12
$\begingroup$

What is an example of perfect collinearity in terms of the design matrix $X$?

I would like an example where $\hat \beta = (X'X)^{-1}X'Y$ can't be estimated because $(X'X)$ is not invertible.

$\endgroup$
  • $\begingroup$ I went through the recommended post for Colinearity and felt that it was sufficient to understand, but a simple example using data would add clarity. $\endgroup$ – TsTeaTime Jul 3 '16 at 11:20
  • 2
    $\begingroup$ What do you mean by "in terms of X and Y"? Colinearity exists among X variables, Y has nothing to do with it. $\endgroup$ – gung - Reinstate Monica Jul 3 '16 at 12:42
  • 1
    $\begingroup$ I have adjusted the question to be more specific in regards to $X$ $\endgroup$ – TsTeaTime Jul 3 '16 at 12:48
  • 1
    $\begingroup$ As multicollineariry of $X$ shows in singularity of $X'X$ you might want also to read this question: stats.stackexchange.com/q/70899/3277. $\endgroup$ – ttnphns Jul 4 '16 at 20:43
10
$\begingroup$

Here is an example with 3 variables, $y$, $x_1$ and $x_2$, related by the equation

$$ y = x_1 + x_2 + \varepsilon $$

where $\varepsilon \sim N(0,1)$

The particular data are

         y x1 x2
1 4.520866  1  2
2 6.849811  2  4
3 6.539804  3  6

So it is evident that $x_2$ is a multiple of $x_1$ hence we have perfect collinearity.

We can write the model as

$$ Y = X \beta + \varepsilon$$

where:

$$ Y = \begin{bmatrix}4.52 \\6.85 \\6.54\end{bmatrix}$$

$$ X = \begin{bmatrix}1 & 1 & 2\\1 & 2 & 4 \\1 & 3 & 6\end{bmatrix}$$

So we have

$$ XX' = \begin{bmatrix}1 & 1 & 2\\1 & 2 & 4 \\1 & 3 & 6\end{bmatrix} \begin{bmatrix}1 & 1 & 1\\1 & 2 & 3 \\2 & 4 & 6\end{bmatrix} = \begin{bmatrix}6 & 11 & 16\\11 & 21 & 31 \\16 & 31 & 46\end{bmatrix} $$

Now we calculate the determinant of $XX'$ :

$$ \det XX' = 6\begin{vmatrix}21 & 31 \\31 & 46\end{vmatrix} - 11 \begin{vmatrix}11 & 31 \\16 & 46\end{vmatrix} + 16\begin{vmatrix}11 & 21 \\16 & 31\end{vmatrix}= 0$$

In R we can show this as follows:

> x1 <- c(1,2,3)

create x2, a multiple of x1

> x2 <- x1*2

create y, a linear combination of x1, x2 and some randomness

> y <- x1 + x2 + rnorm(3,0,1)

observe that

> summary(m0 <- lm(y~x1+x2))

fails to estimate a value for the x2 coefficient:

Coefficients: (1 not defined because of singularities)
            Estimate Std. Error t value Pr(>|t|)
(Intercept)   3.9512     1.6457   2.401    0.251
x1            1.0095     0.7618   1.325    0.412
x2                NA         NA      NA       NA

Residual standard error: 0.02583 on 1 degrees of freedom
Multiple R-squared:      1,     Adjusted R-squared:  0.9999 
F-statistic: 2.981e+04 on 1 and 1 DF,  p-value: 0.003687

The model matrix $X$ is:

> (X <- model.matrix(m0))

(Intercept) x1 x2
1           1  1  2
2           1  2  4
3           1  3  6

So $XX'$ is

> (XXdash <- X %*% t(X))
   1  2  3
1  6 11 16
2 11 21 31
3 16 31 46

which is not invertible, as shown by

> solve(XXdash)
Error in solve.default(XXdash) : 
  Lapack routine dgesv: system is exactly singular: U[3,3] = 0

Or:

det(XXdash) [1] 0

$\endgroup$
21
+50
$\begingroup$

Here are a couple of fairly common scenarios producing perfect multicollinearity, i.e. situations in which the columns of the design matrix are linearly dependent. Recall from linear algebra that this means there is a linear combination of columns of the design matrix (whose coefficients are not all zero) which equals zero. I have included some practical examples to help explain why this pitfall strikes so often — I have encountered almost all of them!

