OBJECTIVES: Instrumental Variable (IV) approaches have been advocated to estimate causal treatment effects using observational data in the presence of unmeasured confounding. However, IV methods can be subject to weak instrument bias (and therefore attenuation of the causal effect towards the null). They also rely on strong assumptions, which cannot be tested from the data and have to be justified by background knowledge.

METHODS: Using an observational cohort of 150 patients experiencing exacerbations due to Chronic Obstructive Pulmonary Disease (COPD), we used an IV approach to estimate the effect of treatment with prednisolone and antibiotics compared to prednisolone alone in terms of Forced Expiratory Volume (FEV) post treatment. Three potential IVs were considered: sputum colour, distance from facility and deprivation index. We also undertook a simulation study (based on the characteristics of the cohort study) to compare these IVs with regard to weak instrument bias and to assess the sensitivity of our analyses to violation of IV assumptions.

RESULTS: The three potential IVs displayed varying degrees of strength in this cohort of COPD patients, and our simulation study confirmed that the impact that this variability had on our study estimates, and therefore conclusions, could range from minor to considerable depending upon the weakness of a particular instrument.

CONCLUSIONS: IV approaches to estimating causal treatment effects from observational data are becoming popular. Finding a suitable IV is not always straightforward. Our study illustrates the potential dangers associated with weak (but valid) instruments or with instruments that violate core assumptions. We recommend that the impact on an analysis in any particular context should be explored using a simulation study approach.