0

I'm really scratching my head here. I really don't understand what is going on. This is a MWE, but the actual code and purpose is more complex then this. So the code:

library(dplyr)
ds <- mutate(iris, Species = as.numeric(Species == 'setosa'))

ds %>%
    do_(
        .dots = lazyeval::interp(
            "broom::tidy(stats::glm(form, data = ., family = distr))",
            form = Species ~ Sepal.Length,
            distr = binomial()
        )
    )

Which returns: Error in family$linkfun(mustart) : object 'C_logit_link' not found ... but this code bit works fine:

ds %>%
    do_(
        .dots = lazyeval::interp(
            "broom::tidy(stats::glm(form, data = ., family = distr))",
            form = Sepal.Width ~ Sepal.Length,
            distr = gaussian()
        )
    )

The only difference between the two is the family distribution used (gaussian vs binomial) and the variable used.

So the question: why is it that lazyeval can't find C_logit_link?

lmo
  • 37,904
  • 9
  • 56
  • 69
Luke W. Johnston
  • 954
  • 9
  • 17

1 Answers1

1

When you call interp(x, *), it evaluates the arguments that are to be interpolated into x. In the case of binomial(), the result is a structure that represents the binomial distribution in a GLM.

interp(~x, x=binomial())

#~list(family = "binomial", link = "logit", linkfun = function (mu) 
#.Call(C_logit_link, mu), linkinv = function (eta) 
#.Call(C_logit_linkinv, eta), variance = function (mu) 
#mu * (1 - mu), dev.resids = function (y, mu, wt) 
#.Call(C_binomial_dev_resids, y, mu, wt), aic = function (y, n, 
#    mu, wt, dev) 
#{
#    m <- if (any(n > 1)) 
#    . . .

Buried inside that structure is a function that calls out to compiled C code, via the object C_logit_link. This is an unexported object in the stats package. Normally everything works fine, because the environment of that function is the stats namespace and so it's able to find C_logit_link.

The problem here is that the object you're interpolating is a string, which means that everything interpolated into it is also coerced into a string. That loses the environment information necessary to find C_logit_link.

The solution is to interp a formula instead:

library(dplyr)
ds <- mutate(iris, Species = as.numeric(Species == 'setosa'))

ds %>%
    do_(
        .dots = lazyeval::interp(
            ~broom::tidy(stats::glm(form, data = ., family = distr)),  # formula
            form = Species ~ Sepal.Length,
            distr = binomial()
        )
    )

#          term  estimate std.error statistic      p.value
#1  (Intercept) 27.828521 4.8275611  5.764509 8.189574e-09
#2 Sepal.Length -5.175698 0.8933984 -5.793270 6.902910e-09
Hong Ooi
  • 56,353
  • 13
  • 134
  • 187