Skip to contents

Linear regression association analysis

Source: R/assoc.R
assoc_linear.Rd

Fits one or more linear regression models for each exposure variable and returns a tidy result table. By default, two standard adjustment models are always included:

Usage

assoc_linear(
  data,
  outcome_col,
  exposure_col,
  covariates = NULL,
  base = TRUE,
  conf_level = 0.95
)

assoc_lm(
  data,
  outcome_col,
  exposure_col,
  covariates = NULL,
  base = TRUE,
  conf_level = 0.95
)

Arguments

data

(data.frame or data.table) Analysis dataset.

outcome_col

(character) Name of the continuous numeric outcome column.

exposure_col

(character) One or more exposure variable names.

covariates

(character or NULL) Covariate column names for the Fully adjusted model. Default: NULL.

base

(logical) Include Unadjusted and Age and sex adjusted models. Default: TRUE.

conf_level

(numeric) Confidence level. Default: 0.95.

Value

A data.table with one row per exposure \(\times\) term \(\times\) model combination, and columns:

exposure

Exposure variable name.

term

Coefficient name (e.g. "bmi_categoryObese").

model

Ordered factor: Unadjusted < Age and sex adjusted < Fully adjusted.

n

Participants in model (after NA removal).

beta

Regression coefficient (\(\beta\)).

se

Standard error of \(\beta\).

CI_lower

Lower confidence bound.

CI_upper

Upper confidence bound.

p_value

t-test p-value.

beta_label

Formatted string, e.g. "0.23 (0.05-0.41)".

Details

  • Unadjusted - no covariates (crude).

  • Age and sex adjusted - age + sex auto-detected from the data via UKB field IDs (21022 and 31). Skipped with a warning if either column cannot be found.

  • Fully adjusted - the covariates supplied via the covariates argument. Only run when covariates is non-NULL.

Outcome: intended for continuous numeric variables. Passing a binary (0/1) or logical column is permitted (linear probability model) but will trigger a warning recommending assoc_logistic instead.

CI method: based on the t-distribution via confint.lm(), which is exact under the normal linear model assumption. There is no ci_method argument (unlike assoc_logistic) as profile likelihood does not apply to lm.

SE column: the standard error of \(\beta\) is included to support downstream meta-analysis and GWAS-style summary statistics.

Examples

dt <- ops_toy(scenario = "association", n = 500)
#> ✔ ops_toy: 500 participants | 33 columns | scenario = "association" | seed = 42

res <- assoc_linear(
  data         = dt,
  outcome_col  = "p21001_i0",
  exposure_col = "p20116_i0",
  covariates   = c("bmi_cat", "tdi_cat"),
  base         = FALSE
)
#> 
#> ── assoc_linear ────────────────────────────────────────────────────────────────
#> ℹ 1 exposure x 1 model = 1 linear regression
#> ℹ Input cohort: 500 participants (n reflects each model's actual analysis set)
#> 
#> ── p20116_i0 ──
#> 
#> ✔   Fully adjusted | p20116_i0Previous: beta 0.10 (-0.31-0.51), p = 0.628
#> ✔   Fully adjusted | p20116_i0Current: beta -0.05 (-0.59-0.49), p = 0.856
#> ✔ Done: 2 result rows across 1 exposure and 1 model.