How will anonymization of simulated clinical data affect the data utility of pharmacoepidemiological studies?

Abstract

Background: The pressure to share more data and being more transparency of clinical study reports has grown and becomes an important topic in recent years. Before clinical data and clinical results can be shared they must undergo anonymization. How anonymization of clinical data affects the utility is poorly-studied, especially in pharmacoepidemiology. Objective: The aim of the study is to describe and evaluate how anonymization of simulated clinical data will affect the data utility of pharmacoepidemiological analyses of these data. Method: We have simulated five clinical datasets with different characteristics, associations, types of outcome and study populations. Suppression, generalization, randomization and k-anonymity were used as our anonymization approaches. These methods will be evaluated by the change in the data and statistical results before and after anonymization. Result: K-anonymity and suppression were the methods that affected the simulated clinical data the most, while generalization and randomization affected the data least. With k-anonymity and suppression there is a risk to overestimating the clinical results due to the elimination of unique records. On the other hand, generalization and randomization preserved the most data utility but they were less effective in anonymizing the data. Conclusion: Our study revealed that different anonymization approaches can affect the clinical results differently. The more we anonymize a record or attribute, the less utility is provided. It is therefore important to construct a balance of data utility and effectiveness of anonymization before the clinical data are published. More investigations about how anonymization of clinical data affects data utility are needed in order to maximize the benefit of using anonymized clinical data to improve public health.