Premise

When the relationship between two variable is (potentially) dependent on a third, categorical variable ANCOVA (analysis of covariance), or some variant, is commonly used. There are several approaches to testing for differences in slope/intercepts (in the case of a simple linear model) between levels of the stratifying variable. In R the following formula notation is usually used to test for interaction between levels of a factor (f) and the relationship between two continuous variables x and y: y ~ x * f. A simple graphical exploration of this type of model can be done through examination of confidence intervals computed for slope and intercept terms, for each level of our grouping factor (f). An example of a fictitious dataset is presented below. Note that this a rough approximation for testing differences in slope/intercept within a multi-level model. A more robust approach would take into account that we are trying to make several pair-wise comparisons, i.e. something akin to Tukey's HSD. Something like this can be done with the multcomp package. For any real data set you should always consult a real statistician.

Example Multi-Level Model: each panel represents a model fit to y ~ x, for group f

Example Multi-Level Data

Implementation

Example Multi-Level Model: Confidence Intervals: parameter estimates along with 95% confidence interval, within each level of our grouping factor (f).

Automated Plotting of Parameter Confidence Intervals

Formal Evaluation with lm()

The first two lines in the output below are testing the hypothesis that the slope and intercept term for level 'a' are not different than 0. Subsequent hypothesis tests are relative to the first 'level' in the dataset. In this case we are testing the hypothesis that intercept and slope terms for levels 'b' through 'j' are not different than the corresponding terms for level 'a'. From the output below we can see that none of the intercept terms (levels 'b' through 'j') are different than for 'a', and that the slope term for level 'd' is only marginally "different" (p=0.0625) than the slope term for 'a'.

Testing Model Terms