\newpage
Analytical Plan for Sensitivity of mortality rates to the imputation of missing socioeconomic data: cohort study
Document version
Version | Alterations |
---|---|
01 | Initial version |
Imputing the missing zip codes will decrease the missingness in the dataset and improve the model fit.
The raw data table was created by merging the TBI database with the DCI table, using the Zip codes as merging key. The raw data base had 711 variables collected with 76,665 observations from 19,303 individuals.
From the raw table, multiple analytical datasets will be created by applying various imputation methods to the Zip code values. The creation of the analytical datasets is described in the next section and the imputation procedures are described in section 5.1.4.
Many datasets will be created for this sensitivity analysis, and the many-datasets approach will be used to apply the statistical model (defined in section 5.1.3) to each dataset so that a sensitivity analysis can be performed. The datasets created under this approach will be created in steps, and stored in a single object to which specific code instructions can be applied to all datasets in a single command. This approach will allow for the simultaneous application of the following instructions to all datasets:
After the cleaning process 24 variables were included in the analysis. The total number of observations excluded due to incompleteness and exclusion criteria will be reported in the analysis. Table 1 shows the structure of the analytical dataset.
id | exposure | outcome | Time | SexF | Race | Mar | AGE | PROBLEMUse | EDUCATION | EMPLOYMENT | RURALdc | PriorSeiz | SCI | Cause | RehabPay1 | ResDis | DAYStoREHABdc | FIMMOTD | FIMCOGD | FollowUpPeriod | FIMMOTD4 | FIMCOGD4 |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | ||||||||||||||||||||||
2 | ||||||||||||||||||||||
3 | ||||||||||||||||||||||
… | ||||||||||||||||||||||
N |
Table: Table 1 Analytical dataset structure after variable selection and cleaning.
All variables in the analytical set were labeled according to the raw data provided and values were labeled according to the data dictionary for the preparation of production-quality results tables and figures.
This is a retrospective analysis of a prospective cohort study.
Inclusion criteria
Exclusion criteria
SES of the neighborhood to which the participant was discharged. The SES measure was stratified into its quintiles, and labelled according to the data dictionary to facilitate interpretation of the results.
Specification of outcome measures (Zarin, 2011):
Primary outcome
Death after a brain injury.
The epidemiological profile of the study participants will be described. Demographic and clinical variables will be described as mean (SD) or as counts and proportions (%), as appropriate. The distributions of participants’ characteristics will be summarized in tables and visualized in exploratory plots.
All inferential analyses will be performed in the statistical models (described in the next section).
This analysis will evaluate the sensitivity to the model specification chosen in SAR-2023-016-BH to changes in the SES data (defined as the exposure of that analysis). The model specification used for the sensitivity analysis will be the best model selected in that associated report.
For reference, the specification defined there regresses the hazard on the SES controlling for all covariates listed in section 4.5, except “previous seizure”.
This model specification will be applied on all datasets created from the imputation approaches described in section 5.1.4, and the Schoenfeld test will be applied to verify the proportional hazards assumption on all model terms.
A couple of simple imputation approaches will be applied on missing values for Zip codes, before the DCI data is merged into the TBI database. An LOCF-based imputation will be applied to impute future Zip codes based on the last known value for each individual. An additional dataset will be created by applying both NOCB- and LOCF-based imputations on missing values, with the intention of increasing the proportion of location data before the DCI data is merged and inclusion/exclusion criteria are applied, in particular the criterion that selects only the last valid observation of the individual in the study period. The non-imputed complete case dataset will be used as the control for the evaluation of the LOCF and the NOCB+LOCF datasets.
All analyses will be performed using the significance level of 5%. All significance hypothesis tests and confidence intervals computed will be two-tailed.
N/A
This analysis will be performed using statistical software R
version 4.3.0.
Recommended reporting guideline
The adoption of the EQUATOR network (http://www.equator-network.org/) reporting guidelines have seen increasing adoption by scientific journals. All observational studies are recommended to be reported following the STROBE guideline (von Elm et al, 2014).
This document was elaborated following recommendations on the structure for Statistical Analysis Plans (Gamble, 2017) for better transparency and clarity.
All documents from this consultation were included in the consultant’s Portfolio.
The portfolio is available at:
https://philsf-biostat.github.io/SAR-2023-017-BH/
This analysis is part of a larger project and is supported by other analyses, linked below.
Effect of socioeconomic status in mortality rates after brain injury: cohort study
https://philsf-biostat.github.io/SAR-2023-004-BH/
Time-adjusted effect of socioeconomic status in mortality rates after brain injury: cohort study