Skip to content

Passing formula and weights error

4 messages · John Smith, Duncan Murdoch, William Dunlap

John Smith
# 
Dear R-help:

I am writing a function based on glm and would like some variations of
weights. In the code below, I couldn't understand why the second glm
function fails and don't know how to fix it:

Error in eval(extras, data, env) : object 'newweights' not found
 Calls: print ... eval -> <Anonymous> -> model.frame.default -> eval -> eval
 Execution halted

### R code
y <- rnorm(100)
 x <- rnorm(100)
 data <- data.frame(cbind(x, y))
 weights <- rep(1, 100)
 n <- 100
 myglm <- function(formula, data, weights){
     ## this works
     print(glm(formula, data, family=gaussian(), weights))
     ## this is not working
     newweights <- rep(1, n)
     glm(formula, data, family=gaussian(), weights=newweights)
 }
 myglm(y~., data, weights)
Duncan Murdoch
# 
This came up recently in a discussion of lm() on the R-devel list.  I'd 
assume the same issue applies to glm.

The problem is that the argument to weights is evaluated in the same way 
as arguments in the formula:  first in data, then in the environment of 
the formula.  The latter will eventually lead back to the global 
environment, but won't lead to the local evaluation frame in myglm().

The easiest solution is to add newweights to the data argument, but 
there are a few gotchas here.

First, if newweights is already a column in data, you'll mess things up. 
  So be sure to use a name that can't be there.  That's okay in your 
example.

The second problem is that a dot in the formula will cause problems, 
because it will try to include newweights as a predictor variable.  It's 
possible to work around this, but it's probably better to use a more 
complicated solution instead:  modify the formula environment so it 
starts with a small environment holding newweights.  You don't want to 
add newweights directly to environment(formula), because that will have 
side effects outside your function.

This version of your function takes this more complicated approach:

  myglm <- function(formula, data, weights){
      ## this works
      print(glm(formula, data, family=gaussian(), weights))
      env <- new.env(parent = environment(formula))
      env$newweights <- rep(1, n)
      environment(formula) <- env

      glm(formula, data, family=gaussian(), weights=newweights)
  }

Duncan Murdoch
On 28/08/2020 11:32 a.m., John Smith wrote:
William Dunlap
# 
Note that neither call to glm in your myglm function really works -
the first one is using the 'weights' object from the global
environment, not the weights argument.  E.g., in the fresh R session,
where I avoid making unneeded assignments and use fixed x and y for
repeatability,

  > n <- 16
  > data <- data.frame(x = log2(1:n), y = 1:n)
  > myglm2 <- function(formula, data, weights)
      {
          glm(formula, data=data, family=gaussian(), weights=weights)
      }
  > myglm2(y~., data=data, weights=1/(1:n))
  Error in model.frame.default(formula = formula, data = data, weights
= weights,  :
    invalid type (closure) for variable '(weights)'

The error arises because glm finds stats::weights, a function, not the
argument called weights.  glm(), lm() and their ilk evaluate their
weights and subset arguments in the environment of the formula.  In
this case environment(y~.) is .GlobalEnv, not the function's
environment.  The following function gives one way to deal with this,
by giving formula a new environment that inherits from its original
environment and contains the extra variables.

  > myglm3 <- function(formula, data, weights)
      {
          envir <- list2env(list(weights=weights), parent=environment(formula))
          environment(formula) <- envir
          glm(formula, data=data, family=gaussian(), weights=weights)
      }
  > myglm3(y~., data=data, weights=1/(1:n))

  Call:  glm(formula = formula, family = gaussian(), data = data,
weights = weights)

  Coefficients:
  (Intercept)            x
     -0.09553      2.93352

  Degrees of Freedom: 15 Total (i.e. Null);  14 Residual
  Null Deviance:      60.28
  Residual Deviance: 7.72         AIC: 70.42

This is the same result you get with a direct call to
  glm(y~., data=data, weights=1/(1:n))

This is a common problem and I don't know if there is a FAQ on it or a
standard function to deal with it.

Bill Dunlap
TIBCO Software
wdunlap tibco.com
On Fri, Aug 28, 2020 at 8:33 AM John Smith <jswhct at gmail.com> wrote:
John Smith
# 
Thanks to Duncan and Bill for very helpful tips.
On Fri, Aug 28, 2020 at 11:38 AM William Dunlap <wdunlap at tibco.com> wrote: