pbnftest function

Modified BNF--Durbin--Watson Test and Baltagi--Wu's LBI Test for Panel Models

Tests for AR(1) disturbances in panel models.

pbnftest(x, ...)

## S3 method for class 'panelmodel'
pbnftest(x, test = c("bnf", "lbi"), ...)

## S3 method for class 'formula'
pbnftest(
  x,
  data,
  test = c("bnf", "lbi"),
  model = c("pooling", "within", "random"),
  ...
)

Arguments

• x: an object of class "panelmodel" or of class "formula",
• ...: only relevant for formula interface: further arguments to specify the model to test (arguments passed on to plm()), e.g., effect.
• test: a character indicating the test to be performed, either "bnf" or "lbi" for the (modified) BNF statistic or Baltagi--Wu's LBI statistic, respectively,
• data: a data.frame (only relevant for formula interface),
• model: a character indicating on which type of model the test shall be performed ("pooling", "within", "random", only relevant for formula interface),

Returns

An object of class "htest".

Details

The default, test = "bnf", gives the (modified) BNF statistic, the generalised Durbin-Watson statistic for panels. For balanced and consecutive panels, the reference is Bhargava/Franzini/Narendranathan (1982). The modified BNF is given for unbalanced and/or non-consecutive panels (d1 in formula 16 of \insertCite BALT:WU:99;textualplm).

test = "lbi" yields Baltagi--Wu's LBI statistic \insertCite BALT:WU:99plm, the locally best invariant test which is based on the modified BNF statistic.

No specific variants of these tests are available for random effect models. As the within estimator is consistent also under the random effects assumptions, the test for random effect models is performed by taking the within residuals.

No p-values are given for the statistics as their distribution is quite difficult. \insertCite BHAR:FRAN:NARE:82;textualplm supply tabulated bounds for p = 0.05 for the balanced case and consecutive case.

For large N, \insertCite BHAR:FRAN:NARE:82plm suggest it is sufficient to check whether the BNF statistic is < 2 to test against positive serial correlation.

Examples

data("Grunfeld", package = "plm")

# formula interface, replicate Baltagi/Wu (1999), table 1, test case A:
data_A <- Grunfeld[!Grunfeld[["year"]] %in% c("1943", "1944"), ]
pbnftest(inv ~ value + capital, data = data_A, model = "within")
pbnftest(inv ~ value + capital, data = data_A, test = "lbi", model = "within")

# replicate Baltagi (2013), p. 101, table 5.1:
re <- plm(inv ~ value + capital, data = Grunfeld, model = "random")
pbnftest(re)
pbnftest(re, test = "lbi")

References

\insertRef BALT:13plm

\insertRef BALT:WU:99plm

\insertRef BHAR:FRAN:NARE:82plm

See Also

pdwtest() for the original Durbin--Watson test using (quasi-)demeaned residuals of the panel model without taking the panel structure into account. pbltest(), pbsytest(), pwartest() and pwfdtest() for other serial correlation tests for panel models.

Author(s)

Kevin Tappe