# Is there any way the adjusted $R^2$ might decrease by adding predictors?

Let's consider a multiple linear regression formula:

$\hat{y} = \beta_0 + \beta_1 \hat{x}_1 + \beta_2 \hat{x}_2$ (1)

which produces adjusted $R^2 = r_1$.

Now I want to add to one predictor to the (1) which turns into:

$\hat{y} = \beta_0 + \beta_1 \hat{x}_1 + \beta_2 \hat{x}_2 + \beta_3 \hat{x}_3$ (2)

which produces adjusted $R^2 = r_2$.

If the data fed into (1) and (2) are exactly the same, is there any way to explain $r_2 < r_1$ apart from a code bug?

• i think you might be confusing {R-sq} and {adjusted R-sq} here. Feb 19, 2014 at 2:02
• @charles indeed Feb 19, 2014 at 2:31
• The very large number of threads on our site discussing adjusted R-squared are worth reviewing in this context.
– whuber
Feb 19, 2014 at 15:24

## 1 Answer

Yes, it's definitely possible for adjusted $R^2$ to decrease when you add parameters.

Ordinary $R^2$ can't decrease, but adjusted-$R^2$ certainly can. We can write the relationship between the two like so:

$R_{adj}^2 = R^2-(1-R^2)\frac{p}{n-p-1}$

Note that both terms in the product $(1-R^2)\cdot\frac{p}{n-p-1}$ are positive (unless $R^2=1$), so if $R^2<1$, $R_{adj}^2 < R^2$.

If $R^2<\frac{p}{n-1}$, then adjusted-$R^2$ will be negative.

$R_{adj}^2$ will decrease if the $R^2$ for a model with an additional term if the second model's $R^2$ didn't increase from that for the first model by at least as much as would be expected for an unrelated variable.

We can see this happen quite easily: I just generated three unrelated variables in R (via)
x1=rnorm(20);x2=rnorm(20);y=rnorm(20)

1. If we fit a linear regression with just the first $x$ (lm(y~x1)), the adjusted $R^2$ is smaller than with the null model (which is 0):
Multiple R-squared: 0.0007048, Adjusted R-squared: -0.05481

2. If we fit both independent variables (lm(y~x1+x2)), the adjusted $R^2$ goes down again (and the $R^2$ – necessarily – goes up):
Multiple R-squared: 0.00199, Adjusted R-squared: -0.1154

For adjusted $R^2$ to increase, its addition has to explain more additional variation in the data than would be expected from an unrelated variable; it's possible for an unrelated variable to add less than it would be expected to, just by chance.