Collinearity can be, but is not always , a property of just a pair of variables and this is especially true when dealing with categorical variables. So although a high correlation coefficient would be sufficient to establish that collinearity might be a problem, a bunch of pairwise low to medium correlations is not a sufficient test for lack of collinearity. The usual method for continuous mixed or categorical collections for variables is to look at the variance inflation factors (which my memory tells me are proportional to the eigenvalues of the variance-covariance-matrix). At any rate this is the code for the vif-function in package:rms:
vif <-
function (fit)
{
v <- vcov(fit, regcoef.only = TRUE)
nam <- dimnames(v)[[1]]
ns <- num.intercepts(fit)
if (ns > 0) {
v <- v[-(1:ns), -(1:ns), drop = FALSE]
nam <- nam[-(1:ns)]
}
d <- diag(v)^0.5
v <- diag(solve(v/(d %o% d)))
names(v) <- nam
v
}
The reason that categorical variables have a greater tendency to generate collinearity is that the three-way or four-way tabulations often form linear combinations that lead to complete collinearity. You example case is an extreme case of collinearity but you can also get collinearity with
A B C D
1 1 0 0
1 0 1 0
1 0 0 1
Notice that this is collinear because A == B+C+D in all rows. None of pairwise correlations would be high, but the system together causes complete collinearity.
After putting your data into an R object and running lm() on it, it becomes apparent that there is another way to determine collinearity with R and that is because lm will drop factor variables from the results when they are "aliased", which is just another term for being completely collinear.
Here is an example for @Alex demonstrating highly collinear data and the output of vif in that situation. Generally you hope to see variance inflation factors below 10.
> set.seed(123)
> dat2 <- data.frame(res = rnorm(100), A=sample(1:4, 1000, repl=TRUE)
+ )
> dat2$B<-dat2$A
> head(dat2)
res A B
1 -0.56047565 1 1
2 -0.23017749 4 4
3 1.55870831 3 3
4 0.07050839 3 3
5 0.12928774 2 2
6 1.71506499 4 4
> dat2[1,2] <- 2
#change only one value to prevent the "anti-aliasing" routines in `lm` from kicking in
> mod <- lm( res ~ A+B, dat2)
> summary(mod)
Call:
lm(formula = res ~ A + B, data = dat2)
Residuals:
Min 1Q Median 3Q Max
-2.41139 -0.58576 -0.02922 0.60271 2.10760
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.10972 0.07053 1.556 0.120
A -0.66270 0.91060 -0.728 0.467
B 0.65520 0.90988 0.720 0.472
Residual standard error: 0.9093 on 997 degrees of freedom
Multiple R-squared: 0.0005982, Adjusted R-squared: -0.001407
F-statistic: 0.2984 on 2 and 997 DF, p-value: 0.7421
> vif ( mod )
A B
1239.335 1239.335
If you make a fourth variable "C" that is independent of the first two perdictors (admittedly a bad name for a variable since C is also an R function), you get a more desirable result from vif:
dat2$C <- sample(1:4, 1000, repl=TRUE)
vif ( lm( res ~ A + C, dat2) )
#---------
A C
1.003493 1.003493
Edit: I realized that I had not actually created R-representations of a "categorical variable" despite sampling from 1:4. The same sort of result occurs with factor versions of that "sample":
> dat2 <- data.frame(res = rnorm(100), A=factor( sample(1:4, 1000, repl=TRUE) ) )
> dat2$B<-dat2$A
> head(dat2)
res A B
1 -0.56047565 1 1
2 -0.23017749 4 4
3 1.55870831 3 3
4 0.07050839 3 3
5 0.12928774 2 2
6 1.71506499 4 4
> dat2[1,2] <- 2
> #change only one value to prevent the "anti-aliasing" routines in `lm` from kicking in
> mod <- lm( res ~ A+B, dat2)
> summary(mod)
Call:
lm(formula = res ~ A + B, data = dat2)
Residuals:
Min 1Q Median 3Q Max
-2.43375 -0.59278 -0.04761 0.62591 2.12461
Coefficients: (2 not defined because of singularities)
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.11165 0.05766 1.936 0.0531 .
A2 -0.67213 0.91170 -0.737 0.4612
A3 0.01293 0.08146 0.159 0.8739
A4 -0.04624 0.08196 -0.564 0.5728
B2 0.62320 0.91165 0.684 0.4944
B3 NA NA NA NA
B4 NA NA NA NA
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 0.9099 on 995 degrees of freedom
Multiple R-squared: 0.001426, Adjusted R-squared: -0.002588
F-statistic: 0.3553 on 4 and 995 DF, p-value: 0.8404
Notice that two of the factor levels are omitted from the calculation of coefficints. ... because they are completely collinear with the corresponding A levels. So if you want to see what vif returns for factor variables that are almost collinear, you need to change a few more values:
> dat2[1,2] <- 2
> dat2[2,2] <-2; dat2[3,2]<-2; dat2[4,2]<-4
> mod <- lm( res ~ A+B, dat2)
> summary(mod)
Call:
lm(formula = res ~ A + B, data = dat2)
Residuals:
Min 1Q Median 3Q Max
-2.42819 -0.59241 -0.04483 0.62482 2.12461
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.11165 0.05768 1.936 0.0532 .
A2 -0.67213 0.91201 -0.737 0.4613
A3 -1.51763 1.17803 -1.288 0.1980
A4 -0.97195 1.17710 -0.826 0.4092
B2 0.62320 0.91196 0.683 0.4945
B3 1.52500 1.17520 1.298 0.1947
B4 0.92448 1.17520 0.787 0.4317
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 0.9102 on 993 degrees of freedom
Multiple R-squared: 0.002753, Adjusted R-squared: -0.003272
F-statistic: 0.4569 on 6 and 993 DF, p-value: 0.8403
#--------------
> library(rms)
> vif(mod)
A2 A3 A4 B2 B3 B4
192.6898 312.4128 308.5177 191.2080 312.5856 307.5242
