[R] attributable cost estimation using aggregate data

n.mitsakakis at utoronto.ca n.mitsakakis at utoronto.ca
Thu Dec 9 20:49:06 CET 2010


I am facing with an unusual problem of using aggregate data in order  
to estimate the attributable cost of a disease, for different stages.  
My data set consist of mean and std estimates of the cost outcome  
corresponding to strata coming from cross-classification of a set of  
factors (age group, gender, co morbidity etc.), as well as the number  
of observations in those strata. Those estimates are separate for the  
controls and cases (of more than one disease levels). Some strata have  
only controls or only cases, and some have only one observation, so no  
estimate for sd. So in most cases (except for the ?atomic? strata)  
individual patient data are not available. For example, the data set  
is something like

disease.level stratum   cost.mean    cost.sd  n.cases
             2    STR1  156359.070         NA        1
             0    STR1    6298.799   6995.153       53
             0    STR2    9892.051  11378.500       38
             1    STR3   24264.470  35450.673       14
             0    STR4   10946.446  15472.971       81
             0    STR5   17095.066  20558.138       50
             2    STR5   44130.380         NA        1
             0    STR6   15979.599  17771.120       41

where disease level 0 indicates control.

I am interested in the estimation of the coefficients for the  
difference disease levels. Since cost is usually very skewed to the  
right, gamma or log-normal is usually preferred to normal  
distribution. There is also known heteroscedasticity (higher mean =  
higher variance) and heterogeneity between strata. I was thinking of  
applying some of the approaches for meta-analysis, and perhaps a  
random effects model, addressing those issues. I was referred to lme  
but I am not sure if it is appropriate (I have no experience with it),  
or if other methods (e.g. Bayesian hierarchical models with WinBUGS)  
would be preferred.
Any lead or suggestion would be highly appreciated.


More information about the R-help mailing list