---
title: "lm function"
date: today
date-format: long
bibliography: ../inst/REFERENCES.bib
vignette: >
%\VignetteIndexEntry{lm function}
%\VignetteEngine{quarto::pdf}
%\VignetteEncoding{UTF-8}
---
**OLS** is by far the most important estimation method used in
econometrics. It is implemented in **R** by the `lm` function. A deep
understanding of the way this function is implemented, its arguments,
its results and how to use them is particularly useful as other
estimation functions (for example `glm` for generalized linear models)
have a lot in common with `lm`. Moreover, if one is willing to
implement its own estimation function, it is advisable to use at least
some of the elements of `lm`.
For illustration, we use the `random_group` data set of the **micsr**
package that analyzes the effect of a
training on wage.
```{r }
#| echo: true
#| message: false
library(micsr)
random_group |> head()
```
The response is hourly wage `wage` and
the main covariate is a dummy for training. The `group` variable
contains integers from -2 to 3, with negative values for treated
individuals and positive values for untreated individuals. We first
create a dummy for treatment, called `treated`:
```{r }
random_group <- transform(random_group, treated = ifelse(group < 0, 1, 0))
```
## A basic model
We then estimate the model, using as control the number of children
`child` and the `age`.
```{r }
mod0 <- lm(log(wage) ~ treated + age + child, data = random_group)
```
The result is a an object of class `lm`, with the 12 following elements:
```{r }
names(mod0)
```
We won't say much about `"qr"` and `"effects"` which are technical
elements of the computation of least squares. `"coefficients"`,
`"residuals"` and `"fitted.values"` don't deserve further
comments. The length of `"residuals"` and `"fitted.values"` is $N$,
the one of `"coefficients"` is $K+1$ (if there is an
intercept). `"rank"` is the rank of the model matrix. If the matrix
has full rank, the rank equals the number of columns of the model
matrix ($K + 1$), which is also the number of fitted
coefficients.^[Actually the number of not `NA` coefficients.]
```{r }
#| collapse: true
mod0$rank
length(mod0$coefficients)
```
`"df.residual"` is then $N - K - 1$. `"terms"` is a term
object, which is a kind of representation of the formula of the model:
```{r }
mod0$terms
```
`"xlevels"` is for this model a useless list of length 0. We'll see
latter the usefulness of this element. `assign` is a numeric vector
indicating the relation between the variables and the coefficients:
```{r }
#| collapse: true
mod0$assign
```
Once again, in our simple case, it is useless: the intercept is at
position 0, the first covariate at position 1 and so on.
All the elements of the model can be extracted using the `$` operator,
for example `mod0$coefficients` for the coefficients. A safer method
to extract elements of the model is to use generic functions with
homonym names:
```{r }
#| collapse: true
coef(mod0)
df.residual(mod0)
resid(mod0) |> head()
fitted(mod0) |> head()
```
```{r }
#| results: false
terms(mod0)
```
Several other generics enables to extract information of the model
that are not elements of the results, for example the covariance
matrix of the estimates, the residual standard error and the number of
observations:
```{r }
#| collapse: true
vcov(mod0)
sigma(mod0)
nobs(mod0)
```
The next two sections are devoted to the last two elements of the `lm`
object, `"model"` and `"call"`.
## Model frame
The element `"model"` of the `lm` object is the model frame. It's a
special data frame:
```{r }
head(mod0$model, 3)
```
Instead of using the `$` operator, the `model.frame` function can be
used to extract the model frame:
```{r }
#| results: false
model.frame(mod0)
```
For this basic model, it looks like a subset of columns of the
original data frame, but note that the first column contains the response
which is `log(wage)` and not `wage`. A model frame has several attributes:
```{r }
#| collapse: true
names(attributes(mod0$model))
```
The only interesting attribute is `"terms"`, the two other
being the rows and columns' names (like for ordinary data frames).
From the model frame, the elements of the model can be extracted. For
this basic model, there are two elements, the matrix of covariates $X$
and the vector of response $y$. To extract the model matrix, we use
`model.matrix` with two arguments, the model terms and the model
frame:
```{r }
model.matrix(terms(mod0), model.frame(mod0)) |> head(3)
```
More simply, the same model matrix can be extracted using the fitted
object as the unique argument:
```{r }
#| results: hide
#| collapse: true
model.matrix(mod0)
```
The model matrix looks like the model frame, except that the response
has been removed and a column of ones (called `"(Intercept)"`) has been
added. Actually there is a fundamental difference between the model
frame and the model matrix, even in this very simple example. A model
frame is a data frame, which is a list of vectors of the same
length. Because all the vectors have the same length, it has a tabular
representation and therefore looks like a matrix. A model matrix is a
matrix, which is for **R** a vector with an attribute indicating the
dimension, ie., the number of rows and of columns.
