Identify confounders and effect modifiers

Review whether one or more candidate variables may act as confounders or effect modifiers for one or more exposures.

Usage

identify_confounder(
  data,
  outcome = NULL,
  exposure,
  potential_confounder,
  approach = "logit",
  time = NULL,
  event = NULL,
  distribution = "weibull",
  method = "change",
  threshold = 10,
  emm_threshold = 10,
  emm_test = c("interaction", "both", "estimate"),
  interaction_alpha = 0.05,
  format = c("flextable", "gt"),
  theme = c("minimal")
)

Arguments

data: A data frame.
outcome: Outcome variable name. Quoted and bare names are accepted. Required for ordinary regression approaches. Leave unset for Cox and parametric survival approaches and supply time and event instead.
exposure: Exposure variable name(s). Can be a character scalar or vector. Quoted names are recommended in scripts, and bare names are also accepted.
potential_confounder: Candidate confounder/effect-modifier variable name(s). Can be a character scalar or vector. Quoted names are recommended in scripts, and bare names are also accepted.
approach: Regression approach. One of "logit", "logbinomial", "poisson", "robpoisson", "linear", "negbin", "cox", or "survreg".
time: Survival time variable name for approach = "cox" or approach = "survreg". Quoted and bare names are accepted.
event: Event indicator variable name for survival approaches. Values may be coded as 0/1, 1/2, logical, or a two-level factor/character variable.
distribution: Parametric survival distribution for approach = "survreg". One of "weibull", "exponential", "lognormal", or "loglogistic".
method: Confounding assessment method. One of "change", "mh", or "both". "change" compares crude and adjusted model estimates. "mh" compares crude and Mantel-Haenszel pooled estimates and is available for binary outcome, binary exposure, and categorical strata. "both" uses either method.
threshold: Percent change threshold for confounding assessment.
emm_threshold: Threshold for relative spread in stratum-specific estimates when using estimate-based effect-modification screening.
emm_test: One of "interaction", "both", or "estimate".
interaction_alpha: Alpha threshold for interaction p-values.
format: Output table format. One of "flextable" (default) or "gt".
theme: Table theme preset or primitives.

Value

If a single exposure-candidate pair is supplied, returns a detailed list.

If multiple combinations are supplied, returns a list with:

summary: A tibble with one row per exposure-candidate combination.
details: A named list of detailed results for each combination.

Details

The function first assesses possible effect modification using stratum-specific estimates and/or an interaction test. If no important effect modification is detected, it then assesses confounding using the selected method.

This is a screening aid for viewing and organising results. Confounding and effect modification should be interpreted using subject-matter knowledge, study design, and causal diagrams such as DAGs. Automated change-in-estimate and interaction checks should not be used as the sole basis for model adjustment.

For each exposure-candidate pair, the interaction screen fits outcome ~ exposure + potential_confounder + exposure:potential_confounder. It does not adjust for other covariates. Consequently, the reported interaction p-value is a candidate-only screening result, not a final adjusted interaction analysis. Use interaction_models(covariates = ...) when testing a planned, covariate-adjusted interaction.

Use this function when you want to screen one or more candidate variables and organise crude, adjusted, Mantel-Haenszel, and effect-modification signals in one place. For a focused comparison of models with and without a planned exposure-by-modifier interaction term, use interaction_models().

Examples

birthwt_data <- data_birthwt |>
  dplyr::mutate(
    low = factor(low, levels = c(0, 1), labels = c("Normal BW", "Low BW")),
    smoke = factor(smoke, levels = c(0, 1), labels = c("No", "Yes")),
    race = factor(race, levels = c(1, 2, 3),
                  labels = c("White", "Black", "Other"))
  )

identify_confounder(
  data = birthwt_data,
  outcome = low,
  exposure = smoke,
  potential_confounder = race,
  approach = logit
)

lung_data <- data_lungcancer |>
  dplyr::mutate(
    trt = factor(trt, levels = c(1, 2),
                 labels = c("Standard treatment", "Test treatment")),
    prior = factor(prior, levels = c(0, 10), labels = c("No", "Yes"))
  )

identify_confounder(
  data = lung_data,
  time = time,
  event = status,
  exposure = trt,
  potential_confounder = prior,
  approach = cox
)

Usage

Arguments

Value

Details

See also

Examples