new package: eha

Göran Broström gb at
Mon Jun 16 11:20:07 CEST 2003

A few days ago I uploaded to CRAN a new package called 'eha', which 
stands for 'Event History Analysis'. Its main focus is on proportional 
hazards modeling in survival analysis, and in that respect  eha  can
be regarded as a complement and an extension to the 'survival'
package. In fact  eha  requires  survival. Eha contains three functions
for proportional hazards analysis:

1. 'coxreg': Performs Cox regression, almost as 'coxph' in survival. 
There are two methods, 'efron' (default) and 'breslow', exactly as in
coxph. There are two extensions, compared to  coxph: (i) Sampling of 
survivors in risk sets (at event times), which can be useful with 
huge data sets and few events. (ii) The so-called 'weird bootstrap':
For the fitted model, new events are drawn in each risk set with 
probabilities given by the fitted model, independently between 
risk sets (that's the 'weird' part). This is repeated  R  times
and the output is two Rxp matrices, one with the bootstrap estimates 
of the regression coefficients, and one with the corresponding 
standard errors. The analysis is up to the user for now.
The 'boot' package?

2. 'mlreg': A discrete time proportional hazards model is fitted along 
the lines of Kalbfleisch & Prentice (1980, pp. 98--103). See also 
Broström (2002): "Cox regression; Ties withot tears", Communications 
in Statistics, Theory & Methods 31, 285--297. This function has two methods;
"ML", the purely discrete model with one parameter per observed distinct 
event time, and "MPPL", which is a hybrid between Cox regression and 
the discrete model: Only tied event times are associated to a unique
parameter; the untied event times contributes a "Cox regression term".
For completely untied data this results in ordinary Cox regression.
"MPPL" can be regarded as an attempt to handle tied data in Cox regression, 
comparable to the 'efron' method. This method does not break down because 
of too heavily tied data, which the  efron  method might do.    
3. 'weibreg': Weibull regression for left truncated and right censored 
data. Allows for stratification with different shape and scale parameters 
in the strata. 

Moreover, there are functions for extracting subsamples as 'rectangles'
in the Lexis diagram, including external ('communal') covariates in a
'survival data frame', extracting information from risk sets, summary 
statistics from the Lexis diagram, etc, etc.

 Göran Broström                    tel: +46 90 786 5223
 Department of Statistics          fax: +46 90 786 6614
 Umeå University         
 SE-90187 Umeå, Sweden             e-mail: gb at

More information about the R-announce mailing list