It seems that often in social science, race is examined in causal terms, as researchers are interested in the differences between various ethnic groups in outcomes when controlling for other covariates. However, my understanding is that we actually can't use race for causal inference due to the omitted variable bias and the fact that essentially everything you control for is like a "post-treatment effect".

So why do social scientists do this? Is it valid in certain contexts?

  • 1
    $\begingroup$ VanderWeele & Robinson (2014) has a discussion about this very matter. $\endgroup$ – Noah Sep 13 '18 at 2:08
  • $\begingroup$ They keep using race despite its definition is fuzzy at best. What is exactly a race? I don't consider myself to belong to any of the races that drop down in surveys and forms. $\endgroup$ – Aksakal Oct 10 '18 at 17:04

Race and ethnicity are variables that cannot be "controlled" in experiments, since it is not possible for the researcher to assign or change this characteristic of the study participant.$^\dagger$ For this reason, causal inference relating to race and ethnicity cannot generally rely on randomised controlled trials, and must instead fall back on uncontrolled observational studies. As with other uncontrolled studies on any other topic, this comes with all the regular drawbacks and caveats on causal interpretation of results, including the possibility that there may be omitted "lurking variables" that affect analysis. As a general principle, causal inference from uncontrolled studies is not reliable, and tends to be reasonable only in cases where the predictor in question is shown to have a statistical relationship conditional on a wide array of covariates, and tends to retain its predictive ability under variations in covariates that are not themselves intermediate causes.

Many studies in the social sciences include race/ethnicity as covariates, and the goal is to filter out these variables to find some other causal or statistical relationship. There may be some studies where race/ethnicity is of direct interest as a predictor, and in this case the researcher needs to be careful to distinguish predictive efffects from causal effects, as in any uncontrolled observational study. There is certainly no scientific problem with including race/ethnicity as variables in social science studies; the problems, if any, arise in regard to interpretation of results. There is a good discussion of the causal interpretation of race variables in VanderWeele and Robinson (2014).

For the most part, all of this is just a matter of applying general statistical principles to a particular set of variables. However, one issue that arises specifically in regard to causal inference regarding race and ethnicity is competing theories of whether any causality is direct (i.e., genetic/hereditary) or indirect (i.e., due to discrimination). This aspect of the problem has been discussed at length by the economist Thomas Sowell in a series of books discussing statistical disparities among racial groups (see esp., Sowell 1975, Sowell 2013, Sowell 2018). Sowell notes that historically, there was an excessive tendency to ascribe all racial disparities to genetic causes in the nineteenth and early twentieth centuries, and since the late twentieth century there is now an excessive tendency to ascribe all racial disparities to discrimination. Both of these constitute a failure to properly apply statistical reasoning relating to causality, and both tend to occur due to a conflation of correlation and cause. In any case, if you have not already read these works, they may give you a better understanding of the difficulties that arise in making causal inferences from statistical disparities among racial and ethnic groups.

It is difficult to answer your specific question without seeing a particular example of the kind of inference that concerns you. There are a wide variety of cases where social science researchers "use race for causal inference" and the validity would depend on the nature of the data and the resulting inference. (It is not even clear from that framing of the question whether race is the predictor of interest or just a covariate.)

$^\dagger$ Note that there are some randomised experiments where the appearance of race is controlled via some experimental mechanism. For example, many studies on ethnic discrimination in employment use randomised 'correspondence tests' where the researchers control (and randomise) the markers of race and ethnicity in submitted job applications (see e.g., Zschirnt and Ruedin 2015).

  • 1
    $\begingroup$ Thank you for the thoughtful answer. As to what kind of situation I am imagining, I am thinking in particular to statistical analyses of (say) wage gaps between genders or between race. There is also the Stanford Open Policing data set, in which car-search outcomes and arrests were analyzed via regression when controlling for other covariates -- however, race was the primary interest. In these contexts, I am imagining social science in the perspective of trying to analyze something like discrimination (by looking at outcomes) $\endgroup$ – Marcel Sep 11 '18 at 4:14
  • 1
    $\begingroup$ I think you summarize this line of reasoning very well. The usual econometric counterargument is that this conflates the definition of an effect and the acts of identification and inference about it, which are logically separate. Just because there are empirical difficulties in estimating an effect doesn't imply that the effect is ill-defined. $\endgroup$ – Dimitriy V. Masterov Sep 11 '18 at 16:59
  • 1
    $\begingroup$ @Dimitriy: Thanks for your nice comment. I certainly agree that causality can exist even when it is hard to identify/infer, but I can't see anywhere where I've conflated those things (since I only talk about causal inference). If you can point out where I've conflated them I'd be happy to edit the answer. $\endgroup$ – Ben Sep 11 '18 at 23:38
  • 2
    $\begingroup$ @Ben I did not mean to imply that you conflated these. I was thinking that if "no causation without manipulation" is true, than I think that would undermine any meaningful attempts at causal inference about gender and race. $\endgroup$ – Dimitriy V. Masterov Sep 12 '18 at 0:10
  • 1
    $\begingroup$ "it is not possible for the researcher to assign or change this characteristic of the study participant." This doesn't matter at all. If it did, we could not make causal conclusions about the effect of the sun or the season or the day and night cycle. Randomized trials are another kind of intervention. $\endgroup$ – Neil G Sep 24 '18 at 23:50

