MichaelChirico - 1 year ago 66

R Question

I'm trying to create a more parsimonious version of this solution, which entails specifying the RHS of a formula in the form

`d1 + d1:d2`

Given that

`*`

`d1 * d2`

`d1 + d2 + d1:d2`

`%+:%`

`"%+:%" <- function(d1,d2) d1 + d2 + d1:d2`

However, this predictably fails because I haven't been careful about evaluation; let's introduce an example to illustrate my progress:

`set.seed(1029)`

v1 <- runif(1000)

v2 <- runif(1000)

y <- .8*(v1 < .3) + .2 * (v2 > .25 & v2 < .8) -

.4 * (v2 > .8) + .1 * (v1 > .3 & v2 > .8)

With this example, hopefully it's clear why simply writing out the two terms might be undesirable:

`y ~ cut(v2, breaks = c(0, .25, .8, 1)) +`

cut(v2, breaks = c(0, .25, .8, 1)):I(v1 < .3)

One workaround which is close to my desired output is to define the whole formula as a function:

`plus.times <- function(outvar, d1, d2){`

as.formula(paste0(quote(outvar), "~", quote(d1),

"+", quote(d1), ":", quote(d2)))

}

This gives the expected coefficients when passed to

`lm`

`d1`

`d2`

`out1 <- lm(y ~ cut(v2, breaks = c(0, .25, .8, 1)) +`

cut(v2, breaks = c(0, .25, .8, 1)):I(v1 < .3))

out2 <- lm(plus.times(y, cut(v2, breaks = c(0, .25, .8, 1)), I(v1 < .3)))

any(out1$coefficients != out2$coefficients)

# [1] FALSE

names(out2$coefficients)

# [1] "(Intercept)" "d1(0.25,0.8]" "d1(0.8,1]" "d1(0,0.25]:d2TRUE"

# [5] "d1(0.25,0.8]:d2TRUE" "d1(0.8,1]:d2TRUE"

So this is less than optimal.

Is there any way to define the adjust the code so that the infix operator I mentioned above works as expected? How about altering the form of

`plus.times`

I've been poking around (

`?formula`

`?"~"`

`?":"`

`getAnywhere(formula.default)`

`*`

Answer Source

You do not need to define a new operator in this case: in a formula `d1/d2`

expands to `d1 + d1:d2`

. In other words `d1/d2`

specifies that `d2`

is nested within `d1`

. Continuing your example:

```
out3 <- lm(y ~ cut(v2,breaks=c(0,.25,.8,1))/I(v1 < .3))
all.equal(coef(out1), coef(out3))
# [1] TRUE
```

*Further comments*

Factors may be *crossed* or *nested*. Two factors are crossed if it possible to observe every combination of levels of the two factors, e.g. sex and treatment, temperature and pH, etc. A factor is nested within another if each level of that factor can only be observed within one of the levels of the other factor, e.g. town and country, staff member and store etc.

These relationships are reflected in the parametrization of the model. For crossed factors we use `d1*d2`

or `d1 + d2 + d1:d2`

, to give the main effect of each factor, plus the interaction. For nested factors we use `d1/d2`

or `d1 + d1:d2`

to give a separate submodel of the form `1 + d2`

for each level of `d1`

.

The idea of nesting is not restricted to factors, for example we may use `sex/x`

to fit a separate linear regression on `x`

for males and females.

In a formula, `%in%`

is equivalent to `:`

, but it may be used to emphasize the nested, or hierarchical structure of the data/model. For example, `a + b %in% a`

is the same as `a + a:b`

, but reading it as "a plus b within a" gives a better description of the model being fitted. Even so, using `/`

has the advantage of simplifying the model formula at the same time as emphasizing the structure.