4
$\begingroup$

In a recent paper, Edelman et al. examine (amongst other things) how discrimination on AirBnB varies with the characteristics of hosts. First, they conduct a field experiment which involves sending a large number of fictitious applications, some of which are associated with African-American names and some of which are not. Then, they regress whether an applicant is accepted on the guest race, a host characteristic and the interaction. For example, they regress acceptance on guest race, host race and the interaction of the two. The idea is that if African-American hosts (for instance) are less likely to discriminate, then the coefficient on the interaction term should be non-zero since the effect of receiving an application from an African-American should depend on whether the host is African-American.

So far, so good. However, given that host characteristics are not randomly assigned, we should control for them. Otherwise, we might find that African-Americans are less likely to discriminate even though the real explanation is that they tend to offer expensive properties, and people who offer expensive properties are less likely to discriminate (made up example to illustrate the point). They seem aware of this, since they throw in various host and location characteristics as controls. But wouldn't it have been better to use the interactions of these covariates with applicant ethnicity as the controls? After all, we want to control for the fact that people with certain characteristics might respond to African-American applicants differently, and this means controlling for interaction effects.

$\endgroup$
  • $\begingroup$ By "interactions" you mean to include the product of two controls also as a regressor I guess? $\endgroup$ – Alecos Papadopoulos Apr 23 '18 at 21:41
  • $\begingroup$ Exactly. Specifically, I am referring to the product of the each covariate (say property price) with the dummy that equals 1 if the fictitious applicant is African-American. $\endgroup$ – afreelunch Apr 24 '18 at 10:34
1
$\begingroup$

I leave for another thread the issue "What is the justification of using the arithmetic operation of multiplication and so the product of two variables in order to model the concept of interaction?"

Practically, product-terms come from the tradition of the Translog regression specification, which is a second-order approximation to any actual specification, and produces such products. They are therefore considered "second-order" effects, meaning that they are included in a specification alongside the stand-alone variables. This means that their importance is judged after the effect of the stand-alone variables has been taken into account.

I have not seen in the literature a situation where "interaction terms" are included alone, without the stand-alone presence of the two variables forming the product.

So for the specific study mentioned by the OP, conventional practice would ask "why not include also interaction terms?"

Nevertheless, it appears intriguing to argue convincingly that in a certain case, we should only include the "interaction term" of two variables.

$\endgroup$

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.