02.Causation03.Instrumental variables - sporedata/researchdesigneR GitHub Wiki
- When researchers want to establish a causal association regarding an intervention and outcomes while using an observational dataset
An instrument is a variable strongly associated with the exposure and not associated with the outcome. In other words, the instrument should affect the outcome only through the exposure.
Instead of experimentally randomizing individuals to an intervention, if researchers find an instrument, they can control for confounding without having information on the confounders. However, it is difficult for a variable to satisfy the instrumental variable conditions.
The distance from home to the clinic is an example of an instrument frequently used in health services research. Some patients live nearby hospitals with specific services while others live far away from them. One could compute the excess travel time as the time that patients would take to arrive at the nearest hospital with a certain service minus the time that it would take them to arrive at the nearest hospital without that service. In an emergency, patients will most likely receive care in a nearby hospital. In this case, the excess travel time will be related to the type of treatment that the patient receives (the exposure) but unrelated to the outcome. However, excess travel time would be close to zero if most patients live nearby clinics with any kind of service (like in a big city). To strengthen the instrumental variable, researchers can restrict the population (for example, not using the information from big cities) or perform nonbipartite matching. The latter procedure creates pairs of individuals with varying degrees of similarity between their covariates and the instrument. The ideal situation would be a set of pairs with very similar covariates but different instruments [5].
Instrumental variables (IVs) are a solution to address the unmeasured confounders. It is necessary to follow some parameters to select an IV: i) the IV must be highly correlated to the treatment, ii) it can only affect the outcome through treatment, and iii) it must not be associated with unmeasured confounders after controlling for measured confounders. If these assumptions hold, cause-and-effect interpretations and estimates can be drawn based on the analysis of RCD (routinely collected data) using IVs [6].
When the IV is weak, even if it is a valid IV, treatment effect estimates based on IV methods have some limitations, such as large variance even with large samples, which can lead to bias in treatment effect estimates [6].
Instrumental variables are like nature flipping a coin.
- Books
- Articles
- US Food and Drug Administration Approvals of Drugs and Devices Based on Nonrandomized Clinical Trials [3].
- Using instrumental variables to address bias from unobserved confounders [4].
- Common references for causation
Using an instrumental variable for causal analysis means taking advantage of a natural experiment.
For a variable to be considered an instrument, it should be connected to the intervention but not to the outcome:
See
- Avoiding bias from weak instruments in Mendelian randomization studies [1].
- Association of Antibiotic Treatment With Outcomes in Patients Hospitalized for an Asthma Exacerbation Treated With Systemic Corticosteroids [2].
- Guidelines for performing Mendelian randomization investigations
[1] Burgess S, Thompson SG, CRP CHD Genetics Collaboration. Avoiding bias from weak instruments in Mendelian randomization studies. International journal of epidemiology. 2011 Jun 1;40(3):755-64.
[2] Stefan MS, Shieh MS, Spitzer KA, Pekow PS, Krishnan JA, Au DH, Lindenauer PK. Association of Antibiotic Treatment With Outcomes in Patients Hospitalized for an Asthma Exacerbation Treated With Systemic Corticosteroids. JAMA internal medicine. 2019 Mar 1;179(3):333-9.
[3] Razavi M, Glasziou P, Klocksieben FA, Ioannidis JP, Chalmers I, Djulbegovic B. US Food and Drug Administration Approvals of drugs and devices based on nonrandomized clinical trials: a systematic review and meta-analysis. JAMA network open. 2019 Sep 4;2(9):e1911111-.
[4] Maciejewski ML, Brookhart MA. Using instrumental variables to address bias from unobserved confounders. Jama. 2019 Jun 4;321(21):2124-5.
[5] Rosenbaum, P. R.Design of observational studies. Springer. 2020; 10: 159-165
[6] Ezzalfani M, Porcher R, Savignoni A, Delaloge S, Filleron T, Robain M, ... & ESME Group. Addressing the issue of bias in observational studies: Using instrumental variables and a quasi-randomization trial in an ESME research project. Plos one. 2021 Sep 15;16(9):e0255017.