```{r }
dim(model.matrix(mod0))
```
Being a vector, a matrix contains only elements of the same mode, for
example numeric.
The model response is extracted using `model.response` with the model
frame as unique argument:
```{r }
#| collapse: true
mod0 |> model.frame() |> model.response() |> head()
```
## Call and update
`"call"` is an object of class `"call"`, that corresponds to
the call to the function that resulted in the `mod0` object:
```{r }
#| collapse: true
class(mod0$call)
mod0$call
```
Note that all the arguments are named, although we didn't name the first
argument (`formula`) while calling `"lm"`. The call is very useful as
it enables to update the model. The default `update` function is:
`update(object, formula., ..., evaluate = TRUE)`. Its first argument is
a fitted model and the second one is a
formula. For example, to update `mod0` by adding the `female`,
`single`, `migrant` and `temp` covariates (dummies for females,
singles, migrants and temporary jobs), instead of rewriting the whole
command:
```{r }
mod1 <- lm(log(wage) ~ treated + age + child + female + single + migrant +
temp, data = random_group)
```
we can update `mod0` with a new formula:
```{r }
mod1 <- mod0 |> update(log(wage) ~ treated + age + child +
female + single + migrant + temp)
```
or even more simply, we can use dots in the formula, which means "as
before" and enables to add or to remove variables from the initial
formula:
```{r }
mod1 <- update(mod0, . ~ . + female + single + migrant + temp)
```
We have seen previously that the default `update` function has a `...`
argument. This means that we can use in a call to update an argument
of `lm` which is then passed to `lm` while updating the model. As an
example, `lm` has a `subset` argument which is a logical expression
that select a subset of the sample for which the estimation has to be
performed. For example, starting from `mod0`, if one wishes to add the
four dummies as previously and also to select the individuals aged at
least 20, we can use:
```{r }
#| eval: false
lm(log(wage) ~ treated + age + child + female + single + migrant + temp,
data = random_group, subset = age >= 20)
```
Or much more simply:
```{r }
#| eval: false
update(mod0, . ~ . + female + single + migrant + temp, subset = age >= 20)
```
## Missing values
When values of some variables used in the regression are missing for
some observations, one has to indicate how to deal with this
problem. The default behavior corresponds to the `na.action` element
of `options()` which is a list containing options:
```{r }
#| collapse: true
options()$na.action
```
`na.omit` means remove from the model frame all the lines for which
there is at least a missing value. There are a couple of other
possible values, one of them being `na.fail` which stops the
estimation and returns an error. `lm` has a `na.action` argument where
the desired action can be indicated. For example, if we update the
model with `na.action` set to `"na.fail"`:
```{r }
#| results: false
mod1 |> update(na.action = "na.fail")
```
we get exactly the same results as there are no missing values for
the subset of variables we used. Now consider adding the `ten`
covariate which is the tenure in month and has some missing
values. Setting as previously `na.action = "na.fail"`, we'll then get
an error message:
```{r }
#| eval: false
mod1 |> update(. ~ . + ten, na.action = "na.fail")
```
Using the default `"na.omit"` value, we get:
```{r }
mod2 <- mod1 |> update(. ~ . + ten)
names(mod2)
```
and there is now a 13^th^ element named `na.action`:
```{r }
unname(mod2$na.action)
```
This is an object of class `"omit"` containing a vector of integers
indicating the positions of the lines that have been removed because
of missing values. We used `unname` because actually this is a named
vector and, as in our case, when there is no row names, the "name" is
the number (so that the name of the first element of the vector, 61,
is "61"). This object is also returned as an attribute of the model frame:
```{r }
#| collapse: true
mod2$model |> attributes() |> names()
```
# Factors
Now we introduce two further covariates `fsize` (firm size) and `edu`
(education). These covariates are **factors** (categorical variables) and
the modalities are called **levels**. The modalities can be extracted
using the `levels` function:
```{r }
#| collapse: true
levels(random_group$fsize)
levels(random_group$edu)
```
We introduce these two factors in the regression and we also introduce
a quadratic term for `age`. This is done using the `poly` function
inside the formula:
```{r }
mod3 <- lm(log(wage) ~ treated + poly(age, 2) + child + fsize + edu +
female + single + migrant + temp + ten,
data = random_group)
```
There is now a supplementary element called `"contrasts"`. But first
consider the already existing `"assign"` and `"xlevels"` elements:
```{r }
mod3$assign
```
`"assign"` is now a vector of length 14. The first element is 0 (for
the intercept), the 13 remaining indicate the link between the
position of the coefficients and the position of the covariate in the
formula:
```{r }
names(coef(mod3))
```
for example, the 3^rd^ and 4^th^ values of `"assign"` are 2, which is the
position of the `"age"` covariate, with now two coefficients. The
8^th^ and 9^th^ values of `"assign"` are 5, they correspond to the
two dummies introduced in the regression for the 5^th^ covariate in
the formula which is `edu`. Let's now have a look to the `"xlevels"`
element which used to be in the previous example a list of length 0:
```{r }
mod3$xlevels
```
This is now a named list containing character vectors indicating the
different levels of the factors. Note that for a factor with $J$
levels, only $J-1$ dummies are introduced in the regression, the one
corresponding to the first level being omitted.
