key: cord-0024998-m0xx0d3s
authors: McIntosh, Cameron N
title: RE: “CONSIDER THIS BEFORE USING THE SARS-COV-2 PANDEMIC AS AN INSTRUMENTAL VARIABLE IN AN EPIDEMIOLOGIC STUDY”
date: 2021-09-15
journal: Am J Epidemiol
DOI: 10.1093/aje/kwab229
sha: 2388c3e6d01a082a342c3ff85e928a9694d7910a
doc_id: 24998
cord_uid: m0xx0d3s

nan

I read with interest the recent commentary by Dimitris and Platt (1) on potential pitfalls in the use of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic as an instrumental variable in regression models. The authors correctly point out that the tenability of the exclusion restriction-that is, the assumption that the instrument affects the outcome solely via the exposure variableis dubious in this case. In particular, given the existence of a complex set of pathways from the pandemic "shock" to the broad array of social, health, and economic outcomes of interest, the SARS-CoV-2 instrument may not be a valid instrument for exposures.

However, I worry that researchers following Dimitris and Platt's advice might end up throwing the baby out with the bathwater, as the commentary misses a number of approaches for handling violations of the exclusion restriction. First, and perhaps contrary to popular belief, the exclusion restriction can be tested if the researcher is willing to assume a specific value for the correlation between the instrument and the error term of the model (2) . Fortunately, in the case of an exogenous shock like SARS-CoV-2, this correlation could defensibly be set to 0, thereby allowing an unambiguous test of exclusion. If exclusion does not hold, it is possible to perform a sensitivity analysis using prespecified values of the exclusion restriction to determine how large the violation would need to be in order to overturn the inferences (3). If it takes an extremely large violation to reverse the conclusions and the researcher can make a reasoned argument that the true extent of the violation is unlikely to be that high, one can have more confidence in the findings.

Further, one can potentially relax the exclusion condition entirely. One way to do this is to obtain an estimate of the direct effect of the instrument on the outcome from a subset of observations where the instrument has no effect on the exposure (4, 5) . In addition, the exclusion restriction can potentially be replaced with an alternative InSIDE (instrument strength independent of direct effect) assumption, which states that the direct effect of the instrument on the outcome is uncorrelated with the instrument's effect on the exposure (6). This assumption is widely used in Mendelian randomization studies to deal with horizontal pleiotropy, that is, when genetic instruments affect the outcome through pathways other than the exposure. The InSIDE assumption is most justifiable in a 2-sample setting in which the instrument-exposure association and the instrumentoutcome association are estimated in different samples.

Because SARS-CoV-2 is likely to be a strong instrument for various exposures, such as online schooling in Dimitris and Platt's example (1), numerous opportunities may be missed by considering it nonviable due to questionable exclusion restrictions. Coupled with the methods noted above, the SARS-CoV-2 instrument could potentially help to identify causal effects in several contexts. Of course, none of the approaches mentioned here is a silver bullet, and each needs to be justified on a case-by-case basis. However, with careful application, these methods can facilitate causal inference in situations where randomized controlled trials are infeasible. I hope that this note helps to balance Dimitris and Platt's critique, as well as raise awareness among researchers of different techniques for dealing with deviations from the exclusion restriction in instrumental-variable models.

Consider this before using the SARS-CoV-2 pandemic as an instrumental variable in an epidemiologic study

Testing the impossible: identifying exclusion restrictions

Plausibly exogenous

Beyond plausibly exogenous

Testing and relaxing the exclusion restriction in the control function approach

The use of two-sample methods for Mendelian randomization analyses on single large datasets

Conflict of interest: none declared.