Fast (Weighted) F-test for Linear Models (with Factors)

fFtest computes an R-squared based F-test for the exclusion of the variables in exc, where the full (unrestricted) model is defined by variables supplied to both exc and X. The test is efficient and designed for cases where both exc and X may contain multiple factors and continuous variables. There is also an efficient 2-part formula method.

Usage

fFtest(...) # Internal method dispatch: formula if is.call(..1) || is.call(..2)

# S3 method for default
fFtest(y, exc, X = NULL, w = NULL, full.df = TRUE, ...)

# S3 method for formula
fFtest(formula, data = NULL, weights = NULL, ...)

Arguments

y: a numeric vector: the dependent variable.
exc: a numeric vector, factor, numeric matrix or list / data frame of numeric vectors and/or factors: variables to test / exclude.
X: a numeric vector, factor, numeric matrix or list / data frame of numeric vectors and/or factors: covariates to include in both the restricted (without exc) and unrestricted model. If left empty (X = NULL), the test amounts to the F-test of the regression of y on exc.
w: numeric. A vector of (frequency) weights.
formula: a 2-part formula: y ~ exc | X, where both exc and X are expressions connected with +, and X can be omitted. Note that other operators (:, *, ^, -, etc.) are not supported, you can interact variables using standard functions like finteraction/itn or magrittr::multiply_by inside the formula e.g. log(y) ~ x1 + itn(x2, x3) | x4 or log(y) ~ x1 + multiply_by(x2, x3) | x4.
data: a named list or data frame.
weights: a weights vector or expression that results in a vector when evaluated in the data environment.
full.df: logical. If TRUE (default), the degrees of freedom are calculated as if both restricted and unrestricted models were estimated using lm() (i.e. as if factors were expanded to matrices of dummies). FALSE only uses one degree of freedom per factor.
...: other arguments passed to fFtest.default or to fhdwithin. Sensible options might be the lm.method argument or further control parameters to fixest::demean, the workhorse function underlying fhdwithin for higher-order centering tasks.

Details

Factors and continuous regressors are efficiently projected out using fhdwithin, and the option full.df regulates whether a degree of freedom is subtracted for each used factor level (equivalent to dummy-variable estimator / expanding factors), or only one degree of freedom per factor (treating factors as variables). The test automatically removes missing values and considers only the complete cases of y, exc and X. Unused factor levels in exc and X are dropped.

Note that an intercept is always added by fhdwithin, so it is not necessary to include an intercept in data supplied to exc / X.

Value

A 5 x 3 numeric matrix of statistics. The columns contain statistics:

the R-squared of the model
the numerator degrees of freedom i.e. the number of variables (k) and used factor levels if full.df = TRUE
the denominator degrees of freedom: N - k - 1.
the F-statistic
the corresponding P-value

The rows show these statistics for:

the Full (unrestricted) Model (y ~ exc + X)
the Restricted Model (y ~ X)
the Exclusion Restriction of exc. The R-squared shown is simply the difference of the full and restricted R-Squared's, not the R-Squared of the model y ~ exc.

If X = NULL, only a vector of the same 5 statistics testing the model (y ~ exc) is shown.

Examples

## We could use fFtest as a simple seasonality test:
fFtest(AirPassengers, qF(cycle(AirPassengers)))         # Testing for level-seasonality
#>   R-Sq.     DF1     DF2 F-Stat. P-value 
#>   0.106      11     132   1.424   0.169 
fFtest(AirPassengers, qF(cycle(AirPassengers)),         # Seasonality test around a cubic trend
        poly(seq_along(AirPassengers), 3))
#>                    R-Sq. DF1 DF2 F-Stat. P-Value
#> Full Model         0.965  14 129 250.585   0.000
#> Restricted Model   0.862   3 140 291.593   0.000
#> Exclusion Rest.    0.102  11 129  33.890   0.000
fFtest(fdiff(AirPassengers), qF(cycle(AirPassengers)))  # Seasonality in first-difference
#>   R-Sq.     DF1     DF2 F-Stat. P-value 
#>   0.749      11     131  35.487   0.000 