  1. One variable is a multiple of another, regardless of whether there is an intercept term: perhaps because you have recorded the same variable twice using different units (e.g. "length in centimetres" is precisely 100 times larger than "length in metres") or because you have recorded a variable once as a raw number and once as a proportion or percentage, when the denominator is fixed (e.g. "area of petri dish colonized" and "percentage of petri dish colonized" will be exact multiples of each other if the area of each petri dish is the same). We have collinearity because if $w_i = ax_i$ where $w$ and $x$ are variables (columns of your design matrix) and $a$ is a scalar constant, then $1(\vec w) - a(\vec x)$ is a linear combination of variables that is equal to zero.

  2. There is an intercept term and one variable differs from another by a constant: this will happen if you center a variable ($w_i = x_i - \bar x$) and include both raw $x$ and centered $w$ in your regression. It will also happen if your variables are measured in different unit systems that differ by a constant, e.g. if $w$ is "temperature in kelvin" and $x$ as "temperature in °C" then $w_i = x_i + 273.15$. If we regard the intercept term as a variable that is always $1$ (represented as a column of ones, $\vec 1_n$, in the design matrix) then having $w_i = x_i + k$ for some constant $k$ means that $1(\vec w) - 1(\vec x) - k(\vec 1_n)$ is a linear combination of the $w$, $x$ and $1$ columns of the design matrix that equals zero.

  3. There is an intercept term and one variable is given by an affine transformation of another: i.e. you have variables $w$ and $x$, related by $w_i = ax_i + b$ where $a$ and $b$ are constants. For instance this happens if you standardize a variable as $z_i = \frac{x_i - \bar x}{s_x}$ and include both raw $x$ and standardized $z$ variables in your regression. It also happens if you record $w$ as "temperature in °F" and $x$ as "temperature in °C", since those unit systems do not share a common zero but are related by $w_i = 1.8x_i + 32$. Or in a business context, suppose there is fixed cost $b$ (e.g. covering delivery) for each order, as well as a cost $\$a$ per unit sold; then if $\$w_i$ is the cost of order $i$ and $x_i$ is the number of units ordered, we have $w_i = ax_i + b$. The linear combination of interest is $1(\vec w) - a(\vec x) - b(\vec 1_n) = \vec 0$. Note that if $a=1$, then (3) includes (2) as a special case; if $b=0$, then (3) includes (1) as a special case.

  4. There is an intercept term and the sum of several variables is fixed (e.g. in the famous "dummy variable trap"): for example if you have "percentage of satisfied customers", "percentage of dissatisfied customers" and "percentage of customers neither satisfied nor dissatisfied" then these three variables will always (barring rounding error) sum to 100. One of these variables — or alternatively, the intercept term — needs to be dropped from the regression to prevent collinearity. The "dummy variable trap" occurs when you use indicator variables (more commonly but less usefully called "dummies") for every possible level of a categorical variable. For instance, suppose vases are produced in red, green or blue color schemes. If you recorded the categorical variable "color" by three indicator variables (red, green and blue would be binary variables, stored as 1 for "yes" and 0 for "no") then for each vase only one of the variables would be a one, and hence red + green + blue = 1. Since there is a vector of ones for the intercept term, the linear combination 1(red) + 1(green) + 1(blue) - 1(1) = 0. The usual remedy here is either to drop the intercept, or drop one of the indicators (e.g. leave out red) which becomes a baseline or reference level. In this case, the regression coefficient for green would indicate the change in the mean response associated with switching from a red vase to a green one, holding other explanatory variables constant.

