Error in Overall Weights Calculation and Logistic Regression Error in "level" + "eventstudy"

[nina-straub]

Hi there,

First of all, many thanks for this exciting extension of the DiD framework to continuous treatment, and for providing the contdid package!

I am currently trying to implement cont_did in a staggered adoption setup with a continuous treatment where realised doses for treated units range roughly from 1 to 94. In the process, I encountered two issues that might be worth flagging.

---

Issue 1: Error in Overall Weights Calculation

Error message:

Error in overall_weights(att_gt, ...) : something's going wrong calculating overall weights

Occurs in:

• target_parameter = "slope" with aggregation = "dose"
• target_parameter = "level" with aggregation = "dose"

All other parameters were left at their default values.

Note:
This may be related to numerical instability (e.g., floating point precision) in the sum(out_weight) != 1 condition. The issue has already been discussed and illustrated with an example in this previous report: #6

---

Issue 2: Logistic Regression Error in "level" + "eventstudy"

Error message:

Error in eval(family$initialize) : y values must be 0 <= y <= 1

Occurs in:

• target_parameter = "level"
• aggregation = "eventstudy"

Note:
This configuration is the only one calling ptetools::did_attgt internally, which made be think that it is related to DRDID::drdid_panel trying to estimate the propensity score with binary GLM on a treatment-dose not in the [0, 1] intervall.

Minimal Reproducible Example:
Indeed, I was able to reproduce the error by modifying the simulate_contdid_data function such that the treatment dose variable can sample values above 1, and subsequently running the package example with this modyfied setup:

# Change D <- runif(n, 0, 1) to D <- runif(n, 0, 1.1)
sim_contdid_dat_changed <- function(
    n = 5000,
    num_time_periods = 4,
    num_groups = num_time_periods,
    pg = rep(1 / num_groups, num_groups - 1),
    pu = 1 / (num_groups),
    dose_linear_effect = 0,
    dose_quadratic_effect = 0) {
  if (!requireNamespace("tidyr", quietly = TRUE)) {
    stop("Package 'tidyr' is required for this function but is not installed.
         Please install it with install.packages('tidyr').", call. = FALSE)
  }
  time_periods <- 1:num_time_periods
  groups <- c(0, time_periods[-1])
  p <- c(pu, pg)
  G <- sample(groups, n, replace = TRUE, prob = p)
  D <- runif(n, 0, 1.1)
  eta <- rnorm(n, mean = G)
  time_effects <- 1:num_time_periods
  Y0t <- sapply(1:num_time_periods, function(tp) {
    time_effects[tp] + eta + rnorm(n)
  })
  Y1t <- sapply(1:num_time_periods, function(tp) {
    dose_linear_effect * D + dose_quadratic_effect * D^2 + time_effects[tp] + eta + rnorm(n)
  })
  post_mat <- sapply(1:num_time_periods, function(tp) {
    1 * ((G <= tp) & G != 0)
  })
  Y <- post_mat * Y1t + (1 - post_mat) * Y0t
  df <- as.data.frame(Y)
  colnames(df) <- paste0("Y_", 1:num_time_periods)
  df$id <- 1:n
  df$G <- G
  df$D <- D
  df2 <- tidyr::pivot_longer(df,
                             cols = tidyr::starts_with("Y"),
                             names_to = "time_period",
                             names_prefix = "Y_",
                             names_transform = list(time_period = as.numeric),
                             values_to = "Y"
  ) |> as.data.frame()
  df2$D[df2$G == 0] <- 0
  df2
}

set.seed(1234)
# Simulate data (same setup as in example)
df <- sim_contdid_dat_changed(
  n = 5000,
  num_time_periods = 4,
  num_groups = 4,
  dose_linear_effect = 0,
  dose_quadratic_effect = 0
)

# Try to estimate level - eventstudy setting (same as in example)
cd_res <- cont_did(
  yname = "Y",
  tname = "time_period",
  idname = "id",
  dname = "D",
  data = df,
  gname = "G",
  target_parameter = "level",
  aggregation = "eventstudy",
  treatment_type = "continuous",
  control_group = "notyettreated",
  biters = 100,
  cband = TRUE,
  num_knots = 1,
  degree = 3,
)

This leads to the error above.
However, rescaling the treatment back into [0,1] resolves the issue:

# Rescale D
df <- df %>% mutate(D_bin = D / (1 + D))

# Re-run cont_did with level + eventstudy
cd_res <- cont_did(
  yname = "Y",
  tname = "time_period",
  idname = "id",
  dname = "D_bin",
  data = df,
  gname = "G",
  target_parameter = "level",
  aggregation = "dose",
  treatment_type = "continuous",
  control_group = "notyettreated",
  biters = 100,
  cband = TRUE,
  num_knots = 1,
  degree = 3,
)

After this transformation, the function runs again without errors. At least deriving from this, it seems like in the "level" + "eventstudy" setup, only doses between 0 and 1 are supported.

I apologise if this is a known limitation or if I've missed something in the documentation - in that case, a small note on the expected scale of the dose variable would be very helpful for future users.

---

Minor Code Detail

• There seems to be a small duplication/typo in contdid/R/cont_did.R (in lines 348 and 350), where
  dose_est_method = dose_est_method is included twice.

  GitHub reference:
  https://github.com/bcallaway11/contdid/blob/main/R/cont_did.R

---

In any case, I hope this is of some use, and thanks again for all the work that has gone into this package!