EvalTest: Evaluate an estimated PSTR model

Description

EvalTest provides post-estimation evaluation tests for an estimated PSTR model. It supports two null hypotheses:

Parameter constancy: No time variation in parameters (labelled "time-varying").
No remaining nonlinearity: No remaining nonlinearity/heterogeneity given a candidate transition variable (labelled "heterogeneity").

Usage

EvalTest(use, type = c("time-varying", "heterogeneity"), vq = NULL)
WCB_TVTest(use, iB = 100, parallel = FALSE, cpus = 4)
WCB_HETest(use, vq, iB = 100, parallel = FALSE, cpus = 4)

Value

Invisibly returns use with evaluation results added.

tv: A list of parameter-constancy (time-varying) test results, one element per $m$.
ht: A list of no remaining nonlinearity (heterogeneity) test results, one element per $m$.
wcb_tv: A numeric matrix of WB/WCB p-values for parameter-constancy tests (one row per $m$).
wcb_ht: A numeric matrix of WB/WCB p-values for no remaining nonlinearity tests (one row per $m$).

The individual list elements in tv and ht contain LM-type test statistics and p-values (including HAC variants), consistent with the output from LinTest.

Arguments

use: An object of class "PSTR" returned by EstPSTR. The model must be estimated (nonlinear PSTR) before evaluation tests can be run.
type: Character vector. Which evaluation tests to run in EvalTest. Must be a subset of c("time-varying","heterogeneity"). Default is both.
vq: Numeric vector. Candidate transition variable used by the no remaining nonlinearity test. Required if "heterogeneity" is included in type, and required for WCB_HETest.
iB: Integer. Number of bootstrap replications. Default is 100.
parallel: Logical. Whether to use parallel computation (via the snowfall backend).
cpus: Integer. Number of CPU cores used if parallel = TRUE.

Details

Wild bootstrap (WB) and wild cluster bootstrap (WCB) versions are available via WCB_TVTest (parameter constancy) and WCB_HETest (no remaining nonlinearity).

Two equivalent interfaces are available for each test:

Wrapper function, for example EvalTest(use = obj, ...).
R6 method, for example obj$EvalTest(...).

Each wrapper calls the corresponding R6 method and returns use invisibly.

The bootstrap variants are computationally intensive. WB is robust to heteroskedasticity, while WCB is additionally robust to within-individual dependence (cluster dependence). Parallel execution can be enabled via parallel and cpus.

Examples

Run this code

# \donttest{
pstr <- NewPSTR(Hansen99, dep = "inva", indep = 4:20,
               indep_k = c("vala","debta","cfa","sales"),
               tvars = c("vala"), iT = 14)

# estimate first
pstr <- EstPSTR(use = pstr, im = 1, iq = 1, useDelta = TRUE, par = c(.63, 0), method = "CG")

# evaluation tests
pstr <- EvalTest(
  use = pstr,
  type = c("time-varying","heterogeneity"),
  vq = as.matrix(Hansen99[,'vala'])[,1]
)
print(pstr, mode = "evaluation")

# bootstrap variants (requires snowfall)
library(snowfall)
pstr <- WCB_TVTest(
    use = pstr, iB = 4,
    parallel = TRUE, cpus = 2)
pstr <- WCB_HETest(
    use = pstr,
    vq = as.matrix(Hansen99[,'vala'])[,1],
    iB = 4, parallel = TRUE, cpus = 2)
print(pstr, mode = "evaluation")
# }