ECON 371 Class Recording 10/29
Overview
The class develops interaction terms in regression, showing how a variable’s effect can differ across groups and how to calculate group-specific marginal effects relative to an omitted baseline. Examples in Stata use female and married wage differences, Houston traffic-stop outcomes by driver and officer race, and a beauty rating interacted with gender; the lecture then connects continuous-variable interactions to turning points and nonlinearities.
Key takeaways
- In a regression with two dummy variables and their product, the effect of either indicator depends on the value of the other; for example, marriage has effect β₁ for men and β₁ + β₃ for women.
- With unmarried men omitted, the predicted group differences are β₂ for unmarried women, β₁ for married men, and β₁ + β₂ + β₃ for married women.
- A statistically significant interaction coefficient indicates that a variable’s effect differs across groups; an insignificant interaction does not provide evidence of different group-specific effects.
- For a continuous variable interacted with a dummy, the continuous-variable marginal effect is β₁ + β₃D, while the dummy’s marginal effect is β₂ + β₃X.
- In a log-wage model, a one-unit increase in a continuous predictor is approximately associated with a 100β percent wage change; an interaction can make that percentage effect differ by group.
- A calculated turning point is substantively relevant only if it falls within the observed range of the variable, such as the 0-to-10 beauty scale used in the wage example.
Chapters
- The nonlinearity quiz is due tomorrow, and the related problem set is due next Tuesday.
- The coming week is reserved for interaction terms, beginning with dummy-variable interactions and later moving to continuous-variable interactions.
- Returned exams are available in class.
- A regression with separate female and married indicators assumes each group effect is constant across the other indicator.
- Female-married, male-married, female-unmarried, and male-unmarried workers form four combinations that may have different wage outcomes.
- An interaction between female and married tests whether the combined group effect differs from adding the two separate effects.
- For wages modeled with married, female, and married-by-female terms, the marginal effect of marriage is β₁ + β₃ for women.
- For men, the female indicator is zero, so the marriage effect is β₁ alone.
- The interaction coefficient β₃ measures how the marriage effect differs for women relative to men.
- The marginal effect of being female is β₂ + β₃ for married people and β₂ for unmarried people.
- A married woman’s predicted wage difference from the baseline includes the marriage coefficient, female coefficient, and interaction coefficient.
- With unmarried men as the omitted group, unmarried women differ by β₂, married men by β₁, and married women by β₁ + β₂ + β₃, apart from the shared intercept and controls.
- Stata’s factor-variable interaction syntax can include both dummy indicators and their interaction with `##`; dummy variables do not need the continuous-variable `c.` prefix.
- The unmarried male group is the reference category, so its indicator effect is absorbed into the intercept.
- A statistically significant interaction means the group-specific effects differ; an insignificant interaction is consistent with the effect of marriage being the same for men and women.
- A Houston traffic-stop example models whether a stopped driver receives a ticket using driver race, officer race, and their interaction.
- Without an interaction, the driver-white coefficient is not statistically significant, while white officers issue tickets at higher rates.
- Adding the white-driver-by-white-officer interaction changes the estimated driver-white effect and indicates a higher ticket probability for white drivers stopped by white officers.
- Controls for the reason for the stop, such as speeding or an arrest warrant, make comparisons between otherwise similar stops more meaningful.
- The same race interaction is applied to a different dependent variable: whether the driver is arrested.
- White drivers have a lower predicted arrest probability in the example, and the combined white-driver and white-officer effect remains negative.
- Unlike the ticket result, the arrest interaction is not statistically significant, so the data do not establish that the arrest effect varies by officer race.
- The examples show why ticketing and arrest outcomes should be analyzed separately rather than treated as the same measure of enforcement.
- The next model interacts a continuous beauty rating with a female indicator to test whether beauty’s wage association differs by gender.
- In a log-wage model, a one-unit increase in beauty corresponds approximately to 100 times its coefficient percent change in wages.
- The beauty effect is β₁ for men and β₁ + β₃ for women, making β₃ the gender difference in the beauty-wage slope.
- On a graph, the female indicator changes the starting point while the interaction changes the slope of predicted wages against beauty.
- In the illustrated regression, the beauty coefficient is about 0.04 and the interaction adds about 0.04 to the beauty slope for women, nearly doubling the estimated slope.
- The female effect at a given beauty level is β₂ + β₃ times beauty, so it can become less negative as the beauty rating rises.
- Setting the female marginal effect to zero identifies a potential turning point, but that point may lie outside the observed scale, such as a beauty rating bounded from 0 to 10.
- The same reasoning applies to other continuous measures, such as income interacted with whether a family has children.
- The next class will revisit continuous-by-dummy interactions and examine marginal effects in more detail; continuous-by-continuous interactions are planned for later.
- Students are asked to review the interaction-term slides before Tuesday because the marginal-effect calculations are challenging.
- Additional Zoom office hours are tentatively planned for the next day from 9:30 to 11:30, while Monday office hours may change.
- An announcement will share details about a Johnson & Johnson recruiting event, and students who have not collected their exams can pick them up.
Transcript excerpt
0:00 All right. So, [clears throat] before we get started here, um that quiz for kind of over all of the nonlinearity stuff that we've talked about, that's due at the end of the day tomorrow. Um that state of problem set due at the end of the day next Tuesday, right? So, that all should be kind of nonlinearity stuff that we covered last week and then on Tuesday's class. So, this week we'll kind of start with interaction or this week. Today we'll start with interaction terms that'll carry over into next week. Both classes next week we'll kind of keep talking about these. We might even have a bit of an abbreviated class. I might have said that Tuesday and then I lied, but I think today we might have a little bit of an abbreviated class. Um just based off how quickly we go through some of this material. Um like I said,
0:47 all of next week will be reserved just for interaction terms. Okay. Um, other than that, I did bring the exams. I didn't get yours back uh at the end of class if you still want it and I've got them with me. Otherwise, any questions for me before you start here? I'll showed up even though it's a gorgeous day out instead of being outside. Okay. Um, so got the Zoom going. So, we're going to talk about interaction terms and some of the stuff we talked about with nonlinearities. Technically, it applies here and a lot of it will be carryover. Um, initially it won't seem like that, but towards the end of today's class and definitely next week, I think you'll start to be able to
1:33 make those connections. So, we have [clears throat] kind of these simple linear regressions that we've been doing
This is the start of the transcript. Save this video free for the whole transcript, searchable and time-coded.
In this course
Related notes
Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, Professor Lantis.