Skip to contents

Variable types: rd vs ri

Source: vignettes/rowdep-rowind-concept.Rmd
rowdep-rowind-concept.Rmd
library(swereg)
#> swereg 27.3.6
#> https://papadopoulos-lab.github.io/swereg/
library(data.table)
#> 
#> Attaching package: 'data.table'
#> The following object is masked from 'package:base':
#> 
#>     %notin%

Introduction

In longitudinal registry data analysis with swereg, every derived variable is either row-dependent (its value can change across rows for the same person) or row-independent (fixed per person). swereg uses short name prefixes to make the distinction visible at the call site:

  • rd_: row-dependent (also called “rowdep”). Variables that can change over time for a person.
  • ri_: row-independent (also called “rowind”). Variables that are fixed per person.

Many transformations for analysis-ready datasets convert rd_ variables into ri_ variables. Each one takes the value at a specific moment (“age at first diagnosis”, “year of death”).

The concept

Row-dependent (rd_*) variables

These variables can have different values across time periods for the same person:

  • Education level (rd_education): can improve over time
  • Income (rd_income_inflation_adjusted): changes annually
  • Had diagnosis this week (e11_diag): TRUE/FALSE depending on the specific week
  • Current age (rd_age_continuous): increases continuously
  • Civil status (rd_civil_status): can change on marriage, divorce, bereavement
  • Region of residence (rd_region): changes on relocation

Row-independent (ri_*) variables

These variables have the same value across all time periods for a person:

  • Age at first diagnosis (ri_age_first_dx): fixed once it happens
  • Year of first diagnosis (ri_isoyear_first_dx): fixed once it happens
  • Birth country (ri_birthcountry): never changes
  • Cohort classification (ri_register_tag): person’s role in the study (case, control, sibling, etc.)
  • Age at death (ri_age_death): fixed once death occurs
  • Sex at birth (ri_sex): constant

Why this matters

In longitudinal data, you often need to:

  1. Identify when something first happened (first diagnosis, first prescription).
  2. Capture characteristics at specific time points (education level when first diagnosed).
  3. Create person-level summaries that don’t vary by time period.

A conversion from rd_ to ri_ gives a person-level characteristic that has the same value on every row. Later analyses use it, for example as a baseline covariate in a target trial emulation. There, the enrollment time zero sets which row’s rd_ values apply.

Common transformation patterns

Pattern 1: First-occurrence transformations

The most common transformation finds the first time a condition is TRUE and takes a value from that time point. The condition can be time-based, value-based, or any combination of the two.

# Age at first F64 diagnosis restricted to AFAB individuals
make_rowind_first_occurrence(skeleton,
                             condition = "f64_diag == TRUE & ri_is_amab == FALSE",
                             value_var = "rd_age_continuous",
                             new_var   = "ri_age_first_f64_afab")

# Education level at the time of first diagnosis
make_rowind_first_occurrence(skeleton,
                             condition = "isoyearweek == ri_isoyearweek_first_dx",
                             value_var = "rd_education",
                             new_var   = "ri_education_at_first_dx")

Manual approach (for understanding)

make_rowind_first_occurrence() is a thin wrapper. The manual equivalent is:

skeleton[f64_diag == TRUE & ri_is_amab == FALSE, temp := rd_age_continuous]
skeleton[, ri_age_first_f64_afab := swereg::first_non_na(temp), by = .(id)]
skeleton[, temp := NULL]

The helper function runs this three-line pattern and always removes the temporary column.

Pattern 2: Simple renaming (already row-independent)

When a variable is already row-independent but needs consistent naming:

# If date of birth is truly the same for all rows (it should be)
data.table::setnames(skeleton, "dob", "ri_dob")

# If birth country was added via add_onetime() and is person-level
data.table::setnames(skeleton, "birth_country", "ri_birthcountry")

Use this only when the variable has the same value on all rows for each person. If values differ, fix the data integration first.

Practical example with fake data 

# Create a small skeleton for demonstration
ids <- swereg::fake_demographics$lopnr[1:5]
skeleton <- create_skeleton(ids, 2020, "2020-14")

# Add demographic data (creates ri_ variables)
fake_demographics <- swereg::fake_demographics |>
  data.table::copy() |>
  swereg::make_lowercase_names(date_columns = "doddatum")
#> Found additional date columns not in date_columns: fodelseman. Consider adding them for automatic date parsing.
add_onetime(skeleton, fake_demographics, "lopnr")

