r2 function

R2s of fixest models

Reports different R2s for fixest estimations (e.g. feglm or feols).

r2(x, type = "all", full_names = FALSE)

Arguments

• x: A fixest object, e.g. obtained with function feglm or feols.

• type: A character vector representing the R2 to compute. The R2 codes are of the form: "wapr2" with letters "w" (within), "a" (adjusted) and "p" (pseudo) possibly missing. E.g. to get the regular R2: use type = "r2", the within adjusted R2: use type = "war2", the pseudo R2: use type = "pr2", etc. Use "cor2" for the squared correlation. By default, all R2s are computed.

• full_names: Logical scalar, default is FALSE. If TRUE then names of the vector in output will have full names instead of keywords (e.g. Squared Correlation

  instead of cor2, etc).

Returns

Returns a named vector.

Details

The pseudo R2s are the McFaddens R2s, that is the ratio of log-likelihoods.

For R2s with no theoretical justification, like e.g. regular R2s for maximum likelihood models -- or within R2s for models without fixed-effects, NA is returned. The single measure to possibly compare all kinds of models is the squared correlation between the dependent variable and the expected predictor.

The pseudo-R2 is also returned in the OLS case, it corresponds to the pseudo-R2 of the equivalent GLM model with a Gaussian family.

For the adjusted within-R2s, the adjustment factor is (n - nb_fe) / (n - nb_fe - K)

with n the number of observations, nb_fe the number of fixed-effects and K

the number of variables.

Examples

# Load trade data
data(trade)

# We estimate the effect of distance on trade (with 3 fixed-effects)
est = feols(log(Euros) ~ log(dist_km) | Origin + Destination + Product, trade)

# Squared correlation:
r2(est, "cor2")

# "regular" r2:
r2(est, "r2")

# pseudo r2 (equivalent to GLM with Gaussian family)
r2(est, "pr2")

# adjusted within r2
r2(est, "war2")

# all four at once
r2(est, c("cor2", "r2", "pr2", "war2"))

# same with full names instead of codes
r2(est, c("cor2", "r2", "pr2", "war2"), full_names = TRUE)

Author(s)

Laurent Berge