原始数据在来自小鼠的汇总评估中的作用（R包）？

Question

I'm wondering what is the role of the original data set when using the mice package in R for imputed data. I need to impute my data and then compute some additional variables before turning the long data set back into an as.mids object. I noticed that when computing my additional variable ("total" in the code below) that whether I used na.rm=TRUE affected by estimates and from my understanding, it shouldn't. Here's a reproducible example:

# Add required package 
 require(mice)

# Impute data and compute summary with na.rm=T 
 imp1 <- mice(nhanes, seed = 123) 
 com1 <- complete(imp1, "long", include = TRUE) 
 head(com1) 
 com1$total <- rowSums(com1[4:6],na.rm=T)
 imp2 <- as.mids(com1)

# Fit model with data using na.rm=T 
 fit <- with(imp2, lm(bmi ~ age)) 
 round(summary(pool(fit)), 2)

Notice that my variable "total" is the rowSums of 3 variables and I've used na.rm=TRUE . However, as only the original data set (denoted by the variable ".imp" in the long data set contains NA values, this extra bit of code should only be relevant for the original data. Removing na.rm=TRUE shows that this is not true:

# Impute data and compute summary without na.rm=T 
 imp3 <- mice(nhanes, seed = 123) 
 com2 <- complete(imp3, "long", include = TRUE) 
 head(com2) 
 com2$total <- rowSums(com2[4:6]) 
 imp4 <- as.mids(com2)

# Fit model with data without using na.rm=T 
fit2 <- with(imp4, lm(bmi ~ age)) 
round(summary(pool(fit2)), 2)

Again, notice that leaving out na.rm=TRUE leads to different estimates. The only difference here is that the variable "total" now has NA values when the variable .imp is equal to zero (ie, the original data set).

What am I missing? I would have thought that only the imputed data would have affected the pooled estimates, while I just showed that values in the original data set do (ie, those from .imp = 0). What is the role of the original data set in getting pooled estimates from mice?

NOTE: EDITED FOR CLARITY

Answer 1

I would imagine that the original (raw) data plays no role. According to the as.mids help page it is only needed to signify where the missing data is. I ran your script and noticed there was an error when creating imp2 . You call on the object com which should be com1 . After correction get the exact same results for the two approaches:

# Add required package 
require(mice)

# Impute data and compute summary with na.rm=T 
imp1 <- mice(nhanes, seed = 123) 
com1 <- complete(imp1, "long", include = TRUE) 
head(com1) 
com1$total <- rowSums(com1[4:6],na.rm=T)
imp2 <- as.mids(com1)

# Fit model with data using na.rm=T 
fit <- with(imp2, lm(bmi ~ age)) 

# Impute data and compute summary without na.rm=T 
imp3 <- mice(nhanes, seed = 123) 
com2 <- complete(imp3, "long", include = TRUE) 
head(com2) 
com2$total <- rowSums(com2[4:6]) 
imp4 <- as.mids(com2)

# Fit model with data without using na.rm=T 
fit2 <- with(imp4, lm(bmi ~ age)) 

The results:

> round(summary(pool(fit)), 2)
            estimate std.error statistic    df p.value
(Intercept)    29.76      1.86     15.98 18.61    0.00
age            -1.73      0.95     -1.83 19.50    0.08

> round(summary(pool(fit2)), 2)
            estimate std.error statistic    df p.value
(Intercept)    29.76      1.86     15.98 18.61    0.00
age            -1.73      0.95     -1.83 19.50    0.08

In short I think the different results may be due to an error in your code. I used mice 3.0.9