## A more classical example with only continuous variables
fFtest(mpg ~ cyl + vs | hp + carb, mtcars)
#>                   R-Sq. DF1 DF2 F-Stat. P-Value
#> Full Model        0.750   4  27  20.261   0.000
#> Restricted Model  0.605   2  29  22.175   0.000
#> Exclusion Rest.   0.145   2  27   7.858   0.002
fFtest(mtcars$mpg, mtcars[c("cyl","vs")], mtcars[c("hp","carb")])
#>                   R-Sq. DF1 DF2 F-Stat. P-Value
#> Full Model        0.750   4  27  20.261   0.000
#> Restricted Model  0.605   2  29  22.175   0.000
#> Exclusion Rest.   0.145   2  27   7.858   0.002
 
## Now encoding cyl and vs as factors
fFtest(mpg ~ qF(cyl) + qF(vs) | hp + carb, mtcars)
#>                   R-Sq. DF1 DF2 F-Stat. P-Value
#> Full Model        0.756   5  26  16.140   0.000
#> Restricted Model  0.605   2  29  22.175   0.000
#> Exclusion Rest.   0.152   3  26   5.395   0.005
fFtest(mtcars$mpg, lapply(mtcars[c("cyl","vs")], qF), mtcars[c("hp","carb")])
#>                   R-Sq. DF1 DF2 F-Stat. P-Value
#> Full Model        0.756   5  26  16.140   0.000
#> Restricted Model  0.605   2  29  22.175   0.000
#> Exclusion Rest.   0.152   3  26   5.395   0.005

## Using iris data: A factor and a continuous variable excluded
fFtest(Sepal.Length ~ Petal.Width + Species | Sepal.Width + Petal.Length, iris)
#>                    R-Sq. DF1 DF2 F-Stat. P-Value
#> Full Model         0.867   5 144 188.251   0.000
#> Restricted Model   0.840   2 147 386.386   0.000
#> Exclusion Rest.    0.027   3 144   9.816   0.000
fFtest(iris$Sepal.Length, iris[4:5], iris[2:3])
#>                    R-Sq. DF1 DF2 F-Stat. P-Value
#> Full Model         0.867   5 144 188.251   0.000
#> Restricted Model   0.840   2 147 386.386   0.000
#> Exclusion Rest.    0.027   3 144   9.816   0.000

## Testing the significance of country-FE in regression of GDP on life expectancy
fFtest(log(PCGDP) ~ iso3c | LIFEEX, wlddev)
#>                      R-Sq.   DF1   DF2   F-Stat.   P-Value
#> Full Model           0.955   199  8822   943.424     0.000
#> Restricted Model     0.602     1  9020 13653.865     0.000
#> Exclusion Rest.      0.353   198  8822   350.373     0.000
fFtest(log(wlddev$PCGDP), wlddev$iso3c, wlddev$LIFEEX)
#>                      R-Sq.   DF1   DF2   F-Stat.   P-Value
#> Full Model           0.955   199  8822   943.424     0.000
#> Restricted Model     0.602     1  9020 13653.865     0.000
#> Exclusion Rest.      0.353   198  8822   350.373     0.000
 
## Ok, country-FE are significant, what about adding time-FE
fFtest(log(PCGDP) ~ qF(year) | iso3c + LIFEEX, wlddev)
#>                     R-Sq.  DF1  DF2  F-Stat.  P-Value
#> Full Model          0.963  258 8763  876.312    0.000
#> Restricted Model    0.955  199 8822  943.424    0.000
#> Exclusion Rest.     0.008   59 8763   30.126    0.000
fFtest(log(wlddev$PCGDP), qF(wlddev$year), wlddev[c("iso3c","LIFEEX")])
#>                     R-Sq.  DF1  DF2  F-Stat.  P-Value
#> Full Model          0.963  258 8763  876.312    0.000
#> Restricted Model    0.955  199 8822  943.424    0.000
#> Exclusion Rest.     0.008   59 8763   30.126    0.000

# Same test done using lm:
data <- na_omit(get_vars(wlddev, c("iso3c","year","PCGDP","LIFEEX")))
full <- lm(PCGDP ~ LIFEEX + iso3c + qF(year), data)
rest <- lm(PCGDP ~ LIFEEX + iso3c, data)
anova(rest, full)
#> Analysis of Variance Table
#> 
#> Model 1: PCGDP ~ LIFEEX + iso3c
#> Model 2: PCGDP ~ LIFEEX + iso3c + qF(year)
#>   Res.Df        RSS Df  Sum of Sq     F    Pr(>F)    
#> 1   8822 3.0044e+11                                  
#> 2   8763 2.5097e+11 59 4.9475e+10 29.28 < 2.2e-16 ***
#> ---
#> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1