Common-factor exploration and prediction from stored development results

Two explicit computational helpers retained from existing procedures.

Usage

parallel_analysis(data, variables, independent_units = FALSE,
  continuous_measurements = FALSE, B = 199L, seed = 104729L,
  quantile_probability = .95, complete_cases = FALSE, permutations = NULL)
predict_inference(result, newdata, interval = "none", conf.level = NULL)

Arguments

data: Data frame with unique id and numeric measurement columns.
variables: Explicit measurement column names.
independent_units, continuous_measurements: Required declarations for the permutation design.
B: Prespecified number of replicates, subject to the method's computational limits.
seed: Fixed reproducible seed; caller RNG state is restored.
quantile_probability: Prespecified null-spectrum quantile, using quantile type 6.
complete_cases: Explicit complete-case choice; missingness assumptions remain the analyst's responsibility.
permutations: Optional validated column permutation indices.
result: A finalized prediction_model_linear or prediction_model_logistic result.
newdata: New numeric predictor values matching the fitted model.
interval: "none"; for linear predictions "mean" or "prediction"; for logistic predictions "risk".
conf.level: Confidence level. NULL retains the development result's requested level.

Details

Parallel analysis replaces the correlation diagonal by squared multiple correlations and uses independent column permutations. It supplies exploratory leading-factor guidance without fitting or selecting an ML factor model. Failed replicates are retained and prevent a selected count.

Linear "mean" intervals are pointwise normal-model t confidence intervals for the conditional mean. Linear "prediction" intervals instead cover one independent new outcome and include new-outcome residual variation. Common-row estimands and interval-method labels distinguish these targets; metadata also records the interval target, type and native interpretation. Point predictions without requested intervals have no interval scope. Logistic risk intervals use already fitted bootstrap models and type-6 percentile limits, requiring at least five expected replicates per tail. They concern risk estimation, not a binary outcome, validation performance, or Monte Carlo precision. No model is refitted by predict_inference().

Value

A common result with finalized rows, metadata and full native helper output in payload$native.

Examples

# Parallel analysis of four measurements, three sharing one common factor
set.seed(1)
n <- 60
f <- rnorm(n)
dat <- data.frame(id = seq_len(n),
  v1 = f + rnorm(n, sd = 0.5), v2 = f + rnorm(n, sd = 0.5),
  v3 = f + rnorm(n, sd = 0.5), v4 = rnorm(n))
pa <- parallel_analysis(dat, c("v1", "v2", "v3", "v4"),
  independent_units = TRUE, continuous_measurements = TRUE, B = 100L)
result_data(pa)[, c("estimand", "estimate", "status")]
#>                       estimand   estimate status
#> 1 SMC common-factor eigenvalue  2.1147528     ok
#> 2 SMC common-factor eigenvalue  0.2090136     ok
#> 3 SMC common-factor eigenvalue -0.1008882     ok
#> 4 SMC common-factor eigenvalue -0.1819309     ok

# Predictions with intervals from a finalized linear development result
x1 <- rep(-4:3, 5)
dev <- data.frame(id = seq_along(x1), x1 = x1, x2 = rep(c(0, 1), each = 20),
  y = 1 + 2 * x1 + rep(c(-1, 0, 1, 0.5, -0.5), 8))
fit <- run_inference(method_spec("prediction_model_linear"), data = dev,
  predictors = c("x1", "x2"), independent_units = TRUE, B = 200L)
pred <- predict_inference(fit, data.frame(x1 = c(0, 2), x2 = c(0, 1)),
  interval = "prediction")
result_data(pred)[, c("estimate", "conf.low", "conf.high", "status")]
#>   estimate   conf.low conf.high status
#> 1        1 -0.5293634  2.529363     ok
#> 2        5  3.4578615  6.542138     ok