stats (version 3.6.2)

influence.measures: Regression Deletion Diagnostics

Description

This suite of functions can be used to compute some of the regression (leave-one-out deletion) diagnostics for linear and generalized linear models discussed in Belsley, Kuh and Welsch (1980), Cook and Weisberg (1982), etc.

Usage

influence.measures(model, infl = influence(model))
rstandard(model, …)
# S3 method for lm
rstandard(model, infl = lm.influence(model, do.coef = FALSE),
          sd = sqrt(deviance(model)/df.residual(model)),
          type = c("sd.1", "predictive"), …)
# S3 method for glm
rstandard(model, infl = influence(model, do.coef = FALSE),
          type = c("deviance", "pearson"), …)
rstudent(model, …)
# S3 method for lm
rstudent(model, infl = lm.influence(model, do.coef = FALSE),
         res = infl$wt.res, …)
# S3 method for glm
rstudent(model, infl = influence(model, do.coef = FALSE), …)
dffits(model, infl = , res = )
dfbeta(model, …)
# S3 method for lm
dfbeta(model, infl = lm.influence(model, do.coef = TRUE), …)
dfbetas(model, …)
# S3 method for lm
dfbetas(model, infl = lm.influence(model, do.coef = TRUE), …)
covratio(model, infl = lm.influence(model, do.coef = FALSE),
         res = weighted.residuals(model))
cooks.distance(model, …)
# S3 method for lm
cooks.distance(model, infl = lm.influence(model, do.coef = FALSE),
               res = weighted.residuals(model),
               sd = sqrt(deviance(model)/df.residual(model)),
               hat = infl$hat, …)
# S3 method for glm
cooks.distance(model, infl = influence(model, do.coef = FALSE),
               res = infl$pear.res,
               dispersion = summary(model)$dispersion,
               hat = infl$hat, …)
hatvalues(model, …)
# S3 method for lm
hatvalues(model, infl = lm.influence(model, do.coef = FALSE), …)
hat(x, intercept = TRUE)

Arguments

model

an R object, typically returned by lm or glm.

infl

influence structure as returned by lm.influence or influence (the latter only for the glm method of rstudent and cooks.distance).

res

(possibly weighted) residuals, with proper default.

sd

standard deviation to use, see default.

dispersion

dispersion (for glm objects) to use, see default.

hat

hat values \(H_{ii}\), see default.

type

type of residuals for rstandard, with different options and meanings for lm and glm. Can be abbreviated.

x

the \(X\) or design matrix.

intercept

should an intercept column be prepended to x?

…

further arguments passed to or from other methods.

Details

The primary high-level function is influence.measures which produces a class "infl" object tabular display showing the DFBETAS for each model variable, DFFITS, covariance ratios, Cook's distances and the diagonal elements of the hat matrix. Cases which are influential with respect to any of these measures are marked with an asterisk.

The functions dfbetas, dffits, covratio and cooks.distance provide direct access to the corresponding diagnostic quantities. Functions rstandard and rstudent give the standardized and Studentized residuals respectively. (These re-normalize the residuals to have unit variance, using an overall and leave-one-out measure of the error variance respectively.)

Note that for multivariate lm() models (of class "mlm"), these functions return 3d arrays instead of matrices, or matrices instead of vectors.

Values for generalized linear models are approximations, as described in Williams (1987) (except that Cook's distances are scaled as \(F\) rather than as chi-square values). The approximations can be poor when some cases have large influence.

The optional infl, res and sd arguments are there to encourage the use of these direct access functions, in situations where, e.g., the underlying basic influence measures (from lm.influence or the generic influence) are already available.

Note that cases with weights == 0 are dropped from all these functions, but that if a linear model has been fitted with na.action = na.exclude, suitable values are filled in for the cases excluded during fitting.

For linear models, rstandard(*, type = "predictive") provides leave-one-out cross validation residuals, and the “PRESS” statistic (PREdictive Sum of Squares, the same as the CV score) of model model is

   PRESS <- sum(rstandard(model, type="pred")^2)

The function hat() exists mainly for S (version 2) compatibility; we recommend using hatvalues() instead.

