Differences in dataset for ITT analysis

[gravesti]

Using the same parameters for the ITT analysis in SAS and in R, I noticed I got different results. I have identified that the dataset that R generated has more records than SAS. It seems like the datasets share many rows and that R has some in addition. I'll look further into the differences and try to find the cause in the code.

[gravesti]

It looks like the expanded dataset doesn't get filtered for eligibility. So many trials are included where the patient is ineligible are included.

II wonder if this line should be elgcount==1? Because it seems like expand is always 1. https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L245

If I make this change, I get the same regression results as SAS.

---

Or could this line only expand the eligible==1 periods? I'm not really sure how this works with data.table. It would probably be a bit faster to avoid doing all the expanding work if we can. https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L200

[RoonakR]

It looks like the expanded dataset doesn't get filtered for eligibility. So many trials are included where the patient is ineligible are included.

II wonder if this line should be elgcount==1? Because it seems like expand is always 1. https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L245

If I make this change, I get the same regression results as SAS.

Or could this line only expand the eligible==1 periods? I'm not really sure how this works with data.table. It would probably be a bit faster to avoid doing all the expanding work if we can. https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L200

Hi Isaac, I will have a look. I don't think the problem is any of these but I will have a look now and let you know. Thank you so much for letting me know.

[RoonakR]

I edited something and I think the problem should be fixed now.

[gravesti]

I don't think that's the right fix. It seems like this only keeps the records which have eligibility==1 in sw_data, but then we have lost all the follow-up rows from periods which weren't eligible to start a trial.

I'm pretty sure the change has to be in expand(). It seems like https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L177 creates the right expand indicator but that gets overwritten at https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L225 So the filtering at the end only applies to the second expand variable. https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L245 I think we need to filter on the first expand variable before it is overwritten.

[gravesti]

I've made some test data.

# 10 patients observed for periods 0 - 10. Only eligible in the first period.
# I expect that the expanded dataset only contains the trial starting in period==0
# i.e. 10 patients * 1 eligible trial * 11 periods = 110 rows

dummy_data <- expand.grid(t = 0:10, id = 1:10) 
dummy_data$treatment <- ifelse(dummy_data$id < 5, 1, 0)
dummy_data$eligible <- ifelse(dummy_data$t == 0,1,0)
dummy_data$outcome <- ifelse(1 < dummy_data$id  & dummy_data$id <= 6 & dummy_data$t==10, 1, 0)

initiators(data_path = my_csv,
                   id = "id",
                   period = "t",
                   treatment = "treatment",
                   outcome = "outcome",
                   eligible = "eligible",
                   model_var = "assigned_treatment",
                   data_dir ="./",
                   numCores = 1)

I either get 660 rows before your change and 10 rows after

[RoonakR]

I don't think that's the right fix. It seems like this only keeps the records which have eligibility==1 in sw_data, but then we have lost all the follow-up rows from periods which weren't eligible to start a trial.

I'm pretty sure the change has to be in expand(). It seems like https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L177

creates the right expand indicator but that gets overwritten at https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L225

So the filtering at the end only applies to the second expand variable. https://github.com/RoonakR/RandomisedTrialsEmulation/blob/5b603ed8d6c73edbe7769ee42ed9d9def86e2241/R/lr_utils.R#L245

I think we need to filter on the first expand variable before it is overwritten.

I updated the code and tested it with the dummy data you sent and it seems like now it works. However, I got an error from parglm because there are fewer observations than covariates. So, I am looking at it.

[gravesti]

Excellent. Thanks @RoonakR! This has fixed the data issue and I get the same data as SAS.

I updated the code and tested it with the dummy data you sent and it seems like now it works. However, I got an error from parglm because there are fewer observations than covariates. So, I am looking at it.

It was a small dataset, so it's possible that it can't be solved.