  5. There are at least two subsets of variables, each having a fixed sum, regardless of whether there is an intercept term: suppose the vases in (4) were produced in three sizes, and the categorical variable for size was stored as three additional indicator variables. We would havelarge + medium + small = 1. Then we have the linear combination 1(large) + 1(medium) + 1(small) - 1(red) - 1(green) - 1(blue) = 0, even when there is no intercept term. The two subsets need not share the same sum, e.g. if we have explanatory variables $u, v, w, x$ such that every $u_i + v_i = k_1$ and $x_i + y_i = k_2$ then $k_2(\vec u) + k_2(\vec v) - k_1(\vec w) - k_1(\vec x) = \vec 0$.

  6. One variable is defined as a linear combination of several other variables: for instance, if you record the length $l$, width $w$ and perimeter $p$ of each rectangle, then $p_i = 2l_i + 2w_i$ so we have the linear combination $1(\vec p) - 2(\vec l) - 2(\vec w) = \vec 0$. An example with an intercept term: suppose a mail-order business has two product lines, and we record that order $i$ consisted of $u_i$ of the first product at unit cost $\$a$ and $v_i$ of the second at unit cost $\$b$, with fixed delivery charge $\$c$. If we also include the order cost $\$x$ as an explanatory variable, then $x_i = a u_i + b v_i + c$ and so $1(\vec x) - a(\vec u) - b(\vec v) -c(\vec 1_n) = \vec 0$. This is an obvious generalization of (3). It also gives us a different way of thinking about (4): once we know all bar one of the subset of variables whose sum is fixed, then the remaining one is their complement so can be expressed as a linear combination of them and their sum. If we know 50% of customers were satisfied and 20% were dissatisfied, then 100% - 50% - 20% = 30% must be neither satisfied nor dissatisfied; if we know the vase is not red (red=0) and it is green (green=1) then we know it is not blue (blue = 1(1) - 1(red) - 1(green) = 1 - 0 - 1 = 0).

  7. One variable is constant and zero, regardless of whether there is an intercept term: in an observational study, a variable will be constant if your sample does not exhibit sufficient (any!) variation. There may be variation in the population that is not captured in your sample, e.g. if there is a very common modal value: perhaps your sample size is too small and was therefore unlikely to include any values that differed from the mode, or your measurements were insufficiently accurate to detect small variations from the mode. Alternatively, there may be theoretical reasons for the lack of variation, particularly if you are studying a sub-population. In a study of new-build properties in Los Angeles, it would not be surprising that every data point has AgeOfProperty = 0 and State = California! In an experimental study, you may have measured an independent variable that is under experimental control. Should one of your explanatory variables $x$ be both constant and zero, then we have immediately that the linear combination $1(\vec x)$ (with coefficient zero for any other variables) is $\vec 0$.

  8. There is an intercept term and at least one variable is constant: if $x$ is constant so that each $x_i = k \neq 0$, then the linear combination $1(\vec x) - k(\vec 1_n) = \vec 0$.

  9. At least two variables are constant, regardless of whether there is an intercept term: if each $w_i = k_1 \neq 0$ and $x_i = k_2 \neq 0$, then the linear combination $k_2(\vec w) - k_1(\vec x) = \vec 0$.

  10. Number of columns of design matrix, $k$, exceeds number of rows, $n$: even when there is no conceptual relationship between your variables, it is mathematically necessitated that the columns of your design matrix will be linearly dependent when $k > n$. It simply isn't possible to have $k$ linearly independent vectors in a space with a number of dimensions lower than $k$: for instance, while you can draw two independent vectors on a sheet of paper (a two-dimensional plane, $\mathbb R^2$) any further vector drawn on the page must lie within their span, and hence be a linear combination of them. Note that an intercept term contributes a column of ones to the design matrix, so counts as one of your $k$ columns. (This scenario is often called the "large $p$, small $n$" problem: see also this related CV question.)

Data examples with R code

Each example gives a design matrix $X$, the matrix $X'X$ (note this is always square and symmetrical) and $\det (X'X)$. Note that if $X'X$ is singular (zero determinant, hence not invertible) then we cannot estimate $\hat \beta = (X'X)^{-1}X'y$. The condition that $X'X$ be non-singular is equivalent to the condition that $X$ has full rank so its columns are linearly independent: see this Math SE question, or this one and its converse.

