Statistical Peculiarities in RWD Analyses on AIT
Non-randomized studies are subject to confounding since demographic and clinical patient characteristics influencing physicians’ prescribing choices or affecting treatment outcomes may systematically differ between patient cohorts, resulting in a biased estimation of treatment effects (45). Established confounders are patient sex and age, differences in AR symptoms and disease severity before index, duration of AIT, and length of analysis periods. Other factors potentially causing the confounding variables not being properly balanced are the lack of blinding and randomization, residual monitoring bias and confounding by indication (28). Matching methods attempt to approximate the ideal of randomized controlled trials despite using observational data. Two common matching methods are ‘exact matching’ (EM) and ‘propensity score matching’ (PSM) (45). EM uses the complete dataset to identify an exact match covering all confounding variables resulting in smaller variance of the treatment effect but greater danger of excluding cases. With increasing numbers of matching variables, the variability of the patient population is decreased but also the resulting study sample size is diminished. Hence, EM is applied when datasets covering a high patient count are available. On the other hand, with PSM, patients are matched on a single propensity score in order to identify not the exact but the nearest neighbor, i.e. the probability of receiving the exposure of interest given the observed baseline characteristics (45). This leads to an improved applicability in datasets with a high number of confounding variables available but bares the danger of greater variance (46). Both methods were used in the real-world analyses summarized on hand, EM in 6 studies (15,17–19,21,22), PSM in one study (25).