In CAM-Roche/RandomisedTrialsEmulation: Causal Analysis of Observational Time-to-Event Data

knitr::opts_chunk$set(echo = TRUE, eval = FALSE)

User interface

After expansion, we need to fit the outcome model. Before that we need to prepare the outcome data set. This might involve sampling, or some other filtering or variable derivation.

trial_itt <- trial_sequence(...) |> ... |> set_outcome_model(...) |> expand_trials()

simple case

with small data it should just work

fit_msm(trial_itt)

This would load the full expanded data and fit the model

with sampling

-> Maybe delete sample_controls and just have a single method like load_trial_data()

This should be done simply

trial_itt <- sample_controls(trial_itt, p_control = 0.3, seed = 444)
fit_msm(trial_itt)

simple filtering

with small data it should just work

trial_itt <- load_trial_data(trial_itt, period = 1:100, subset_condition = "age > 30", p_control = NULL, seed = NULL)
fit_msm(trial_itt)

some complex filtering

model_data(trial_itt) <- model_data(trial_itt) |> group_by() |> mutate()
fit_msm(trial_itt)

Internal Representation

Currently the input data has a slot @data and a class te_data which has some validation, eg on columns.

We could have a second slot eg @model_data. We would need a different or modified te_data class, at least the validation. Alternatively we could directly have the data.table in the slot and validation could occur in the getter/setter methods.

Option: data.table

Remove te_data class and do all validation in setter methods.

• lose class validation
• lose N and Nobs slots
• lose te_data_unset class
• don't need @data@data

Option: te_data classes

Modify class

• don't validate in class, but have a required columns slot
• can store vector of periods
• lose te_data_unset class

Option: te_data_model, te_data_input

More subclasses which inherit from te_data

• natural in OO
• can have all validation built in
• have slots with N, nobs, period etc

Internal flow

If we have the two data slots, @data, @model_data

• sample_controls() would save the data.table to this slot
• could have a getter/setter model_data(object)/model_data(object)<- to access @model_data@data
• and maybe the save for the input data, input_data(object)/input_data(object)<- to access @data@data
• If trial_msm() is called, but @model_data is empty, it will be read from the te_datastore.
• do we need to check that expansion has been done? --> further consider different class post expansion
• If is(@model_data, "te_data_unset") give error: run data expansion
• If is(@outcome_model, "te_outcome_unset") give error: run set outcome model

Chosen Design

• New class te_input_data which inherits from te_data. This used in @data and created by set_data()
• New class te_outcome_data which inherits from te_data. This used in @outcome_data and created by read_expanded_data, sample_expanded_data. It would also have a slot for trial_periods.
• TrialSequence would still have te_data as the class of the slots to allow te_data_unset to be used.
• Implement new getter and setter for @outcome_data to allow complex modifications.
• set_model_data would check the validity of the data.table to ensure correct columns, as well as update @N, periods
• get_model_data would return data.table

CAM-Roche/RandomisedTrialsEmulation documentation built on April 14, 2025, 7:44 a.m.