This JAMA Guide to Statistics and Methods reviews the use of instrumental variable analysis in observational and randomized studies and how, under specific assumptions, they can provide unbiased estimates of treatment effects even if unobserved confounding exists.
Randomized clinical trials are considered the most reliable source of evidence for the effects of medical interventions, but nonexperimental studies are often used to assess the effectiveness of treatments as they are used in actual clinical practice. In nonexperimental studies, treatment groups may differ by important patient characteristics, such as disease severity, frailty, cognitive function, vulnerability to adverse effects, and ability to pay.1 While statistical adjustment can account for imbalances in observed characteristics between groups, observed imbalances are concerning because they suggest that unobserved differences may also exist. Unobserved patient characteristics that influence both treatment and the outcomes result in “unobserved confounding,” a bias that cannot be removed using standard statistical adjustment.1
Instrumental variable analysis is an approach used to help address unobserved confounding when estimating treatment effects in nonrandomized studies and in randomized studies when protocol nonadherence exists. An instrumental variable is a factor that should effectively randomize some patients to the different groups; it should be correlated with the treatment received and related to outcomes only through its effect on treatment. Under specific assumptions, instrumental variable methods can provide unbiased estimates of treatment effects even if unobserved confounding exists.
In an article published in JAMA Network Open,2 Desai and colleagues used instrumental variable methods to assess the association between initiation of osteoporosis medications and nonvertebral fracture risk in a cohort of 97 169 patients aged 50 years or older who were hospitalized for hip fracture from 2004 to 2015.2 The authors used an instrumental variable analysis to address suspected unobserved confounding from frailty, disease risk, and other factors that might bias the estimate of the treatment effect.
Why Are Instrumental Variables Used?
When patients and clinicians are free to choose treatments, some of the factors that influence treatment choice may also be strongly related to the likelihood of a good outcome. In the study by Desai et al,2 men were less likely to initiate osteoporosis medication than women, and they also may be less likely to develop fractures. If all confounders are observed, then potential confounding influence can be addressed by adjusting for them in a statistical model. If only some confounders are observed, then outcome differences between treatment groups may be driven by treatment alone, confounders alone, or both. For example, if patients with osteoporosis were not given a preventive treatment because they were perceived to be close to the end of life, the group of patients not receiving treatment might be more severely ill or frail than patients receiving treatment. If frailty that is not measured in the study influences patient outcomes ...