Significance and Specification

Practical route to nonparametric significance testing and model-specification testing in np.
Keywords

npsigtest, npcmstest, significance test, model specification test, np tests

This page covers two regression tests:

If you want minimal downloadable scripts for the two main routes on this page, start with:

Current releases use B consistently for public bootstrap counts. npsigtest() also chooses pivotal statistics for continuous predictors and unstandardized statistics for categorical predictors by default; an explicit pivot choice still overrides that automatic policy.

Significance testing: npsigtest

For a fitted regression \(g(x)=E(Y\mid X=x)\), npsigtest() tests whether a selected predictor contributes to \(g(x)\).

## Fit a local-linear regression before calling the nonparametric significance test
library(np)
data(cps71, package = "np")

fit <- npreg(logwage ~ age,
  regtype = "ll",
  bwmethod = "cv.aic",
  gradients = TRUE,
  data = cps71)

npsigtest(fit, B = 99)

The test retains the nonparametric regression specification rather than replacing it with a linear model.

Model specification testing: npcmstest

npcmstest() tests a parametric conditional-mean specification against a consistent nonparametric alternative.

The wage1 example starts with a simple linear specification and then asks whether that specification is too restrictive.

## Compare a simple linear model to a nonparametric alternative
library(np)
data(wage1, package = "np")

model_ols <- lm(lwage ~ female + married + educ + exper + tenure,
  x = TRUE,
  y = TRUE,
  data = wage1)

X <- data.frame(female = wage1$female,
  married = wage1$married,
  educ = wage1$educ,
  exper = wage1$exper,
  tenure = wage1$tenure)

test_out <- npcmstest(model = model_ols,
  xdat = X,
  ydat = wage1$lwage,
  B = 99,
  nmulti = 1)

summary(test_out)

Practical warning

These tests can be computationally demanding because they may involve both bandwidth selection and bootstrap procedures.

Practical advice:

  • begin with a small problem when exploring a specification,
  • budget separately for estimation and testing,
  • preserve the default search tolerances unless a deliberate exploratory run requires otherwise.

When should I use these?

Question Function
Is a regressor significant in a nonparametric regression setting? npsigtest
Does a parametric regression model look misspecified against a nonparametric alternative? npcmstest

How do these fit with the rest of the site?

Back to top