(1) One column is multiple of another

# x2 = 2 * x1
# Note no intercept term (column of 1s) is needed
X <- matrix(c(2, 4, 1, 2, 3, 6, 2, 4), ncol = 2, byrow=TRUE)

X
#     [,1] [,2]
#[1,]    2    4
#[2,]    1    2
#[3,]    3    6
#[4,]    2    4


t(X) %*% X
#     [,1] [,2]
#[1,]   18   36
#[2,]   36   72

round(det(t(X) %*% X), digits = 9)
#0

(2) Intercept term and one variable differs from another by constant

# x1 represents intercept term
# x3 = x2 + 2
X <- matrix(c(1, 2, 4, 1, 1, 3, 1, 3, 5, 1, 0, 2), ncol = 3, byrow=TRUE)

X
#     [,1] [,2] [,3]
#[1,]    1    2    4
#[2,]    1    1    3
#[3,]    1    3    5
#[4,]    1    0    2


t(X) %*% X
#     [,1] [,2] [,3]
#[1,]    4    6   14
#[2,]    6   14   26
#[3,]   14   26   54

round(det(t(X) %*% X), digits = 9)
#0

# NB if we drop the intercept, cols now linearly independent
# x2 = x1 + 2 with no intercept column
X <- matrix(c(2, 4, 1, 3, 3, 5, 0, 2), ncol = 2, byrow=TRUE)

X
#     [,1] [,2]
#[1,]    2    4
#[2,]    1    3
#[3,]    3    5
#[4,]    0    2


t(X) %*% X
#     [,1] [,2]
#[1,]   14   26
#[2,]   26   54
# Can you see how this matrix is related to the previous one, and why?

round(det(t(X) %*% X), digits = 9)
#80
# Non-zero determinant so X'X is invertible

(3) Intercept term and one variable is affine transformation of another

# x1 represents intercept term
# x3 = 2*x2 - 3
X <- matrix(c(1, 2, 1, 1, 1, -1, 1, 3, 3, 1, 0, -3), ncol = 3, byrow=TRUE)

X
#     [,1] [,2] [,3]
#[1,]    1    2    1
#[2,]    1    1   -1
#[3,]    1    3    3
#[4,]    1    0   -3


t(X) %*% X
#     [,1] [,2] [,3]
#[1,]    4    6    0
#[2,]    6   14   10
#[3,]    0   10   20

round(det(t(X) %*% X), digits = 9)
#0

# NB if we drop the intercept, cols now linearly independent
# x2 = 2*x1 - 3 with no intercept column
X <- matrix(c(2, 1, 1, -1, 3, 3, 0, -3), ncol = 2, byrow=TRUE)

X
#     [,1] [,2]
#[1,]    2    1
#[2,]    1   -1
#[3,]    3    3
#[4,]    0   -3


t(X) %*% X
#     [,1] [,2]
#[1,]   14   10
#[2,]   10   20
# Can you see how this matrix is related to the previous one, and why?

round(det(t(X) %*% X), digits = 9)
#180
# Non-zero determinant so X'X is invertible

(4) Intercept term and sum of several variables is fixed

# x1 represents intercept term
# x2 + x3 = 10
X <- matrix(c(1, 2, 8, 1, 1, 9, 1, 3, 7, 1, 0, 10), ncol = 3, byrow=TRUE)

X
#     [,1] [,2] [,3]
#[1,]    1    2    8
#[2,]    1    1    9
#[3,]    1    3    7
#[4,]    1    0   10


t(X) %*% X
#     [,1] [,2] [,3]
#[1,]    4    6   34
#[2,]    6   14   46
#[3,]   34   46  294

round(det(t(X) %*% X), digits = 9)
#0

# NB if we drop the intercept, then columns now linearly independent
# x1 + x2 = 10 with no intercept column
X <- matrix(c(2, 8, 1, 9, 3, 7, 0, 10), ncol = 2, byrow=TRUE)