# Add diagnosis data (creates rd_ variables: the diag column
# varies week-by-week)
fake_diagnoses <- swereg::fake_diagnoses |>
  data.table::copy() |>
  swereg::make_lowercase_names(date_columns = "indatum")
#> Found additional date columns not in date_columns: utdatum. Consider adding them for automatic date parsing.
add_diagnoses(skeleton, fake_diagnoses, "lopnr",
              codes = list(f64_diag = "F64"))

head(skeleton[id == ids[1]], 8)
#>       id isoyear isoyearweek is_isoyear isoyearweeksun personyears fodelseman
#>    <int>   <int>      <char>     <lgcl>         <Date>       <num>     <char>
#> 1:     1    1900     1900-**       TRUE     1900-07-01           1       1956
#> 2:     1    1901     1901-**       TRUE     1901-06-30           1       1956
#> 3:     1    1902     1902-**       TRUE     1902-06-29           1       1956
#> 4:     1    1903     1903-**       TRUE     1903-06-28           1       1956
#> 5:     1    1904     1904-**       TRUE     1904-07-03           1       1956
#> 6:     1    1905     1905-**       TRUE     1905-07-02           1       1956
#> 7:     1    1906     1906-**       TRUE     1906-07-01           1       1956
#> 8:     1    1907     1907-**       TRUE     1907-06-30           1       1956
#>    doddatum f64_diag
#>      <Date>   <lgcl>
#> 1:     <NA>    FALSE
#> 2:     <NA>    FALSE
#> 3:     <NA>    FALSE
#> 4:     <NA>    FALSE
#> 5:     <NA>    FALSE
#> 6:     <NA>    FALSE
#> 7:     <NA>    FALSE
#> 8:     <NA>    FALSE

Now create ri_ variables from the row-dependent f64_diag:

# ISO year of first F64 diagnosis
make_rowind_first_occurrence(skeleton,
                             condition = "f64_diag == TRUE",
                             value_var = "isoyear",
                             new_var   = "ri_isoyear_first_f64")

# ISO year-week of first F64 diagnosis
make_rowind_first_occurrence(skeleton,
                             condition = "f64_diag == TRUE",
                             value_var = "isoyearweek",
                             new_var   = "ri_isoyearweek_first_f64")

head(skeleton[id == ids[1], .(id, isoyear, isoyearweek, f64_diag,
                              ri_isoyear_first_f64,
                              ri_isoyearweek_first_f64)], 8)
#>       id isoyear isoyearweek f64_diag ri_isoyear_first_f64
#>    <int>   <int>      <char>   <lgcl>                <int>
#> 1:     1    1900     1900-**    FALSE                 2000
#> 2:     1    1901     1901-**    FALSE                 2000
#> 3:     1    1902     1902-**    FALSE                 2000
#> 4:     1    1903     1903-**    FALSE                 2000
#> 5:     1    1904     1904-**    FALSE                 2000
#> 6:     1    1905     1905-**    FALSE                 2000
#> 7:     1    1906     1906-**    FALSE                 2000
#> 8:     1    1907     1907-**    FALSE                 2000
#>    ri_isoyearweek_first_f64
#>                      <char>
#> 1:                  2000-**
#> 2:                  2000-**
#> 3:                  2000-**
#> 4:                  2000-**
#> 5:                  2000-**
#> 6:                  2000-**
#> 7:                  2000-**
#> 8:                  2000-**

f64_diag varies by week, so it is row-dependent. Its name has no rd_ prefix because it comes from a code-registry entry. ri_isoyear_first_f64 and ri_isoyearweek_first_f64 are constant for each person.

Best practices

1. Use prefixes consistently

  • rd_ for time-varying variables
  • ri_ for time-invariant variables

Code-registry columns (ov_*, sv_*, os_*, osd_*, rx_*, op_*, etc.) are row-dependent, but they keep their source prefix. You do not need to add an rd_ prefix to them.

2. Always clean up temporary variables

When you transform by hand, always remove the temporary variables:

skeleton[condition == TRUE, temp := value]
skeleton[, new_ri_var := first_non_na(temp), by = .(id)]
skeleton[, temp := NULL]  # critical

3. Use helper functions when possible

make_rowind_first_occurrence() manages the temporary variable for you, which reduces errors.

4. Validate your transformations

Check that your ri_ variables are row-independent:

# Should return 1 for every person -- all rows have the same value
skeleton[, .(unique_values = uniqueN(ri_isoyear_first_f64)), by = .(id)]

How this fits into the swereg pipeline

In the four-phase RegistryStudy$process_skeletons() pipeline (see vignette("skeleton-pipeline")), the rd_ -> ri_ conversion happens inside phase 3 (randvars) steps. A typical pattern:

  1. Phase 1 – framework produces the base time grid, including rd_age_continuous (derived from isoyearweeksun - dob).
  2. Phase 1b – trim is the one phase that MAY delete skeleton rows.
  3. Phase 2 – codes produces the code-derived columns such as os_e11, osd_i21_to_i24, rx_c10aa.
  4. Phase 3 – randvars produces both:
    • rd_* columns from LISA (annual demographics).
    • ri_* columns from first-occurrence transformations.

The ri_* columns are the baseline covariates that target trial emulation uses in vignette("tte-workflow").

Conclusion

The rd_ / ri_ distinction makes person-week skeletons usable as TTE inputs. The ri_ columns hold the per-person baseline characteristics. The enrollment procedure reads the time-varying rd_ columns at each candidate time zero. make_rowind_first_occurrence() creates ri_ columns from rd_ columns.