References

Belsley, D. A., Kuh, E. and Welsch, R. E. (1980). Regression Diagnostics. New York: Wiley.

Cook, R. D. and Weisberg, S. (1982). Residuals and Influence in Regression. London: Chapman and Hall.

Williams, D. A. (1987). Generalized linear model diagnostics using the deviance and single case deletions. Applied Statistics, 36, 181--191. 10.2307/2347550.

Fox, J. (1997). Applied Regression, Linear Models, and Related Methods. Sage.

Fox, J. (2002) An R and S-Plus Companion to Applied Regression. Sage Publ.

Fox, J. and Weisberg, S. (2011). An R Companion to Applied Regression, second edition. Sage Publ; http://socserv.socsci.mcmaster.ca/jfox/Books/Companion/.

See Also

influence (containing lm.influence).

‘plotmath’ for the use of hat in plot annotation.

Examples

# NOT RUN {
require(graphics)

## Analysis of the life-cycle savings data
## given in Belsley, Kuh and Welsch.
lm.SR <- lm(sr ~ pop15 + pop75 + dpi + ddpi, data = LifeCycleSavings)

inflm.SR <- influence.measures(lm.SR)
which(apply(inflm.SR$is.inf, 1, any))
# which observations 'are' influential
summary(inflm.SR) # only these
# }
# NOT RUN {
inflm.SR          # all
# }
# NOT RUN {
plot(rstudent(lm.SR) ~ hatvalues(lm.SR)) # recommended by some
plot(lm.SR, which = 5) # an enhanced version of that via plot(<lm>)

## The 'infl' argument is not needed, but avoids recomputation:
rs <- rstandard(lm.SR)
iflSR <- influence(lm.SR)
identical(rs, rstandard(lm.SR, infl = iflSR))
## to "see" the larger values:
1000 * round(dfbetas(lm.SR, infl = iflSR), 3)
cat("PRESS :"); (PRESS <- sum( rstandard(lm.SR, type = "predictive")^2 ))
stopifnot(all.equal(PRESS, sum( (residuals(lm.SR) / (1 - iflSR$hat))^2)))

## Show that "PRE-residuals"  ==  L.O.O. Crossvalidation (CV) errors:
X <- model.matrix(lm.SR)
y <- model.response(model.frame(lm.SR))
## Leave-one-out CV least-squares prediction errors (relatively fast)
rCV <- vapply(seq_len(nrow(X)), function(i)
              y[i] - X[i,] %*% .lm.fit(X[-i,], y[-i])$coef,
              numeric(1))
## are the same as the *faster* rstandard(*, "pred") :
stopifnot(all.equal(rCV, unname(rstandard(lm.SR, type = "predictive"))))


## Huber's data [Atkinson 1985]
xh <- c(-4:0, 10)
yh <- c(2.48, .73, -.04, -1.44, -1.32, 0)
lmH <- lm(yh ~ xh)
# }
# NOT RUN {
summary(lmH)
# }
# NOT RUN {
im <- influence.measures(lmH)
# }
# NOT RUN {
 im 
# }
# NOT RUN {
plot(xh,yh, main = "Huber's data: L.S. line and influential obs.")
abline(lmH); points(xh[im$is.inf], yh[im$is.inf], pch = 20, col = 2)

## Irwin's data [Williams 1987]
xi <- 1:5
yi <- c(0,2,14,19,30)    # number of mice responding to dose xi
mi <- rep(40, 5)         # number of mice exposed
glmI <- glm(cbind(yi, mi -yi) ~ xi, family = binomial)
# }
# NOT RUN {
summary(glmI)
# }
# NOT RUN {
signif(cooks.distance(glmI), 3)   # ~= Ci in Table 3, p.184
imI <- influence.measures(glmI)
# }
# NOT RUN {
 imI 
# }
# NOT RUN {
stopifnot(all.equal(imI$infmat[,"cook.d"],
          cooks.distance(glmI)))
# }