X
#     [,1] [,2]
#[1,]    2    8
#[2,]    1    9
#[3,]    3    7
#[4,]    0   10

t(X) %*% X
#     [,1] [,2]
#[1,]   14   46
#[2,]   46  294
# Can you see how this matrix is related to the previous one, and why?

round(det(t(X) %*% X), digits = 9)
#2000
# Non-zero determinant so X'X is invertible

(4a) Intercept term with dummy variable trap

# x1 represents intercept term
# x2 + x3 + x4 = 1
X <- matrix(c(1, 0, 0, 1, 1, 1, 0, 0, 1, 0, 1, 0, 1, 1, 0, 0, 1, 0, 1, 0), ncol = 4, byrow=TRUE)

X
#     [,1] [,2] [,3] [,4]
#[1,]    1    0    0    1
#[2,]    1    1    0    0
#[3,]    1    0    1    0
#[4,]    1    1    0    0
#[5,]    1    0    1    0

t(X) %*% X
#     [,1] [,2] [,3] [,4]
#[1,]    5    2    2    1
#[2,]    2    2    0    0
#[3,]    2    0    2    0
#[4,]    1    0    0    1
# This matrix has a very natural interpretation - can you work it out?

round(det(t(X) %*% X), digits = 9)
#0

# NB if we drop the intercept, then columns now linearly independent
# x1 + x2 + x3 = 1 with no intercept column
X <- matrix(c(0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 0), ncol = 3, byrow=TRUE)  

X
#     [,1] [,2] [,3]
#[1,]    0    0    1
#[2,]    1    0    0
#[3,]    0    1    0
#[4,]    1    0    0
#[5,]    0    1    0

t(X) %*% X
#     [,1] [,2] [,3]
#[1,]    2    0    0
#[2,]    0    2    0
#[3,]    0    0    1
# Can you see how this matrix is related to the previous one?

round(det(t(X) %*% X), digits = 9)
#4
# Non-zero determinant so X'X is invertible

(5) Two subsets of variables with fixed sum

# No intercept term needed
# x1 + x2 = 1
# x3 + x4 = 1
X <- matrix(c(0,1,0,1,1,0,0,1,0,1,1,0,1,0,0,1,1,0,1,0,0,1,1,0), ncol = 4, byrow=TRUE)

X
#     [,1] [,2] [,3] [,4]
#[1,]    0    1    0    1
#[2,]    1    0    0    1
#[3,]    0    1    1    0
#[4,]    1    0    0    1
#[5,]    1    0    1    0
#[6,]    0    1    1    0

t(X) %*% X
#     [,1] [,2] [,3] [,4]
#[1,]    3    0    1    2
#[2,]    0    3    2    1
#[3,]    1    2    3    0
#[4,]    2    1    0    3
# This matrix has a very natural interpretation - can you work it out?

round(det(t(X) %*% X), digits = 9)
#0

(6) One variable is linear combination of others

# No intercept term
# x3 = x1 + 2*x2
X <- matrix(c(1,1,3,0,2,4,2,1,4,3,1,5,1,2,5), ncol = 3, byrow=TRUE)

X
#     [,1] [,2] [,3]
#[1,]    1    1    3
#[2,]    0    2    4
#[3,]    2    1    4
#[4,]    3    1    5
#[5,]    1    2    5

t(X) %*% X
#     [,1] [,2] [,3]
#[1,]   15    8   31
#[2,]    8   11   30
#[3,]   31   30   91

round(det(t(X) %*% X), digits = 9)
#0

(7) One variable is constant and zero

# No intercept term
# x3 = 0
X <- matrix(c(1,1,0,0,2,0,2,1,0,3,1,0,1,2,0), ncol = 3, byrow=TRUE)

X
#     [,1] [,2] [,3]
#[1,]    1    1    0
#[2,]    0    2    0
#[3,]    2    1    0
#[4,]    3    1    0
#[5,]    1    2    0

t(X) %*% X
#     [,1] [,2] [,3]
#[1,]   15    8    0
#[2,]    8   11    0
#[3,]    0    0    0

