Last updated on March 11, 2018
I unfortunately have too little time for a proper post this week, but I wanted to make time for a quick post. A grad-school friend and colleague sent a link to an interesting new(-ish) paper by Kahn and Whited (2017) that has been making the rounds in finance, and which is forthcoming in the Review of Corporate Finance Studies.
The title of the article is “Identification Is Not Causality, and Vice Versa.” Here is the abstract:
We distinguish between identification and establishing causality. Identification means forming a unique mapping from features of data to quantities that are of interest to economists. Establishing causality by finding sources of exogenous variation is often considered synonymous with identification, but these two concepts are distinct. Exogenous variation is only sometimes necessary and never sufficient to identify economically interesting parameters. Instead, even for causal questions identification must rest on an underlying economic model. We illustrate these points by analyzing identification in three recent papers and by examining the estimation of a simple dynamic model.
This was of interest to me, because I have been mulling over some thoughts over the last few years about the relationship between causal identification and unbiasedness.
The two seem a priori identical, but upon closer inspection, they seem to me like they are not. First, unbiasedness is possible without causal identification by sheer luck: much like a broken clock is right twice a day, a coefficient that is not identified from exogenous variation might well be equal to the true population coefficient by happenstance. Second, causal identification is possible without unbiasedness: we are often perfectly comfortable with a coefficient that is biased toward zero because of classical measurement error but which is identified from exogenous variation–what I have heard referred to as “the good kind of bias.”*
ht: Gabe Power.
* This is usually based on the reasoning that if you have (i) exogenous variation, (ii) a statistically significant coefficient, and (iii) attenuation bias, then you have an even stronger case for there being a causal relationship than if you only had (i) and (ii).