Personal data is collected and stored by many organizations, and this data is highly valuable for analyses in domains such as health research or fraud detection. In many cases, the data of multiple organizations has to be combined because no single data provider holds a sufficiently large number of samples to support statistically reliable conclusions or because the relevant pieces of information are under distinct data ownership. Linking records from different sources requires a common global identifier, but such identifiers are rarely available. Therefore, records are compared and classified pairwise based on quasi-identifying attributes, such as first name, last name or date of birth, to assign common global IDs to records corresponding to the same real-world entity. Data linkages may potentially harm the privacy of individuals by combining information that can be used against their interests. As a consequence, they are subject to many legal and organizational constraints. In most cases, the identifying information is not needed for the actual data analysis but only for the matching. Privacy-preserving record linkage (PPRL) methods are a promising approach to conduct such linkages without sharing identifying plaintext information between the data owners or to a third party, thereby following the data minimization principle and reducing the privacy risk. The data owners encode the identifying data and the matching is conducted on the encoded data only. A variety of encoding techniques have been proposed, mostly employing similarity-preserving but irreversible transformations, thereby enabling fuzzy matching to achieve a higher linkage quality in the presence of minor discrepancies between duplicates. An attribute-level application of such encodings preserves the frequency distribution of values, which allows an adversary to recover plaintext values, e.g., by matching the most common encoded item to the most common real‑world attribute. Record-level encodings are a recommended approach to reduce the risk of reidentification attacks but impose considerable limitations on the linkage algorithm. In this thesis, we focus on this encoding scheme and study its impact on the linkage outcome. We propose methods for improving the robustness and quality of the PPRL results in order to foster its consideration as a privacy-by-default alternative to conventional record linkage. PPRL algorithms are typically applied to linkage problems where personal identifying attributes are used for comparison. Naturally, such datasets from real-world databases are rarely publicly available, which is why research results in this domain are often reported for a few datasets only. Therefore, we investigate the robustness of PPRL methods with record-level encodings. To this end, we designed a modular evaluation framework that includes components for dataset creation, linkage execution, and subsequent result analysis. The approach supports comprehensive studies of different linkage strategies across a broad range of datasets including both curated subsets of existing real-world data and synthetically constructed linkage scenarios that reflect varying levels of data quality. For the majority of the evaluated scenarios, baseline PPRL methods lag behind a state-of-the-art deep learning entity resolution system that depends on the existence of labeled training data though. Assigning weights to attributes is a fundamental component of most record linkage algorithms. Several encoding schemes have been introduced to embed these weights into record-level representations. However, a detailed investigation on how these weights influence the results in the PPRL context is missing. We propose and evaluate adaptive weighting strategies for record-level encodings with the goal of enhancing the robustness and overall quality of PPRL outcomes. