Rはlmの欠損値をどのように処理しますか？

行列Aの各列に対してベクトルBを回帰させたいと思います。欠損データがない場合、これは簡単ですが、行列Aに欠損値が含まれている場合、Aに対する回帰はすべての行のみを含むように制限されます値が存在します（デフォルトのna.omit動作）。これにより、欠落データのない列に対して誤った結果が生成されます。列行列Bを行列Aの個々の列に対して回帰することはできますが、何千回も回帰を行う必要があり、これは非常に遅く、洗練されていません。na.exclude機能は、このような場合のために設計されているように見えるが、私はそれを動作させることはできません。ここで何が間違っていますか？重要な場合、OSXでR 2.13を使用します。

A = matrix(1:20, nrow=10, ncol=2)
B = matrix(1:10, nrow=10, ncol=1)
dim(lm(A~B)$residuals)
# [1] 10 2 (the expected 10 residual values)

# Missing value in first column; now we have 9 residuals
A[1,1] = NA  
dim(lm(A~B)$residuals)
#[1]  9 2 (the expected 9 residuals, given na.omit() is the default)

# Call lm with na.exclude; still have 9 residuals
dim(lm(A~B, na.action=na.exclude)$residuals)
#[1]  9 2 (was hoping to get a 10x2 matrix with a missing value here)

A.ex = na.exclude(A)
dim(lm(A.ex~B)$residuals)
# Throws an error because dim(A.ex)==9,2
#Error in model.frame.default(formula = A.ex ~ B, drop.unused.levels = TRUE) : 
#  variable lengths differ (found for 'B')

編集：あなたの質問を誤解しました。2つの側面があります。

a) na.omit and na.exclude both do casewise deletion with respect to both predictors and criterions. They only differ in that extractor functions like residuals() or fitted() will pad their output with NAs for the omitted cases with na.exclude, thus having an output of the same length as the input variables.

> N    <- 20                               # generate some data
> y1   <- rnorm(N, 175, 7)                 # criterion 1
> y2   <- rnorm(N,  30, 8)                 # criterion 2
> x    <- 0.5*y1 - 0.3*y2 + rnorm(N, 0, 3) # predictor
> y1[c(1, 3,  5)] <- NA                    # some NA values
> y2[c(7, 9, 11)] <- NA                    # some other NA values
> Y    <- cbind(y1, y2)                    # matrix for multivariate regression
> fitO <- lm(Y ~ x, na.action=na.omit)     # fit with na.omit
> dim(residuals(fitO))                     # use extractor function
[1] 14  2

> fitE <- lm(Y ~ x, na.action=na.exclude)  # fit with na.exclude
> dim(residuals(fitE))                     # use extractor function -> = N
[1] 20  2

> dim(fitE$residuals)                      # access residuals directly
[1] 14  2

b) The real issue is not with this difference between na.omit and na.exclude, you don't seem to want casewise deletion that takes criterion variables into account, which both do.

> X <- model.matrix(fitE)                  # design matrix
> dim(X)                                   # casewise deletion -> only 14 complete cases
[1] 14  2

The regression results depend on the matrices $X^{+} = (X' X)^{-1} X'$ (pseudoinverse of design matrix $X$, coefficients $\hat{\beta} = X^{+} Y$) and the hat matrix $H = X X^{+}$, fitted values $\hat{Y} = H Y$). If you don't want casewise deletion, you need a different design matrix $X$ for each column of $Y$, so there's no way around fitting separate regressions for each criterion. You can try to avoid the overhead of lm() by doing something along the lines of the following:

> Xf <- model.matrix(~ x)                    # full design matrix (all cases)
# function: manually calculate coefficients and fitted values for single criterion y
> getFit <- function(y) {
+     idx   <- !is.na(y)                     # throw away NAs
+     Xsvd  <- svd(Xf[idx , ])               # SVD decomposition of X
+     # get X+ but note: there might be better ways
+     Xplus <- tcrossprod(Xsvd$v %*% diag(Xsvd$d^(-2)) %*% t(Xsvd$v), Xf[idx, ])
+     list(coefs=(Xplus %*% y[idx]), yhat=(Xf[idx, ] %*% Xplus %*% y[idx]))
+ }

> res <- apply(Y, 2, getFit)    # get fits for each column of Y
> res$y1$coefs
                   [,1]
(Intercept) 113.9398761
x             0.7601234

> res$y2$coefs
                 [,1]
(Intercept) 91.580505
x           -0.805897

> coefficients(lm(y1 ~ x))      # compare with separate results from lm()
(Intercept)           x 
113.9398761   0.7601234 

> coefficients(lm(y2 ~ x))
(Intercept)           x 
  91.580505   -0.805897

Note that there might be numerically better ways to caculate $X^{+}$ and $H$, you could check a $QR$-decomposition instead. The SVD-approach is explained here on SE. I have not timed the above approach with big matrices $Y$ against actually using lm().

I can think of two ways. One is combine the data use the na.exclude and then separate data again:

A = matrix(1:20, nrow=10, ncol=2)
colnames(A) <- paste("A",1:ncol(A),sep="")

B = matrix(1:10, nrow=10, ncol=1)
colnames(B) <- paste("B",1:ncol(B),sep="")

C <- cbind(A,B)

C[1,1] <- NA
C.ex <- na.exclude(C)

A.ex <- C[,colnames(A)]
B.ex <- C[,colnames(B)]

lm(A.ex~B.ex)

Another way is to use the data argument and create a formula.

Cd <- data.frame(C)
fr <- formula(paste("cbind(",paste(colnames(A),collapse=","),")~",paste(colnames(B),collapse="+"),sep=""))

lm(fr,data=Cd)

Cd[1,1] <-NA

lm(fr,data=Cd,na.action=na.exclude)

If you are doing a lot of regression the first way should be faster, since less background magic is performed. Although if you need only coefficients and residuals I suggest using lsfit, which is much faster than lm. The second way is a bit nicer, but on my laptop trying to do summary on the resulting regression throws an error. I will try to see whether this is a bug.

The following example shows how to make predictions and residuals that conform to the original dataframe (using the "na.action=na.exclude" option in lm() to specify that NA's should be placed in the residual and prediction vectors where the original dataframe had missing values. It also shows how to specify whether predictions should include only observations where both explanatory and dependent variables were complete (i.e., strictly in-sample predictions) or observations where the explanatory variables were complete, and hence Xb prediction is possible, (i.e., including out-of-sample prediction for observations that had complete explanatory variables but were missing the dependent variable).

I use cbind to add the predicted and residual variables to the original dataset.

## Set up data with a linear model
N <- 10
NXmissing <- 2 
X <- runif(N, 0, 10)
Y <- 6 + 2*X + rnorm(N, 0, 1)
## Put in missing values (missing X, missing Y, missing both)
X[ sample(1:N , NXmissing) ] <- NA
Y[ sample(which(is.na(X)), 1)]  <- NA
Y[ sample(which(!is.na(X)), 1)]  <- NA
(my.df <- data.frame(X,Y))

## Run the regression with na.action specified to na.exclude
## This puts NA's in the residual and prediction vectors
my.lm  <- lm( Y ~ X, na.action=na.exclude, data=my.df)

## Predict outcome for observations with complete both explanatory and
## outcome variables, i.e. observations included in the regression
my.predict.insample  <- predict(my.lm)

## Predict outcome for observations with complete explanatory
## variables.  The newdata= option specifies the dataset on which
## to apply the coefficients
my.predict.inandout  <- predict(my.lm,newdata=my.df)

## Predict residuals 
my.residuals  <- residuals(my.lm)

## Make sure that it binds correctly
(my.new.df  <- cbind(my.df,my.predict.insample,my.predict.inandout,my.residuals))

## or in one fell swoop

(my.new.df  <- cbind(my.df,yhat=predict(my.lm),yhato=predict(my.lm,newdata=my.df),uhat=residuals(my.lm)))