fbetween-fwithin: Fast Between (Averaging) and (Quasi-)Within (Centering) Transformations

Description

fbetween and fwithin are S3 generics to efficiently obtain between-transformed (averaged) or (quasi-)within-transformed (demeaned) data. These operations can be performed groupwise and/or weighted. B and W are wrappers around fbetween and fwithin representing the 'between-operator' and the 'within-operator'.

(B / W provide more flexibility than fbetween / fwithin when applied to data frames (i.e. column subsetting, formula input, auto-renaming and id-variable-preservation capabilities…), but are otherwise identical.)

Usage

fbetween(x, …)
 fwithin(x, …)
       B(x, …)
       W(x, …)
# S3 method for default
fbetween(x, g = NULL, w = NULL, na.rm = TRUE, fill = FALSE, …)
# S3 method for default
fwithin(x, g = NULL, w = NULL, na.rm = TRUE, mean = 0, theta = 1, …)
# S3 method for default
B(x, g = NULL, w = NULL, na.rm = TRUE, fill = FALSE, …)
# S3 method for default
W(x, g = NULL, w = NULL, na.rm = TRUE, mean = 0, theta = 1, …)
# S3 method for matrix
fbetween(x, g = NULL, w = NULL, na.rm = TRUE, fill = FALSE, …)
# S3 method for matrix
fwithin(x, g = NULL, w = NULL, na.rm = TRUE, mean = 0, theta = 1, …)
# S3 method for matrix
B(x, g = NULL, w = NULL, na.rm = TRUE, fill = FALSE, stub = "B.", …)
# S3 method for matrix
W(x, g = NULL, w = NULL, na.rm = TRUE, mean = 0, theta = 1, stub = "W.", …)
# S3 method for data.frame
fbetween(x, g = NULL, w = NULL, na.rm = TRUE, fill = FALSE, …)
# S3 method for data.frame
fwithin(x, g = NULL, w = NULL, na.rm = TRUE, mean = 0, theta = 1, …)
# S3 method for data.frame
B(x, by = NULL, w = NULL, cols = is.numeric, na.rm = TRUE,
  fill = FALSE, stub = "B.", keep.by = TRUE, keep.w = TRUE, …)
# S3 method for data.frame
W(x, by = NULL, w = NULL, cols = is.numeric, na.rm = TRUE,
  mean = 0, theta = 1, stub = "W.", keep.by = TRUE, keep.w = TRUE, …)
# Methods for compatibility with plm:
# S3 method for pseries
fbetween(x, effect = 1L, w = NULL, na.rm = TRUE, fill = FALSE, …)
# S3 method for pseries
fwithin(x, effect = 1L, w = NULL, na.rm = TRUE, mean = 0, theta = 1, …)
# S3 method for pseries
B(x, effect = 1L, w = NULL, na.rm = TRUE, fill = FALSE, …)
# S3 method for pseries
W(x, effect = 1L, w = NULL, na.rm = TRUE, mean = 0, theta = 1, …)
# S3 method for pdata.frame
fbetween(x, effect = 1L, w = NULL, na.rm = TRUE, fill = FALSE, …)
# S3 method for pdata.frame
fwithin(x, effect = 1L, w = NULL, na.rm = TRUE, mean = 0, theta = 1, …)
# S3 method for pdata.frame
B(x, effect = 1L, w = NULL, cols = is.numeric, na.rm = TRUE,
  fill = FALSE, stub = "B.", keep.ids = TRUE, keep.w = TRUE, …)
# S3 method for pdata.frame
W(x, effect = 1L, w = NULL, cols = is.numeric, na.rm = TRUE,
  mean = 0, theta = 1, stub = "W.", keep.ids = TRUE, keep.w = TRUE, …)
# Methods for grouped data frame / compatibility with dplyr:
# S3 method for grouped_df
fbetween(x, w = NULL, na.rm = TRUE, fill = FALSE,
         keep.group_vars = TRUE, keep.w = TRUE, …)
# S3 method for grouped_df
fwithin(x, w = NULL, na.rm = TRUE, mean = 0, theta = 1,
        keep.group_vars = TRUE, keep.w = TRUE, …)
# S3 method for grouped_df
B(x, w = NULL, na.rm = TRUE, fill = FALSE,
  stub = "B.", keep.group_vars = TRUE, keep.w = TRUE, …)
# S3 method for grouped_df
W(x, w = NULL, na.rm = TRUE, mean = 0, theta = 1,
  stub = "W.", keep.group_vars = TRUE, keep.w = TRUE, …)

Arguments

a numeric vector, matrix, data frame, panel series (class pseries of package plm), panel data frame (plm::pdata.frame) or grouped data frame (class 'grouped_df').

a factor, GRP object, atomic vector (internally converted to factor) or a list of vectors / factors (internally converted to a GRP object) used to group x.