The `contrasts` element is:
```{r }
mod3$contrasts
```
It's a named list indicating how the contrasts are computed from the
factors. The default value is `"contr.treatment"` and consist, as we
seen previously, to create $J-1$ dummies for all the levels of the
factor but the first. Other values are possible and can be set
individually for every factor using the `contrasts` argument of `lm`:
```{r }
mod4 <- update(mod3, contrasts = list(edu = "contr.sum",
fsize = "contr.helmert"))
mod4$contrasts
```
The model frame now looks like:
```{r }
head(mod4$model, 3)
```
The `age` column has been replaced by a $N \times 2$ matrix containing
the two terms of the polynomial. The two factors are in their own
column.
The model matrix $X$ is now:
```{r }
head(model.matrix(mod3), 3)
head(model.matrix(mod4), 3)
```
The fundamental difference between the model frame (a list) and the
model matrix (a vector with a dimension) is even clearer than in the
previous simple example. The model frame contains a column called
`"edu"` which is a factor. In the model matrix, this column is
replaced by $J-1=2$ columns that contain numeric and we can see that
the numerical coding of the variable depends on the contrast used.
## Weights and offset
Instead of minimizing the sum of squares residuals $\sum_n (y_n -
\beta ^ \top x_n) ^ 2$, **weights** $w_n$ can be used so that the objective
function becomes $\sum_n w_n (y_n - \beta ^ \top x_n) ^ 2$, leading to the
weighted least squares estimator. Weights can be indicating in `lm`
using the `weights` argument, which is the unquoted name of the
column of the data frame that contains the weights. For the
`random_group` data set, the variable that contains the sample weights
is `samplew`:
```{r }
mod4 <- update(mod3, weights = samplew)
```
The `lm` object now has a 15^th^ element called `"weights"` which
contains the vector of weights of length $N$. The model frame is now:
```{r }
model.frame(mod4) |> head(3)
```
Note the last column called `"(weights)"` that contains the
weights. The weights can be extracted from the model frame using
`model.weights`:
```{r }
model.weights(model.frame(mod4)) |> head(3)
```
An **offset** is a variable added to the regression equation, but with no
associated coefficients (or with a coefficient set to 1). For example, in
the equation: $y_n = \alpha + \beta_1 x_{n1} + \beta_2 x_{n2} + x_{n3} +
\epsilon_n$, $x_{n3}$ is an offset. Note also that the same equation can
be rewritten as: $y_n - x_{n3} = \alpha + \beta_1 x_{n1} + \beta_2
x_{n2} + \epsilon_n$ and that in this case, the new response is $y_n -
x_{n3}$ and there is no offset anymore. We have seen in `mod3` that
the coefficients for the `"Intermediate"` and `"High"` education
values are respectively $0.23$ and $0.50$, which means approximately
25 and 50% more wage compared to the reference level of this factor
which is `"Low"`. Coercing the three levels of the factor to $z = 1,2,3$
(which is actually the internal representation of the factor) and
denoting $v = 3/4 + 1/4 z$, we get the values of 1, 1.25 and
1.50. Therefore, we could get approximately the same results as
`mod3` by removing `"edu"` and adding $v$ as an offset. There are two
equivalent ways to introduce an offset: in the formula, using
`offset(v)` or by setting the `offset` argument of `lm` to `v`:
```{r }
random_group <- transform(random_group, v = 3/4 + 1/4 * as.numeric(edu))
mod5 <- update(mod4, . ~ . + offset(v) - edu)
mod6 <- update(mod4, . ~ . - edu, offset = v)
```
The two fitted models are identical, they contain a supplementary
element called `"offset"`. The offset can be extract from the model
frame using `model.offset`:
```{r }
model.offset(model.frame(mod5)) |> head(3)
```
## The internal
To conclude the presentation of the `lm` function, we'll describe the
internal of the function. We first create a simple `mylm` function
with the same argument as `lm`, but which returns only the call,
obtained using the function `match.call`:
```{r }
mylm <- function (formula, data, subset, weights, na.action, method = "qr",
model = TRUE, x = FALSE, y = FALSE, qr = TRUE,
singular.ok = TRUE, contrasts = NULL, offset, ...){
mf <- match.call()
mf
}
```
We then use this function with our most advanced example:
```{r }
mf <- mylm(log(wage) ~ treated + poly(age, 2) + child + fsize + female +
single + migrant + temp + ten,
random_group,
subset = age > 20,
weights = samplew,
contrasts = list(fsize = "contr.helmert"),
offset = v
)
mf
cl <- mf
```
We saved a copy of the result as we'll need it latter on.
