Lecture 11: Instrumental variables

The closest we get to a "plug and play" approach to causal inference.

Parameter identification

In BNS, a common cgoal is to decompose associations into their individual causal components. In linear models, we can use path coefficients. Sometimes we can identify all path coefficients all at once. This doesn't always work when you have latent variables.

Sometimes we just want to know only specific coefficients, so it doesn't matter if we can't identify all coefficients at once. We can e.g. use the backdoor criterion to identify specific edges.

Residual confounding

  • Causal estimates from observational studies are affected by confounding
  • Backdoor criterion provides a theoretical answer to confounding, but this requires us to know and measure all those confounding variables.
    • This is often not feasible in practice

Slide 11-9: four different analyses on observational data shows healthy effects of vitamin C for both males and females. But the example on the slide shows no effect for a randomized trial (so with experimental data). So despite compensating for confounders, there is apparently still bias in the analysis based on observational data.

This can be explained by unmeasured residual confounders. It's not very hard to come up with may more possible residuals. Some of those are not easy to measure (genetics as a confounder), or will have large measurements error (alcohol consumption based on self-report?).

How can we be sure that we have caught all the confounders from observational data?

You can't be 100% sure. This is also why a medicine will only get accepted using the highest standards, using randomized controlled trials (interventional experimental data).

Main idea of IV; Joshua D. Angrist

Does serving in a war have a negative impact on your lifetime earnings?

This is a scenario that for sure has confounders, e.g. one is more likely to serve if you have a low education, instead of having a high education with a large chance at a good job in your home country. Anyways, the point is that finding all confounders is infeasable, so we need a different approach.

  • Goal: estimate causal effect of X on Y
  • Problem: unobserved U
  • Solution: two-stages least squares using an instrumental variable (IV)

IV has two strict requirements:

  • Exogeneity: I unrelated to any confounders U
  • Exclusion: I only affects Y through X, but not directly

Approach:

  • Regress X on I and obtain prediction $X_{hat}$
    • Keeps the "part of X" predict by I
  • Regress Y on $X_{hat}$

This does widen your confidence interval, which is good to take into account.

X is a combination of I and U. The confounding comes from I -> X <- U -> Y. By predicting X from I only, we remove the information from U about X. We only leave the part of X that can be explained by I. We construct independence this way.

Of course $X_{hat}$ will be a noisy version of X. So the cost of this constructed independence is a loss of precision, widening the confidence interval on our prediction.

How do we find instrumental variables?

"Instruments are like unicorns". It's beautiful when they exist, but I only believe you when you show me.

In the Vietnam example of Angrist, he found an IV: the "draft lottery". Apparently, only people on certain birthdays were asked to draft.

In this case, because the lottery is random by definition, we know it has to be independent from any possible confounder. So we have exogeneity (this is usually very hard to determine)!

Having this call increases the chance of drafting, but does not perfectly explain it. If you get the call, you may still not go to war. (Called imperfect compliance in medicine).

We also assume that later in life you do not have larger lifetime earnings just because you received this independent phone call (independent of whether you went to war or not). This assumption seems very fair, so we have exclusion.

Mendelian Randomization

One group of IVs that has gained traction is "genetic instruments".

2003: George Davey Smith; (Dutch researcher Martijn Katan was earlier though)

A Single Nucleotide Polynorphism (SNP) is a genome variation in a single base-pair. Studies look for correlations between certain SNPs and diseases.

SNPs are inherited and there is randomness in this process. They come randomly from the genotype from your mother or father. (We know this is random because the underlying biological process is known).

A Mendelian randomization (MR) study uses SNPs as IVs.

  • Exogeneity: I is random so there cannot be an incoming arrow. (May be an arrow from I to U though)
  • Exclusion: There may be direct arrows from I to Y

So the assumptions are not guaranteed, but are apparently many sanity checks available for MR.

Use d-separation to ensure exogeneity

We can make an IV instrument from the confounder by applying d-separation. So even though usually the exogeneity is violated, we can try to make a conditional instrumental variable.

Determining whether an I is indeed an unconditional IV is NP-hard. But this is mainly due to edge cases, so in practice we can often do OK.

Dagitty contains a function to get instrumental variables! (Including conditional instrumental variables)