Content area
Instrumental variable (IV) methods can be used to estimate treatment effects in real-world data in the presence of unmeasured confounding but exchange the “no unmeasured confounding” assumption with other unverifiable assumptions. Recently, the trend-in-trend design (TT) was proposed to relax IV assumptions by allowing an IV (calendar time) to be associated with unmeasured confounders in the total study population. However, the TT design comes with additional unverifiable assumptions of its own and no study has evaluated the relative advantages and disadvantages of the TT design in real-world scenarios.
We applied the TT design and calendar time IV in a Medicare population with diabetes comparing older adults who initiated second-line antihyperglycemic initiators of interest (TZD vs. DPP-4i) using the known effect of TZD on congestive heart failure (CHF) and its null effect on cataract as positive and negative control outcomes. We used the active comparator, new user study design to identify TZD and DPP-4i initiators and propensity score weighting methods as the comparison. The TZD vs. DPP-4i setting provides a good opportunity to evaluate these methods since we have strong calendar time trends in prescribing and good equipoise between treatments. We supplemented our empirical example with a Monte-Carlo simulation study assessing scenarios where TT assumptions are (i) upheld and (ii) violated.
We found that TT design provides a biased estimate in the presence of unmeasured confounding and no improvement in validity compared with traditional PS-based methods. TT estimates are also outlying and different from prior literature and estimates from other methods in the empirical example. Precision of TT estimates is also lower than IV and PS-based methods. IV estimates on the other hand provided estimates close to the prior literature in the applied example and unconfounded estimated in simulations where TT assumptions are upheld.