[R] Unable to use `eval(parse(text))' in nlme::lme

[Duncan Murdoch]

On 08/02/2015 3:49 PM, Rolf Turner wrote:
> On 09/02/15 06:46, Ravi Varadhan wrote:
>> Hi,
>>
>> I would like to run lme() on a number of response variables in a
>> dataframe in an automatic manner.  Bu, when I use
>> eval(parse(text=yname)) to denote the LHS of the formula in lme(), I
>> get the following error message:
>>
>>
>>
>>> require(nlme)
>>
>>
>>
>>> mod2 <- with(subset(labdata2, Transplant_type!=0 & time >0),
>>> lme(eval(parse(text=yname)) ~ time +  as.factor(gvhd), random =
>>> ~1|Patient, correlation = corAR1(), method="ML",
>>> na.action=na.omit))
>> Error in model.frame.default(formula = ~Patient + yname + time +
>> gvhd,  : variable lengths differ (found for 'yname')
>>
>> The same usage works well in lme4::lmer without any problems.
>>
>>
>>
>> It seems that there is a problem in how the formula object is
>> evaluated in lme().  Is there an alternative way to do this?
> 
> What about trying some'at lahk:
> 
> fmla <- as.formula(paste(yname,"~ time + as.factor(gvhd)"))
> mod2 <- with(...., lme(fmla, random = ....))
> 
> Also you would probably be better off using the data argument rather 
> then using with(); this could have some impact on the environment in 
> which the formula is evaluated.

Formulas are a little tricky:  effectively they are evaluated twice.  If
you type something like

y ~ x

(or eval(parse(text="y ~ x")), or as.formula("y ~ x")) then a formula
object is created.  That object remembers the environment in which it
was created, so later when the modelling function uses it, the x and y
variables are evaluated in the original context.

In your example, as.formula() will convert the string to a formula, and
attach the current environment.  So you'd better hope that whatever
variable yname names, as well as time and gvhd, are all available there.

Duncan Murdoch

> 
> Just stabbing in the dark here since you did not provide a reproducible 
> example.
> 
> cheers,
> 
> Rolf Turner
>