1
$\begingroup$

I was wondering why, in a regression model including an interaction term, the confidence interval of one of the main effects becomes wider. Consider this Cox regression, where the variable IR_BMI27 is a categorical variable with four categories, which is why we see three categories (Hazard Ratios being expressed relative to the reference category), and the outcome is binary:

First Cox regression

I added an interaction term between and the variable Age and, as you can see, while the point estimate of the HR of the fourth category of IR_BMI27 strongly increases and remains statistically significant, its confidence interval becomes wider (less so in the other categories):

Second Cox regression

Why would that happen? I am curious about the theoretical basis for that. I am familiar with the interpretation of a shift in effect size (or statistical significance) for a main effect when introducing an interaction term, but I wonder whether the sheer change in confidence interval size reflects the same principles. Does it mean that the distribution of age in that category is skewed? Or does it mean that the distribution of the outcome of interest across ages in that category is skewed? This is a table reporting the sample sizes and the CVD events (that is the binary outcome variable) per category of IR_BMI27, all stratified by Age in decennia (39 to 50, 51 to 60, 61 to 70, 71 to 80): I cannot see anything strange.

enter image description here

$\endgroup$
1
$\begingroup$

The errors of the variables may be correlated leading to very large errors in some coefficient when they strongly correlate with others. The matrix $(X^TX)^{−1}$ describes this correlation.

Error in the regression line

The image below shows intuitively how this changes when adding other regressors.

  • On the left the intercept is the error of the mean of the population.
  • On the right the intercept is the error of the regression line intercept.

change of meaning

Confidence regions for correlated parameters

The next image displays the confidence regions (contrasting with confidence intervals) of the above regression in a 2-D plot. Here it takes into account the correlation between the parameters.

The ellipse shows the confidence region which is a based on a multivariate distribution of the slope and intercept which may be related via a correlation matrix. For illustration an alternative type of region is also show. This is depicted by the box which is based on two single variate distributions assuming independence (now the confidence for the single variables is $\sqrt{0.95}$).

By changing the model from $y = a + bx$ to a shifted model $y = a + b(x-35.5)$ we see that the correlation between the slope and intercept changes. Now the "intercept" coincides with the standard error of the line around the point $x=35.5$ which you see in the image above is smaller.

correlation and confidence regions

#used model and data
set.seed(1)

xt <- seq(0,40,0.1)
x <- c(1:10)+30
y <- 10+0.5*x+rnorm(10,0,3)
$\endgroup$
  • $\begingroup$ Thank you - I selected your answer because of the very nice graphical explanation. $\endgroup$ – torwart Sep 4 '18 at 20:48
1
$\begingroup$

Let me take a shot. First I'm going to convert your HR into their original values by taking the log:

table 1

          results    lower    upper
IRBMI274  0.68       0.59     0.77

table 2

              results    lower    upper
IRBMI274      2.60       1.91     3.28
IRBMI274:AGE  -0.03      -0.04    -0.02

The standard error will increase on the main effect because we are adding more correlated variables to the regression (think of the extreme case of perfect collinearity: the standard errors would be massive/infinity). So we shouldn't be surprised that the std. error increases.

Judging from the coefficients above, I don't think you demeaned the age variable (using dubious means, I think the average is ~64 years in the dataset). Try demeaning age first, as that can help reduce possible colinearity.

$\endgroup$
  • $\begingroup$ Thank you - I really appreciated your quick response based on my data and your link to a solution. I only selected Martijn's answer because of the effort he took in making his nice graphical explanation, I am sure you'll understand. $\endgroup$ – torwart Sep 4 '18 at 20:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.