how does R handle NA values vs deleted values with regressions

Question

Say I have a table and I remove all the inapplicable values and I ran a regression. If I ran the exact same regression on the same table, but this time instead of removing the inapplicable values, I turned them into NA values, would the regression still give me the same coefficients?

Answer 1

The regression would omit any NA values prior to doing the analysis (i.e. deleting any row that contains a missing NA in any of the predictor variables or the outcome variable). You can check this by comparing the degrees of freedom and other statistics for both models.

Here's a toy example:

head(mtcars)

# check the data set size (all non-missings)
dim(mtcars) # has 32 rows

# Introduce some missings
set.seed(5)
mtcars[sample(1:nrow(mtcars), 5), sample(1:ncol(mtcars), 5)] <- NA

head(mtcars)

# Create an alternative where all missings are omitted
mtcars_NA_omit <- na.omit(mtcars)

# Check the data set size again
dim(mtcars_NA_omit) # Now only has 27 rows

# Now compare some simple linear regressions
summary(lm(mpg ~ cyl + hp + am + gear, data = mtcars))
summary(lm(mpg ~ cyl + hp + am + gear, data = mtcars_NA_omit))

Comparing the two summaries you can see that they are identical, with the one exception that for the first model, there's a warning message that 5 csaes have been dropped due to missingness, which is exactly what we did manually in our mtcars_NA_omit example.

# First, original model

Call:
lm(formula = mpg ~ cyl + hp + am + gear, data = mtcars)

Residuals:
    Min      1Q  Median      3Q     Max 
-5.0835 -1.7594 -0.2023  1.4313  5.6948 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) 29.64284    7.02359   4.220 0.000352 ***
cyl         -1.04494    0.83565  -1.250 0.224275    
hp          -0.03913    0.01918  -2.040 0.053525 .  
am           4.02895    1.90342   2.117 0.045832 *  
gear         0.31413    1.48881   0.211 0.834833    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 2.947 on 22 degrees of freedom
  (5 observations deleted due to missingness)
Multiple R-squared:  0.7998,    Adjusted R-squared:  0.7635 
F-statistic: 21.98 on 4 and 22 DF,  p-value: 2.023e-07

# Second model where we dropped missings manually    

Call:
lm(formula = mpg ~ cyl + hp + am + gear, data = mtcars_NA_omit)

Residuals:
    Min      1Q  Median      3Q     Max 
-5.0835 -1.7594 -0.2023  1.4313  5.6948 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) 29.64284    7.02359   4.220 0.000352 ***
cyl         -1.04494    0.83565  -1.250 0.224275    
hp          -0.03913    0.01918  -2.040 0.053525 .  
am           4.02895    1.90342   2.117 0.045832 *  
gear         0.31413    1.48881   0.211 0.834833    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 2.947 on 22 degrees of freedom
Multiple R-squared:  0.7998,    Adjusted R-squared:  0.7635 
F-statistic: 21.98 on 4 and 22 DF,  p-value: 2.023e-07