Trouble using `slidingwin` with binomial GLM (proportional data)

[bluewingedsoul]

Dear climwin team,

Firstly, thank you for creating and maintaining this very useful package for climate window analysis.

I am encountering a problem when using slidingwin() with a binomial GLM, where my response is proportional breeding success (successes and failures). According to the documentation, slidingwin supports general GLMs, but I receive the following error:

Error in basewin(exclude = exclude, xvar = xvar[[paste(allcombos[combo,  : 
  NA values present in biological response. Please remove NA values

However, my input data frames do not contain any NA values in the response or predictors, and a baseline binomial GLM runs fine outside of slidingwin. This suggests the error arises from internal data handling or windowing, not my input data.

---

Minimal example structure:

# Baseline model (works)
baseline_glm <- glm(cbind(successes, failures) ~ 1, family = binomial, data = mydata)

# This triggers the error:
output <- slidingwin(
  xvar = list(myclimatevar = climate_data$variable),
  cdate = climate_data$date,
  bdate = mydata$date,
  baseline = baseline_glm,
  cohort = mydata$cohort,
  cinterval = "day",
  range = c(182, 0),
  type = "relative",
  stat = "sum",
  func = "lin"
)

Additional context

• The documentation indicates that slidingwin accepts general GLMs, but in practice it does not seem to work with binomial family models.
• If I run the same analysis with proportional response data (bounded between 0 and 1) using a gaussian error in the GLM, the slidingwin function works fine.
• However, for such proportional data, a binomial family is more accurate and appropriate to use, yet the slidingwin function does not let me proceed with it.
• I would greatly appreciate any guidance or a workaround to enable binomial GLMs for proportional data in slidingwin, or clarification if there is a current limitation.

What I have checked:

• No NA values in input data
• All date ranges and classes are aligned
• No duplicates, no zero or negative values, all responses valid for binomial GLM
• GLM works fine outside slidingwin

Expected behavior
I was hoping that slidingwin would run with binomial GLMs as documented, or else clarify any limitations or provide a recommended workaround.

Session info

• R version: 4.5.0
• climwin version: 1.2.3
• Platform: Windows

---

Thank you very much for your help and for developing climwin!

[gregor-fausto]

I'm not part of the climwin team but troubleshooting some issues as well. I'm working with binomial models with binary response data (0/1) and these have been running fine for me. I think the issue is not the binomial family, per se, but how the response data are defined.

Here's a modified example from the slidingwin documentation that illustrates what I mean. I took the Mass data and generated drew samples for each observation using Mass as the number of trials, then calculated the proportion of 'successes', and then used an arbitrary threshold to create a binary response variable. I then ran three models, all binomial GLMs. Model 1 has the success/failure structure for the response, model 2 has the proportions as the response and includes number of trials as weights, and model 3 has the binary response. Model 1 returns the error (NA values present...) but models 2 and 3 run.

data(Mass)
data(MassClimate)

# draw random samples using 'Mass' as the number of trials
Mass$Mass_n <- Mass$Mass
Mass$Mass_success <- rbinom(length(Mass$Mass),Mass$Mass_n,prob=.5)
Mass$Mass_failure <- Mass$Mass-Mass$Mass_success 

# calculate the proportion of 'successes'
Mass$Mass_p <- Mass$Mass_success/Mass$Mass_n

# use an arbitrary threshold to create a binary response
Mass$Mass_binary <- ifelse(Mass$Mass_p>0.5,1,0)

## MODEL 1: successes/trials, with data in success/failure format
# use the 'Mass' as the number of trials, use rbinom to draw successes
MassWin <- slidingwin(xvar = list(Temp = MassClimate$Temp), 
                      cdate = MassClimate$Date, 
                      bdate = Mass$Date, baseline =  glm(cbind(Mass_success, Mass_failure) ~ 1, data = Mass,family='binomial'),
                      range = c(100, 0),
                      stat = c("mean"), func = "lin",
                      type = "absolute", refday = c(20, 5),
                      cmissing = FALSE, cinterval = "day")

## MODEL 2: successes/trials, with data as proportions with weights for number of trials
# calculate the proportion of successes, fit model with 
MassWin <- slidingwin(xvar = list(Temp = MassClimate$Temp), 
                      cdate = MassClimate$Date, 
                      bdate = Mass$Date, baseline =  glm(Mass_p ~ 1, data = Mass,family='binomial', weights = Mass_n),
                      range = c(100, 0),
                      stat = c("mean"), func = "lin",
                      type = "absolute", refday = c(20, 5),
                      cmissing = FALSE, cinterval = "day")

## MODEL 3: binary response data
# use an arbitrary threshold to create a binary variable
MassWin <- slidingwin(xvar = list(Temp = MassClimate$Temp), 
                      cdate = MassClimate$Date, 
                      bdate = Mass$Date, baseline = glm(Mass_binary ~ 1, data = Mass,family='binomial'),
                      range = c(100, 0),
                      stat = c("mean"), func = "lin",
                      type = "absolute", refday = c(20, 5),
                      cmissing = FALSE, cinterval = "day")