B and W data.frame method: Same as g, but also allows one- or two-sided formulas i.e. ~ group1 or var1 + var2 ~ group1 + group2. See Examples.

a numeric vector of (non-negative) weights. B/W data frame and pdata.frame methods also allow a one-sided formula i.e. ~ weightcol. The grouped_df (dplyr) method supports lazy-evaluation. See Examples.

cols

data.frame method: Select columns to center/average using a function, column names, indices or a logical vector. Default: All numeric variables. Note: cols is ignored if a two-sided formula is passed to by.

na.rm

logical. Skip missing values in x and w when computing averages. If na.rm = FALSE and a NA or NaN is encountered, the average for that group will be NA, and all data points belonging to that group in the output vector will also be NA.

effect

plm methods: Select which panel identifier should be used as grouping variable. 1L takes the first variable in the plm::index, 2L the second etc. Index variables can also be called by name using a character string. If more than one variable is supplied, the corresponding index-factors are interacted.

stub

a prefix or stub to rename all transformed columns. FALSE will not rename columns.

fill

option to fbetween/B: Logical. TRUE will overwrite missing values in x with the respective average. By default missing values in x are preserved.

mean

option to fwithin/W: The mean to center on, default is 0, but a different mean can be supplied and will be added to the data after the centering is performed. A special option when performing grouped centering is mean = "overall.mean". In that case the overall mean of the data will be added after subtracting out group means.

theta

option to fwithin/W: Double. An optional scalar parameter for quasi-demeaning i.e. x - theta * xi.. This is useful for variance components ('random-effects') estimators. see Details.

keep.by, keep.ids, keep.group_vars

B and W data.frame, pdata.frame and grouped_df methods: Logical. Retain grouping / panel-identifier columns in the output. For data frames this only works if grouping variables were passed in a formula.

keep.w

B and W data.frame, pdata.frame and grouped_df methods: Logical. Retain column containing the weights in the output. Only works if w is passed as formula / lazy-expression.

…

arguments to be passed to or from other methods.

Value

fbetween/B returns x with every element replaced by its (groupwise) mean (xi.). Missing values are preserved if fill = FALSE (the default). fwithin/W returns x where every element was subtracted its (groupwise) mean (x - theta * xi. + mean or, if mean = "overall.mean", x - theta * xi. + theta * x..). See Details.

Details

Without groups, fbetween/B replaces all data points in x with their mean or weighted mean (if w is supplied). Similarly fwithin/W subtracts the (weighted) mean from all data points i.e. centers the data on the mean.

With groups supplied to g, the replacement / centering performed by fbetween/B | fwithin/W becomes groupwise. In terms of panel data notation: If x is a vector in such a panel dataset, xit denotes a single data-point belonging to group i in time-period t (t need not be a time-period). Then xi. denotes x, averaged over t. fbetween/B now returns xi. and fwithin/W returns x - xi.. Thus for any data x and any grouping vector g: B(x,g) + W(x,g) = xi. + x - xi. = x. In terms of variance, fbetween/B only retains the variance between group averages, while fwithin/W, by subtracting out group means, only retains the variance within those groups.

The data replacement performed by fbetween/B can keep (default) or overwrite missing values (option fill = TRUE) in x. fwithin/W can center data simply (default), or add back a mean after centering (option mean = value), or add the overall mean in groupwise computations (option mean = "overall.mean"). Let x.. denote the overall mean of x, then fwithin/W with mean = "overall.mean" returns x - xi. + x.. instead of x - xi.. This is useful to get rid of group-differences but preserve the overall level of the data. In regression analysis, centering with mean = "overall.mean" will only change the constant term. See Examples.

If theta != 1, fwithin/W performs quasi-demeaning x - theta * xi.. If mean = "overall.mean", x - theta * xi. + theta * x.. is returned, so that the mean of the partially demeaned data is still equal to the overall data mean x... A numeric value passed to mean will simply be added back to the quasi-demeaned data i.e. x - theta * xi. + mean.

