",fx=TRUE) + rtN.1,data=n1.cal2sDF,family=binomial(link="logit"))
## transform time
N1.Lambda <- transformedTrain(n1S.fitA)
## check out the resulting spike train using the fact
## that transformedTrain objects inherit from spikeTrain
## objects
N1.Lambda
## Use more formal checks
summary(N1.Lambda)
plot(N1.Lambda,which=c(1,2,4,5),ask=FALSE)
## Transform spike trains of neuron 2 and 3
N2.Lambda <- transformedTrain(n1S.fitA,n2.cal2sDF$event)
N3.Lambda <- transformedTrain(n1S.fitA,n3.cal2sDF$event)
## Check interactions
summary(N2.Lambda %frt% N1.Lambda)
summary(N3.Lambda %frt% N1.Lambda)
plot(N2.Lambda %frt% N1.Lambda,ask=FALSE)
plot(N3.Lambda %frt% N1.Lambda,ask=FALSE)

## End(Not run)