There is nothing special about race when it comes to causal inference. You can look for its effects just as you might look for the effect of the season.

The difficulty with a lot of causal models in the social sciences is mediation analysis: you want to know when A causes B directly, but not through C (What Ben means when he said direct versus indirect). For example, in the causal model Gender, Hiring, Education, you want to know if Gender affects Hiring directly, but you concede that Gender might affect Education and we all expect Education to affect Hiring. You cannot simply control for Education without opening back doors. This pattern is ubiquitous.

There are approaches to tackling these problems, but as mentioned in Ben's answer, these are experimental changes. For example, although you cannot control for Education without opening more back doors, you can freeze it by having the women use the men's resumes.

  • $\begingroup$ Sorry, I am not sure what you mean when you say "opening back doors". Do you mean allowing the possibility for more confounding covariates? E.g., you control for education, but then perhaps age becomes a confounding variable for being hired $\endgroup$ – Marcel Sep 25 '18 at 0:59
  • $\begingroup$ @marcel: Right. Age would become an open back door if for example, only older women are educated, and the interviewer prefers older candidates. $\endgroup$ – Neil G Sep 25 '18 at 1:04

The idea that race can be a cause is not without dispute.

In a 1986 JASA article, Paul Holland discussed how he and Don Rubin coined the expression, “no causation without manipulation”.

The idea here is that causal inference requires a strict definition of a cause that identifies an intervention that hypothetically could be implemented -- even if that manipulation is not physically possible or ethically feasible.

So what is a hypothetical intervention that would change someone's race? Perhaps a genetic manipulation? But there is no "race" gene that can be flipped like a digital bit. It is hard to imagine there is some way of changing all the genes that contribute to the phenotypes that define race, while keeping all other phenotypes constant. Perhaps instead a cosmetic procedure could make a white person pass as black or vice versa?

Both of these lines of reasoning lead one to think about how race is defined by how other people perceive an individual. So then is "race" what the researcher is looking for? Perhaps it is racial discrimination? Or a persons professed ethnic identity?

If you talk to Miguel A. Hernán and James M. Robins, authors of the causal inference book cited below, they would tell you race is not a valid cause and that only thinking more deeply about what people actually mean when they talk about race as a cause will lead to better inferences.

On the other hand, some, including Judea Pearl, take issue with position.

Here is a quote (from Twitter) by Hernan:

Pearl believes that any causal effect we can name must also exist. To him, the meaning of “the causal effect of A on death” is self-evident. He says we can quantify, say, the causal effect of race or the causal effect of obesity. I don't think we can.

We cannot estimate "the causal effect of obesity" because we don't know what that means. For the causal effect of A to be well defined, we need a common understanding of the interventions that we would use to change A. Otherwise, the effect is undefined.

If by now you are thinking that this is just another academic debate on the sex of the angels, think again: you beliefs about this issue determine your beliefs about the limits of science and about how to conduct data analyses.

Pearl addresses the issue here.

Hernán MA, Robins JM (2018). Causal Inference. Boca Raton: Chapman & Hall/CRC, forthcoming.”

  • 1
    $\begingroup$ This is an interesting answer, but I think you are misunderstanding the meaning of the expression "no causation without manipulation". This is a statement with respect to causal inference, not the existence of cause (i.e., it is an epistemological rule, not a metaphysical rule). It is certainly possible for something that cannot be manipulated by humans to have a causal effect on another thing. If this were not the case then we could not even claim that the Sun warms the Earth! The expression is a warning that causal inference should be based on manipulation. $\endgroup$ – Ben Nov 29 '18 at 11:50
  • $\begingroup$ Ben is right, but also, we can draw causal conclusions (do causal inference) without manipulation, for example, using "the deconfounder" (Wang, et al). $\endgroup$ – Neil G Jul 18 at 19:34

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.