Now in the case of a linear panel model $y_{i t} = β_{0} + β_{1} X_{i t} + u_{i t}$ with $u_{i t} = α_{i} + ϵ_{i t}$ . If $α_{i} \neq α = c o n s t .$ (there exists individual heterogeneity), then pooled OLS is at least inefficient and inference on $β_{1}$ is invalid. If $E [α_{i} | X_{i t}] = 0$ (mean independence of individual heterogeneity $α_{i}$ ), the variance components or 'random-effects' estimator provides an asymptotically efficient FGLS solution by estimating a transformed model $y_{i t} - θ y_{i .} = β_{0} + β_{1} (X_{i t} - θ X_{i .}) + (u_{i t} - θ u_{i .}$ ), where $θ = 1 - \frac{σ_{α}}{\sqrt{(} σ_{α}^{2} + T σ_{ϵ}^{2})}$ . An estimate of $θ$ can be obtained from the an estimate of ${\hat{u}}_{i t}$ (the residuals from the pooled model). If $E [α_{i} | X_{i t}] \neq 0$ , pooled OLS is biased and inconsistent, and taking $θ = 1$ gives an unbiased and consistent fixed-effects estimator of $β_{1}$ . See Examples.

References

Mundlak, Yair. 1978. On the Pooling of Time Series and Cross Section Data. Econometrica 46 (1): 69-85.

Examples

Run this code

# NOT RUN {
## Simple centering and averaging
head(fbetween(mtcars))
head(B(mtcars))
head(fwithin(mtcars))
head(W(mtcars))
all.equal(fbetween(mtcars) + fwithin(mtcars), mtcars)

## Groupwise centering and averaging
head(fbetween(mtcars, mtcars$cyl))
head(fwithin(mtcars, mtcars$cyl))
all.equal(fbetween(mtcars, mtcars$cyl) + fwithin(mtcars, mtcars$cyl), mtcars)

head(W(wlddev, ~ iso3c, cols = 9:13))    # Center the 5 series in this dataset by country
head(cbind(get_vars(wlddev,"iso3c"),     # Same thing done manually using fwithin..
      add_stub(fwithin(get_vars(wlddev,9:13), wlddev$iso3c), "W.")))

## Using B() and W() for fixed-effects regressions:

# Several ways of running the same regression with cyl-fixed effects
lm(W(mpg,cyl) ~ W(carb,cyl), data = mtcars)                     # Centering each individually
lm(mpg ~ carb, data = W(mtcars, ~ cyl, stub = FALSE))           # Centering the entire data
lm(mpg ~ carb, data = W(mtcars, ~ cyl, stub = FALSE,            # Here only the intercept changes
                        mean = "overall.mean"))
lm(mpg ~ carb + B(carb,cyl), data = mtcars)                     # Procedure suggested by
# ..Mundlak (1978) - partialling out group averages amounts to the same as demeaning the data
# }
# NOT RUN {
 <!-- % No code relying on suggested package -->
plm::plm(mpg ~ carb, mtcars, index = "cyl", model = "within")   # "Proof"..
# }
# NOT RUN {
# This takes the interaction of cyl, vs and am as fixed effects
lm(W(mpg,list(cyl,vs,am)) ~ W(carb,list(cyl,vs,am)), data = mtcars)
lm(mpg ~ carb, data = W(mtcars, ~ cyl + vs + am, stub = FALSE))
lm(mpg ~ carb + B(carb,list(cyl,vs,am)), data = mtcars)

# Now with cyl fixed effects weighted by hp:
lm(W(mpg,cyl,hp) ~ W(carb,cyl,hp), data = mtcars)
lm(mpg ~ carb, data = W(mtcars, ~ cyl, ~ hp, stub = FALSE))
lm(mpg ~ carb + B(carb,cyl,hp), data = mtcars)       # WRONG ! Gives a different coefficient!!

## Manual variance components (random-effects) estimation
res <- HDW(mtcars, mpg ~ carb)[[1]]  # Get residuals from pooled OLS
sig2_u <- fvar(res)
sig2_e <- fvar(fwithin(res, mtcars$cyl))
T <- length(res) / fndistinct(mtcars$cyl)
sig2_alpha <- sig2_u - sig2_e
theta <- 1 - sqrt(sig2_alpha) / sqrt(sig2_alpha + T * sig2_e)
lm(mpg ~ carb, data = W(mtcars, ~ cyl, theta = theta, mean = "overall.mean", stub = FALSE))
# }
# NOT RUN {
 <!-- % No code relying on suggested package -->
# A slightly different method to obtain theta...
plm::plm(mpg ~ carb, mtcars, index = "cyl", model = "random")
# }

Run the code above in your browser using DataLab

Get 50% off unlimited learning