Issue generating adjusted predictions with wbm-models

[Tan2525]

I'm facing some issue generating adjusted predictions with wbm models. My intention is to generate an interaction plot with ggpredict. However, it keeps issuing the following error:

Error in complete.cases(data[[variable]]) : 
  no input has determined the number of cases
In addition: Warning message:
In data.frame(..., check.names = FALSE) :
  row names were found from a short variable and have been discarded

I'm unable to resolve this error. May I know if the ggeffects package is compatible with wbm panel models?

I've included a replication dataset here: dataset__wide.csv.
Below is the code to prepare the dataset and run the model.

# Read csv.
dataset__wide <- read.csv(file = "dataset__wide.csv")

# Pivot data to long.
dataset__long <- dataset__wide %>%
  tidyr::pivot_longer(
    # Exclude the time-invariant variables
    !c(
      ID, 
      Control,
    ),
    names_to = "Variables", values_to = "Values"
  ) %>%
  dplyr::mutate(
    
    # Create a variable keeping track of the waves.
    Wave = case_when(
      str_detect(string = Variables, pattern = "t1") ~ 0,
      str_detect(string = Variables, pattern = "t2") ~ 1,
      TRUE ~ NA_real_
    ),

    # Create a variable to standardize the variable names.
    Variable = case_when(
      !(is.na(Variables)) ~ str_replace_all(string = Variables, pattern = "(_+((t1)|(t2)))", replacement = ""),
      TRUE ~ Variables
    ),
    
  ) %>% 
  dplyr::select(
    !Variables
  ) %>%
  tidyr::pivot_wider(names_from = Variable, values_from = "Values", values_fill = NA_real_)

# Create a panel data frame. 
dataset__long__panel <- panel_data(data = dataset__long, id = ID, wave = Wave)

# Fit the panel model
panel_model <- wbm(
  formula = DV ~ IV + M | Control | IV*M, 
  data = dataset__long__panel,
  family = binomial(link = "logit"),
  use.wave = TRUE,
  wave.factor = TRUE,
  weights = Weights,
  scale = TRUE,
  model = "between",
  control = glmerControl(optimizer = "bobyqa")
)

# Compute adjusted predictions
ggpredict(panel_model, terms = c("IV", "M"), bias_correction = TRUE) %>% plot()

[strengejacke]

This could be an issue in panelr's predict() method. Maybe @jacob-long can help finding out whether this problem is related to predict(), or the ggeffects package?

FWIW, ggemmeans() works.

dataset__wide <- read.csv(file = "~/../Downloads/dataset__wide.csv")
library(panelr)
# Pivot data to long.
dataset__long <- dataset__wide |>
  tidyr::pivot_longer(
    # Exclude the time-invariant variables
    !c(
      ID, 
      Control,
    ),
    names_to = "Variables", values_to = "Values"
  ) |>
  dplyr::mutate(
    
    # Create a variable keeping track of the waves.
    Wave = dplyr::case_when(
      stringr::str_detect(string = Variables, pattern = "t1") ~ 0,
      stringr::str_detect(string = Variables, pattern = "t2") ~ 1,
      TRUE ~ NA_real_
    ),

    # Create a variable to standardize the variable names.
    Variable = dplyr::case_when(
      !(is.na(Variables)) ~ stringr::str_replace_all(string = Variables, pattern = "(_+((t1)|(t2)))", replacement = ""),
      TRUE ~ Variables
    ),
    
  ) |> 
  dplyr::select(
    !Variables
  ) |>
  tidyr::pivot_wider(names_from = Variable, values_from = "Values", values_fill = NA_real_)

# Create a panel data frame. 
dataset__long__panel <- panel_data(data = dataset__long, id = ID, wave = Wave)

# Fit the panel model
panel_model <- wbm(
  formula = DV ~ IV + M | Control | IV*M, 
  data = dataset__long__panel,
  family = binomial(link = "logit"),
  use.wave = TRUE,
  wave.factor = TRUE,
  weights = Weights,
  scale = TRUE,
  model = "between",
  control = glmerControl(optimizer = "bobyqa")
)

d <- expand.grid(lapply(dataset__long__panel[c("IV", "M")], unique))
predict(panel_model, newdata = d)
#> Error in complete.cases(data[[variable]]): no input has determined the number of cases

d <- ggeffects::data_grid(panel_model, c("IV", "M"))
predict(panel_model, newdata = d)
#> Unordered factor wave variable was converted to ordered. You should check
#> that the order is correct.
#> Error in complete.cases(data[[variable]]): no input has determined the number of cases

^{Created on 2024-12-17 with reprex v2.1.1}

[Tan2525]

Thanks @strengejacke for the suggestion to use ggemmeans(). It successfully generated the adjusted predictions and using these, I was able to generate an interaction plot.
.

# Compute estimated predictions, with margin = "marginalmeans".
int_dat <- predict_response(
  model = panel_model,
  terms = c("IV", "M"),
  margin = "marginalmeans"
)

# Generate the interaction plot. 
ggplot(data = int_dat) + 
  geom_line(aes(x = x, y = predicted, colour = group))

You can consider my initial issue (of generating adjusted predictions) solved. However, to shed more light on the earlier problem, I was interested in the methodology applied by this paper: https://doi.org/10.1177/1940161224129270. The author(s) have (graciously) provided the replication dataset [(.Rdata)] (https://drive.google.com/file/d/1a-OPCFA3N0ZC8MNvtA2yp-AzVyXCrV75/view?usp=sharing) and code to generate the interaction plot online. I have extracted the relevant portions of the code below. I find it odd that ggpredict successfully ran for their model/dataset but encountered issues with my case, eventhough the model type is the same (wbm).

####Code for models of trust as a moderator####
library(panelr)
library(ggeffects)
library(ggplot2)

# Load the replication dataset. 
load(file = "replication.RData")

####main models reported in the paper for H3####
mtrust1 <- wbm(aff_pol_PT3 ~ partisan_right2_log + partisan_left2_log + no_partisan2_log + POL_INTEREST| 
                 partisan_right2_log*TRUST_GENERAL + partisan_left2_log*TRUST_GENERAL + no_partisan2_log*TRUST_GENERAL + 
                 GENDER + EDU + AGE + christians + race_rc, use.wave = TRUE, wave.factor = TRUE, data=replication)

ggpredict(mtrust1, terms = c("partisan_right2_log", "TRUST_GENERAL[1,3,5]")) %>% plot() + ylim(-4, 4) + 
  labs(y = "Social polarization", 
       x = "Frequency of use of right-leaning news sources",
       title = "Impact of trust and right-leaning news consumption on social polarization", 
       color = "Trust in news")  +
  scale_fill_manual(values = c("#F8766D", "#619CFF", "#00BA38")) +
  scale_color_manual(
    values = c("#F8766D", "#619CFF", "#00BA38"),
    labels = c("Do not trust at all", "Neither", "Trust completely"))

Anyways, just wanted to share that this package is really valuable to academic research and I would like to express my sincerest thanks to you (and the other maintainers) for all your efforts on this amazing package!