This repo is intended to replicate the study Impact of vaccination on the association of COVID-19 with cardiovascular diseases: An OpenSAFELY cohort study (Cezard et al., 2024). Specifically doing so in order to investigate the variable selection methodology and handling of confounders within the original study. This is done by means of replicating the section of the study on the "pre-vaccination" cohort and then implementing the following:

• Repeating the methodology of the original paper up to and including the fitting of Cox regression models, the hazard ratios from which form the results of the study
• Running a number of different variable selection methods (lasso, lasso_X and lasso_union, see below) to produce alternative variable sets
• Repeating the fitting of the cox regression models under each of these new sets
• Implementing and running the empirical unconfoundedness test provided by Hartwig et al., 2024 to verify the unconfoundedness of these variable sets

Technical details:

• Detailed protocols are in the protocol folder.

• If you are interested in how we defined our code lists, look in the codelists folder.

• Analyses scripts are in the analysis directory:

  • Dataset definition scripts are in the dataset_definition directory:

    • If you are interested in how we defined our variables, we use the variable script variable_helper_fuctions to define functions that generate variables. We then apply these functions in variables_cohorts to create a dictionary of variables for cohort definitions, and in variables_dates to create a dictionary of variables for calculating study start dates and end dates.
    • If you are interested in how we defined study dates (e.g., index and end dates), these vary by cohort and are described in the protocol. We use the script dataset_definition_dates to generate a dataset with all required dates for each cohort. This script imported all variables generated from variables_dates.
    • If you are interested in how we defined our cohorts, we use the dataset definition script dataset_definition_cohorts to define a function that generates cohorts. This script imports all variables generated from variables_cohorts using the patient's index date, the cohort start date and the cohort end date. This approach is used to generate three cohorts: pre-vaccination, vaccinated, and unvaccinated—found in dataset_definition_prevax, dataset_definition_vax, and dataset_definition_unvax, respectively. For each cohort, the extracted data is initially processed in the preprocess data script preprocess data script, which generates a flag variable for pre-existing respiratory conditions and restricts the data to relevant variables.

  • Dataset cleaning scripts are in the dataset_clean directory:

    • This directory also contains all the R scripts that process, describe, and analyse the extracted data.
    • dataset_clean is the core script which executes all the other scripts in this folder
    • fn-preprocess is the function carrying out initial preprocessing, formatting columns correctly
    • fn-modify_dummy is called from within fn-preprocess.R, and alters the proportions of dummy variables to better suit analyses
    • fn-inex is the inclusion/exclusion function
    • fn-qa is the quality assurance function
    • fn-ref is the function that sets the reference levels for factors

  • Modelling scripts are in the model directory:

    • make_model_input.R works with the output of dataset_clean to prepare suitable data subsets for Cox analysis. Combines each outcome and subgroup in one formatted .rds file.
    • fn-prepare_model_input.R is a companion function to make_model_input.R which handles the interaction with active_analyses.rds.
    • cox-ipw is a reusable action which uses the output of make_model_input.R to fit a Cox model to the data.
    • make_model_output.R combines all the Cox results in one formatted .csv file.

  • The script for generating a random 10% sample of the study population is in the generate_subsample directory:

    • generate_subsample.R generates the subsample itself. The subsample is randomly sampled, but for reproducibility-sake, the seed is set in the program.

  • The script for conducting variable selection using a LASSO (Least absolute shrinkage and selection) model is in the lasso_var_selection directory:

    • lasso_var_selection.R fits a cox-regression model (family = "cox") using the subsample data (10% subsample as generated by generate_subsample.R) and applying a LASSO penalty function (alpha = 1). The regularisation parameter lambda is tuned using cross-validation (cv.glmnet) to minimise cvm (mean cross-validated error). The result is a subset of selected variables whose corresponding coefficient does not shrink to zero. For further information please see the documentation for the glmnet and cv.glmnet functions:
      • glmnet
      • cv.glmnet

  • The script for conducting variable selection using a LASSO X (Least absolute shrinkage and selection for exposure) model which takes the exposure (COVID-19) as the response variable is in the lasso_X_var_selection directory:

    • lasso_X_var_selection.R fits a logistic regression (family = "binomial") using binary exposure (X) to COVID-19 as the response variable and excluding the oucomes (Y, acute MI and subarachnoid haemorrhage / haemorrhage stroke) from the dataset. The model is fit using the subsample data (10% subsample as generated by generate_subsample.R). LASSO penalty is applied (alpha = 1). The regularisation parameter lambda is tuned using cross-validation (cv.glmnet) to minimise cvm (mean cross-validated error). The result is a subset of selected variables whose corresponding coefficient does not shrink to zero. For further information please see the documentation for the glmnet and cv.glmnet functions:
      • glmnet
      • cv.glmnet

  • The script for conducting variable selection using a Union LASSO (Least absolute shrinkage and selection) model is in the lasso_union_var_selection directory:

    • lasso_union_var_selection.R takes the union of the two variable sets selected by lasso_var_selection.R and lasso_X_var_selection.R.

  • The script which implements the Hartwig et al., 2024 empirical unconfoundedness test is in the unconfoundedness_test directory:

    • unconfoundedness_test.R performs the empirical unconfoundedness test in the following manner:
      • A cox-regression model taking the oucomes (Y) as the response is fit in the same manner as in lasso_var_selection.R.
      • A logistic regression model taking the exposure (X) as the response is fit in the same manner as in lasso_X_var_selection.R.
      • These two regression models are used to evaluate associations of each confounder (Z) with the exposure (X) and outcome (Y) in the following manner:
        • For every candidate confounder Z, condition (i) is checked (is Z associated with (i.e., not independent of) X given all other covariates?)
        • For every candidate confounder Z, condition (ii) is checked (are Z and Y are conditionally independent given X and all other covariates?)
        • If any covariate Z satisfies both (i) and (ii), then the covariate set is sufficient for confounding adjustment. If not, then the test is inconclusive.
      • Test conditions, coefifcient values, p-values and standard errors are saved for each confounder Z.

• The active_analyses contains a list of active analyses.

• The project.yaml defines run-order and dependencies for all the analysis scripts. This file should not be edited directly. To make changes to the yaml, edit and run the create_project_actions.R script which generates all the actions.

• Descriptive and Model outputs, including figures and tables are in the released_outputs directory.

Output

Outputs follow OpenSAFELY naming conventions related to suppression rules by adding the suffix "_midpoint6". The suffix "_midpoint6_derived" means that the value(s) are derived from the midpoint6 values. Detailed information regarding naming conventions can be found here.

About the OpenSAFELY framework

The OpenSAFELY framework is a Trusted Research Environment (TRE) for electronic health records research in the NHS, with a focus on public accountability and research quality.

Read more at OpenSAFELY.org.

Licences

As standard, research projects have a MIT license.