round(det(t(X) %*% X), digits = 9)
#0

(8) Intercept term and one constant variable

# x1 is intercept term, x3 = 5
X <- matrix(c(1,1,5,1,2,5,1,1,5,1,1,5,1,2,5), ncol = 3, byrow=TRUE)

X
#     [,1] [,2] [,3]
#[1,]    1    1    5
#[2,]    1    2    5
#[3,]    1    1    5
#[4,]    1    1    5
#[5,]    1    2    5

t(X) %*% X
#     [,1] [,2] [,3]
#[1,]    5    7   25
#[2,]    7   11   35
#[3,]   25   35  125

round(det(t(X) %*% X), digits = 9)
#0

(9) Two constant variables

# No intercept term, x2 = 2, x3 = 5
X <- matrix(c(1,2,5,2,2,5,1,2,5,1,2,5,2,2,5), ncol = 3, byrow=TRUE)

X
#     [,1] [,2] [,3]
#[1,]    1    2    5
#[2,]    2    2    5
#[3,]    1    2    5
#[4,]    1    2    5
#[5,]    2    2    5

t(X) %*% X
#     [,1] [,2] [,3]
#[1,]   11   14   35
#[2,]   14   20   50
#[3,]   35   50  125

round(det(t(X) %*% X), digits = 9)
#0

(10) $k > n$

# Design matrix has 4 columns but only 3 rows
X <- matrix(c(1,1,1,1,1,2,4,8,1,3,9,27), ncol = 4, byrow=TRUE)

X
#     [,1] [,2] [,3] [,4]
#[1,]    1    1    1    1
#[2,]    1    2    4    8
#[3,]    1    3    9   27

t(X) %*% X
#     [,1] [,2] [,3] [,4]
#[1,]    3    6   14   36
#[2,]    6   14   36   98
#[3,]   14   36   98  276
#[4,]   36   98  276  794

round(det(t(X) %*% X), digits = 9)
#0
$\endgroup$
4
$\begingroup$

Some trivial examples to help intuition:

  1. $\mathbf{x_1}$ is height in centimeters. $\mathbf{x_2}$ is height in meters. Then:
    • $\mathbf{x_1} = 100 \mathbf{x_2}$, and your design matrix $X$ will not have linearly independent columns.
  2. $\mathbf{x_1} = \mathbf{1}$ (i.e. you include a constant in your regression), $\mathbf{x_2}$ is temperature in fahrenheit, and $\mathbf{x_3}$ is temperature in celsius. Then:
    • $\mathbf{x_2} = \frac{9}{5}\mathbf{x_3} + 32 \mathbf{x_1}$, and your design matrix $X$ will not have linearly independent columns.
  3. Everyone starts school at age 5, $\mathbf{x_1} = \mathbf{1}$ (i.e. constant value of 1 across all observations), $\mathbf{x_2}$ is years of schooling, $\mathbf{x_3}$ is age, and no one has left school. Then:
    • $\mathbf{x_2} = \mathbf{x_3} - 5\mathbf{x_1}$, and your design matrix $X$ will not have linearly independent columns.

There a multitude of ways such that one column of data will be a linear function of your other data. Some of them are obvious (eg. meters vs. centimeters) while others can be more subtle (eg. age and years of schooling for younger children).

Notational notes: Let $\mathbf{x_1}$ denote the first column of $X$, $\mathbf{x_2}$ the second column etc..., and $\mathbf{1}$ denotes a vector of ones, which is what's included in the design matrix X if you include a constant in your regression.

$\endgroup$
  • 1
    $\begingroup$ The schooling and age example is very good, though worth pointing out it that the relationship holds only while everyone is still in school! The logical extension of that is when you have age, years of schooling and years of work, which can continue the relationship beyond graduation. (Of course in practice such multicollinearity rarely tends to be perfect - there are always exceptions, like children who started school at a different age because they came from a different country - but it is often quite severe.) $\endgroup$ – Silverfish Jul 5 '16 at 16:16
  • $\begingroup$ @Silverfish good points! I just made some edits/corrections. $\endgroup$ – Matthew Gunn Jul 5 '16 at 16:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.