The next lines of `lm` work on the call (the `mf` object returned by
our `mylm` function). The call being a list, the idea is first to keep
only the arguments in the call that are useful to compute the model
frame. This is done using the `match` function on the names of the
call:
```{r }
names(mf)
m <- match(c("formula", "data", "subset", "weights", "na.action",
"offset"), names(mf), 0L)
m
```
`m` is an integer vector indicating the position of the arguments we
wish to keep in the call; note that the position of `na.action` is 0
because we didn't use this argument in our call to `mylm`. The first
element of `mf` is unnamed, this is the name of the function
(`mylm`). We then extract from `mf` its first element and
those given by vector `m`:
```{r }
mf <- mf[c(1L, m)]
mf
```
Then, we change the first argument (the name of the function) from
`mylm` to `model.frame`, using the `quote` function:
```{r }
mf[[1L]] <- quote(stats::model.frame)
mf
```
The call is then evaluated using the `eval` function and the result is
the model frame:
```{r }
mf <- eval(mf)
head(mf, 3)
names(attributes(mf))
```
We then extract the terms, and the different components of the model:
```{r }
mt <- attr(mf, "terms")
y <- model.response(mf)
w <- model.weights(mf)
offset <- model.offset(mf)
x <- model.matrix(mt, mf, list(fsize = "contr.helmert"))
```
Note that the model matrix is constructed using as third argument the
`contrasts` argument of `lm`.
The estimation is then performed by `lm.wfit` (if there were no
weights, `lm.fit` would have been used), that take as arguments the
components of the model previously extracted:
```{r }
z <- lm.wfit(x, y, w, offset = offset)
names(z)
```
We can see that the resulting object contains a subset of the elements
returned by `lm`. We then assign to the object the `"lm"` class:
```{r }
class(z) <- "lm"
```
And we add to `z` the elements that are not returned by `lm.wfit`:
```{r }
z$na.action <- attr(mf, "na.action")
z$offset <- offset
z$contrasts <- attr(x, "contrasts")
z$xlevels <- .getXlevels(mt, mf)
z$call <- cl
z$terms <- mt
z$model <- mf
```
`na.action` is an attribute of the model frame, `contrasts` an
attribute of the model matrix and `xlevels` is obtained using the
`.getXlevels` function with the terms and the model frame as arguments.
We then get our `lm` object that contains the fitted model:
```{r }
z
```
Finally, we include in our `mylm` function all the lines we just have described:
```{r }
mylm <- function (formula, data, subset, weights, na.action, method = "qr",
model = TRUE, x = FALSE, y = FALSE, qr = TRUE,
singular.ok = TRUE, contrasts = NULL, offset, ...){
mf <- match.call(expand.dots = FALSE)
m <- match(c("formula", "data", "subset", "weights", "na.action",
"offset"), names(mf), 0L)
mf <- mf[c(1L, m)]
mf[[1L]] <- quote(stats::model.frame)
mf <- eval(mf, parent.frame())
mt <- attr(mf, "terms")
y <- model.response(mf, "numeric")
w <- as.vector(model.weights(mf))
offset <- model.offset(mf)
x <- model.matrix(mt, mf, contrasts)
z <- lm.wfit(x, y, w, offset = offset)
class(z) <- "lm"
z$na.action <- attr(mf, "na.action")
z$offset <- offset
z$contrasts <- attr(x, "contrasts")
z$xlevels <- .getXlevels(mt, mf)
z$call <- cl
z$terms <- mt
z$model <- mf
z
}
```
to get a simplified clone of `lm`:
```{r }
mod6 <- mylm(log(wage) ~ treated + poly(age, 2) + child + fsize + female +
single + migrant + temp + ten,
random_group,
subset = age > 20,
weights = samplew,
contrasts = list(fsize = "contr.helmert"),
offset = v)
```