capnrefsmmat / regressinator Goto Github PK

Suppose you do

pop <- population(...)
d <- sample_x(pop, 100) |> sample_y()

fit <- lm(d$y ~ d$x)

Now sampling_distribution(), model_lineup(), and the rest will not work correctly: they use update(fit, data = ...) to update the data, but fit doesn't use the data argument at all.

We should alert to this possibility. Perhaps, for instance, we can look at the call of fit and check whether there's a data argument provided, and give an informative warning otherwise.

It may be useful (when plotting examples, for instance) to get the mean function for a particular population. For instance:

foo <- population(
  x = predictor("runif", min = 0, max = 10),
  y = response(1 + 2 * x, error_scale = 1)
)

f <- mean_function(foo)
f(x = 2) #=> 5

I suppose this API only works if the population has 1 response variable. An alternative would be to take a data frame of predictors and return a data frame with the mean responses, like sample_y() but without the randomness.

Currently, sample_y() sets size = 1 when sampling binomial responses. It should be possible to simulate binomial outcome data with more than one trial, by somehow linking the number of trials to a column in the data.

Maybe response() can take some kind of weights argument, like glm() does, specifying the counts? Are there any other GLMs this mechanism would be useful for?

Instead of sampling X from the population, fix it according to a design, i.e. a data frame specifying the combinations of predictors to set. Something like this:

design <- some_design_function(...)

pop <- population(
  y = response(x1 + 2 * x2, error_scale = 1)
)

pop |>
  design_x(design) |>
  sample_y()

If one predictor is a factor, predictor_value will be a factor, even for those predictors that are not factors. This will make for confusing plots, since ggplot will treat the axes as factors even for numeric predictors, and try to label every unique value.

Partial residuals for factors do not seem useful, so it may be best to filter out factors when generating the data frame.

If evaluating the response expression in sample_y() fails, we throw a nice error giving the expression and indicating which predictors are available in the environment, since most failures would be from someone typo-ing a variable name.

But if we successfully evaluate the response expression, but then it's not valid for the inverse link function (e.g. it's a string or a list), we just barf an error:

Error in response$family$linkinv(as.numeric(.eval_verbosely(response$response_expr,  : 
  'list' object cannot be coerced to type 'double'

We should have an extra tryCatch() in sample_y() to detect this and show the response expression and its output type.

Reproducer:

library(regressinator)
pop <- population(
x = predictor("rnorm", mean = 5, sd = 4), y = response(
4 + 2 * x, # relationship between X and Y family = gaussian(), # errors are Normal error_scale = 10.0 # errors are N(0, 10^2)
) )

simulated_x <- sample_x(pop, n = 50)
sim_data <- sample_y(simulated_x)

Caused by a copy/paste error, this produces

Error in rnorm(n, mean = y_resp, sd = error_scale) : invalid arguments

This isn't a helpful error. response() should complain properly about the missing error_scale.