Is there an easy way in R to create a linear regression over a model with 100 parameters in R?
Let's say we have a vector Y with 10 values and a dataframe X with 10 columns and 100 rows
In mathematical notation I would write Y = X[[1]] + X[[2]] + ... + X[[100]]
.
How do I write something similar in R syntax?
-
1$\begingroup$ are there 100 or 1000? Also, you would normally have the columns be the variables and the rows be observations (it appears that is reversed here) $\endgroup$– MacroMay 30, 2012 at 13:09
-
$\begingroup$ 100 the extra 0 was a typo $\endgroup$– ChristianMay 30, 2012 at 13:11
-
2$\begingroup$ Really? Are you sure you want to do this? I'd be concerned about overfitting and correlation between linear combinations of the predictors. Not only that, with 100 predictors but only 10 observations, you have $p>n$ and linear regression isn't going to work at all. $\endgroup$– Aaron left Stack OverflowMay 31, 2012 at 1:47
4 Answers
Try this
df<-data.frame(y=rnorm(10),x1=rnorm(10),x2=rnorm(10))
lm(y~.,df)
-
4$\begingroup$ And if you want, say, all second-order interactions, you can write
y ~ . + .^2
. And so on. $\endgroup$ Mar 12, 2015 at 14:46 -
3$\begingroup$ And if you want only some of the second order interactions, something like
y ~ . + .:x1
will get you the interactions of each variable (exceptx1
) withx1
. And so on; you get the idea. $\endgroup$ Mar 12, 2015 at 14:48
Great answers!
I would add that by default, calling formula
on a data.frame
creates an additive formula to regress the first column onto the others.
So in the case of the answer of @danas.zuokas you can even do
lm(df)
which is interpreted correctly.
-
$\begingroup$ Still, this answer does not work if you want to mix in interaction terms. Yours does (+1). $\endgroup$ May 30, 2012 at 13:36
-
6$\begingroup$ I'm continually amazed at how overloaded most of
R
's operators are :) $\endgroup$– MacroMay 30, 2012 at 13:38
If each row is an observation and each column is a predictor so that $Y$ is an $n$-length vector and $X$ is an $n \times p$ matrix ($p=100$ in this case), then you can do this with
Z = as.data.frame(cbind(Y,X))
lm(Y ~ .,data=Z)
If there are other columns you did not want to include as predictors, you would have to remove them from X
before using this trick, or using -
in the model formula to exclude them. For example, if you wanted to exclude the 67th predictor (that has the corresponding name x67
), then you could write
lm(Y ~ .-x67,data=Z)
Also, if you want to include interactions, etc.. you will need to add them manually as (for example)
lm(Y ~ .+X[,1]*X[,2],data=Z)
or make sure they are entered as columns of X
.
You can also use a combination of the formula
and paste
functions.
Setup data: Let's imagine we have a data.frame that contains the predictor variables x1
to x100
and our dependent variable y
, but that there is also a nuisance variable asdfasdf
. Also the predictor variables are arranged in an order such that they are not all contiguous in the data.frame.
Data <- data.frame(matrix(rnorm(102 * 200), ncol=102))
names(Data) <- c(paste("x", 1:50, sep=""),
"asdfasdf", "y", paste("x", 51:100, sep=""))
Imagine also that you have a string containing the names of the predictor variables. In this case, this can easily be created using the paste
function, but in other situations, grep
or some other approach might be used to get this string.
PredictorVariables <- paste("x", 1:100, sep="")
Apply approach: We can then construct a formula as follows:
Formula <- formula(paste("y ~ ",
paste(PredictorVariables, collapse=" + ")))
lm(Formula, Data)
- the
collapse
argument inserts+
between the predictor variables formula
converts the string into an object of class formula suitable for thelm
function.
More generally, I use the following function quite regularly when I want to supply a the predictor variables as vector of variable names.
regression <- function(dv, ivs, data) {
# run a linear model with text arguments for dv and ivs
iv_string <- paste(ivs, collapse=" + ")
regression_formula <- as.formula(paste(dv, iv_string, sep=" ~ "))
lm(regression_formula, data)
}
E.g.,
regression("y", PredictorVariables, Data)
-
2$\begingroup$ +1. I use this technique all the time. Occasionally, however, having the formula stored in a variable causes issues. See stackoverflow.com/a/7668846/210673 for use of
do.call
evaluate the formula before callinglm
. $\endgroup$ May 31, 2012 at 1:44