[R] problem with 'svyby' function from SURVEY package

[Thomas Lumley]

On Thu, 3 Jun 2010, Roni Kobrosly wrote:

> Hello,
>
> I'm using a complex survey dataset and my goal is to simply spit out a bunch of probability-weighted outcome variable means for the different levels of covariate. So I first define the structure of the study design (I'm using the CDC's NHANES data):
>
> dhanes <- svydesign(id=~PSU, strat=~STRATA, weight=~lab_weight, data=final, nest=TRUE)
>
> No problem there.
> Now I use the "svyby" function as follows:
>
> svyby(~outcome, ~covariate, design=dhanes, svymean, na.rm=T) -> haha
> print(haha)
>
>   covariate outcome   se.outcome
> 1        1   0.4961189 0.08828457
> 2        2   0.4474706 0.22214557
> 3        3   0.5157026 0.12076008
> 4        4   0.6773910 0.20605025
> NA      NA   0.8728167 0.15622274
>
> ...and it works just fine. I get a nice table of the mean and standard error for each level of the covariate. I started writing a custom function to automate this and I had problems. Consider this really basic custom function (that does not seem very different from the above code):
>
> this_is_a_test <-function(outcome, covariate)
> {
>
> svyby(~outcome, ~covariate, design=dhanes, svymean, na.rm=T) -> haha
>
> print(hah)
>
>
> 	}
>

You are asking for the mean of a variable called 'outcome', divided up according to a variable called 'covariate'. Presumably you don't have variables with either of those names, so R is getting confused.

Formulas don't work the way you want them to.  As a simpler example with nothing to do with the survey package

this_is_a_simpler_example<-function(outcome){
    ~outcome
}

> this_is_a_simpler_example(test)
~outcome

If you want to substitute a variable into a formula, you need to do it yourself. In your case, you probably want to use make.formula(), from the survey package

> make.formula("test")
~test
> make.formula(c("fred","barney","wilma"))
~fred + barney + wilma

Presumably you want to do something like

approach_that_works <-function(outcome, covariate, design=dhanes,...) svyby(make.formula(outcome),  make.formula(covariate), design,...)

some_outcomes <- colnames(dhanes)[47:63]

some_covariates <- colnames(dhanes)[83:95]

lapply( some_outcomes,
               function(an_outcome) lapply(some_covariates,  approach_that_works, outcome=an_outcome)
            )


For another recent thread using another approach to a related question, see http://tolstoy.newcastle.edu.au/R/e10/help/10/05/5676.html

     -thomas

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle