# Regression Analysis (Line of Best Fit) for Categorical Variables

Brief Background/Motivation: I am looking at an Income vs. Education table that is adapted from a dissertation and was used in developing a curriculum in a social justice mathematics program. In the dissertation, the author discusses using these data (income vs. education level, broken down by gender) to have students create a line of best fit, but does not explain how the categorical variable is treated or transformed into an ordinal variable.

The issue, as I see it, is that education level is categorical and not continuous; I have been unable to find a "standard" or even suggested approach regarding how to treat the categorical variable. I see two different ways:

1. Starting at 1, label each category 1-7. This assumes a uniform/linear step size (i.e., the difference between some high school and completing high school is the same as the difference between a master's degree and a doctorate) which is clearly problematic, but one possibility.

2. Approximate the number of years of schooling for each category. For example, "high school completion" would be 13, bachelor's degree would be 17, masters would be anything from 18 to 19, etc. Then you have to make some decisions about categories such as "some high school": is this a 10, 11 or 12? Also, how should you count the category of "no high school"? Is this a 7 or 8 or 9? This is also clearly subjective and has its own problematics, but is actually roughly the same as (1).

Question: Do either of the two approaches suggested above work? Or is there another, better way to treat these data?

Pointers to relevant papers or resources would be welcome, too.

The application of regression analysis to independent categorical variables is often performed using the so-called dummy variables. Dummy variables are artificial variables that are built to represent an attribute with $\geq2$ categories. Although there are several ways to create these variables, when we deal with a categorical variable having $k$ levels a commonly used approach is to define $k-1$ dichotomous variables to represent the different levels. In particular, the first dummy variable is set to $1$ for all items that are in the first level (in your case, all individuals with $<9^\text{th}$ grade) and to $0$ for all other items, the second is set to $1$ for all items that are in the second level (in your case, all individuals who have from $<9^\text{th}$ to $12^\text{th}$ grade and no completion) and to $0$ for all other items, and so on. The use of $k-1\,$ dummy variables is justified by the fact that one category (usually the last or the first) is treated as the "reference" category, assuming that there is a constant in our regression model (if no constant is assumed, $k$ dummy variables can be used).
This is a classical case of the so called "Two way ANOVA" which is basically a statistical nomenclature for a linear regression model that has only categorical explanatory variables. In your case, one variable is the level of education that consists out of $8$ categories and gender/sex - out of $2$ categories. You don't need to assign any numbers to these categories and simply leave them as $A, B, C..$, and $F, M$ or whatever you want. Formally, let $\alpha_i$ be the level of education such that $i=1,2,3,...,8$ that correspond to the $8$ levels that you have (without any particular order) and $\beta_j$ , $j=1,2$ will be the two "levels" of gender, hence your models is given by $$y_{ijk} = \alpha_i + \beta_j + (\alpha\beta)_{ij} + \epsilon_{ijk},$$ where $\epsilon_{ijk} \sim \mathcal{N}(0,\sigma^2)$. That means that every $k$-th subject comes from the $i$th level of education and $j$th level of gender. The interaction variable suggest that the effect of education may differ by gender. Note that even if your data (the income variable) is not really normal and do not posses equal variances across the levels, ANOVA model will still be robust up to some point. I would suggest to use $\log(income)$ from the start as income itself will have right-asymmetric form. And another important point, where you compare the mean income between the all possible levels (cells) - don't forget to adjust your p.values, otherwise the significance level will decrease very rapidly as a function of the number of conducted comparisons. Common approaches to control the level of significance are Tukey, Bonferroni and the (relatively) novel FDR.