gms | German Medical Science

Introduction: The PRISCUS list assembles evidence on medications that are potentially inappropriate in elderly patients (PIMs) [1]. Efforts to reduce inappropriate prescription behavior could be bolstered up by more specific knowledge about the real world effect of deprescribing specific substances. We evaluated the methodological and practical feasibility of producing such evidence using German routine claims data.

Methods: Following the principles of target trial emulation [2], we designed a cluster-randomized (cRCT) trial. The hypothetical intervention consisted of a medication review that should lead to deprescription of sulfonylurea in favour of DPP4 inhibitors in older patients with diabetes type 2. Methodological challenges included 1) the emulation of cluster-level treatment assignment, 2) confounding associated with prior treatment history and patient clustering within physicians, 3) choosing strategies to deal with intercurrent events (i.e. emulating a real intention-to-treat effect). We selected confounders using a systematic but pragmatic approach that incorporated national guidelines, summaries of product characteristics, expert knowledge and variables considered in relevant RCTs. To adjust for confounding, we used a novel propensity score based method, namely overlap weighting [3].

Results: It was not possible to perfectly emulate the pre-defined hypothetical cRCT. In contrast to our expectation, we found that patients treated with DPP4 inhibitors had a higher rate of combined all-cause hospitalisations and outpatient visits after weighting compared with those treated with sulfonylureas (rate ratio=1.03 [1.02-1.03]) in the total population. However, E-values [4] suggested that this finding could likely be due to bias. We did found a robust protective effect of DPP4 inhibitors on the risk for severe hypoglycaemia in the subgroups of new users (RR=0.51 [0.33, 0.76]) and patients with severe renal insufficiency (RR=0.31 [0.16, 0.61]).

Discussion: The target trial emulation approach was a helpful tool when developing the study analyses but some relevant aspects remain unspecified. There is a lack of concepts for implementing more complex designs like cRCT. Likewise, we found that extending the framework by explicitly defining the estimand is helpful for clarifying the impact of methodical decisions like confounder adjustment or handling of intercurrent events.

We identified two sources of potential confounding bias that we could not adequately adjust for, that is physician preferences and the patients’ history of diabetes treatment. In addition, we could not find a solution to emulate the cluster structure given small cluster sizes, little within-cluster overlap between the groups and few observed physician-level confounders. We found that overlap weighting outperformed inverse probability weighting with respect to balancing covariates representing prior treatment history, but it should be noted that it emulates a different target population that focuses on patients with clinical equipoise [3].

Conclusion: We found that deprescribing of SUs and using DPP4i instead might reduce hypoglycaemia in some groups of elderly people, which agrees with results from RCTs. In contrast our initial hypothesizes that deprescribing reduces provider contacts could not be confirmed. The reasons maybe that we were not able to completely emulate our planned target trial because of a lack of approaches/methods and necessary data (e.g. relevant confounders).

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.