thr_test_het: Threshold Test under Heteroskedasticity
In mlkremer/thrreg: Threshold Regression Model

Description Usage Arguments Details Value Author(s) References See Also Examples

View source: R/thr_test_het.R

Computes a test for a threshold in linear regression under heteroskedasticity.

thr_test_het(
  df,
  yi,
  xi,
  qi,
  var.names = colnames(df),
  trim_per = 0.15,
  rep = 1000,
  cr = 0.95,
  graph = TRUE,
  quick = 2
)

`df`	Data frame.
`yi`	Integer or character; index or column name of dependent (y) variable in `df`.
`xi`	Integer or character vector; indexes or column names of independent (x) variables in `df`.
`qi`	Integer or character; index or column name of threshold (q) variable in `df`.
`var.names`	Character vector; variable names with `length(var.names) == ncol(df)` corresponding to columns in `df` to be used in threshold regression table. Default is `colnames(df)`.
`trim_per`	Numeric; percentage of sample to trim from ends. Default is `trim_per = .15`.
`rep`	Integer; number of bootstrap replications. Default is `rep = 1000`.
`cr`	Numeric; confidence level used to plot the critical value in the graph. It is not used elsewhere in the analysis. Default is `cr = .95`.
`graph`	Logical; graph indicator. Set `TRUE` (default) to view the graph of the likelihood; set `FALSE` otherwise.
`quick`	Integer; indicator of method used for bootstrap. Set `quick = 1` for quick computation of asymptotic distribution. This method is not a proper bootstrap and may result in excess rejections. It also uses more memory. Set `quick = 2` (default) for a better bootstrap procedure, which also uses less memory, but is more time consuming.

Do not include a constant in the independent variables; the function automatically adds an intercept to the regression.
There are two bootstrap methods which the function can use.

The first method, obtained by setting quick = 1, is the method presented in the paper Hansen, B. E. (1996). Inference When a Nuisance Parameter is Not Identified Under the Null Hypothesis. Econometrica, 64(2):413-430. https://www.ssc.wisc.edu/~bhansen/papers/ecnmt_96.pdf, which simulates the asymptotic null distribution. A computational shortcut is also taken which speeds computational time, at the cost of greater memory usage, so may not be possible for large data sets.

The second method, obtained by setting quick = 2, is a "fixed regressor bootstrap", which is quite close. The difference is that the bootstrap procedure calculates the variance-covariance matrix in each bootstrap replication. This results in a better finite sample approximation. The cost is greater computation time.

The function is set by default to use the second method (quick = 2), which has better sampling properties. If computational time is a concern, switch to the first method (quick = 1). If an "out of workspace memory" message appears, switch back to quick = 2.

A list with components:

f_test: the value of Maximal (Quandt) F-statistic.
p_value: the bootstrap p-value.

Marcel Kremer, marcel.kremer@uni-due.de

Bruce E. Hansen, behansen@wisc.edu

Hansen, B. E. (2000). Sample splitting and threshold estimation. Econometrica, 68(3):575–603. https://doi.org/10.1111/1468-0262.00124. https://www.ssc.wisc.edu/~bhansen/papers/ecnmt_00.pdf.

thr_test_hom for threshold test under homoskedasticity, thr_est for threshold estimation.

## Performs part of the empirical work reported in Hansen (2000)
data <- dur_john
output <- thr_test_het(data, 1, 2:5, 6)

output$f_test
output$p_value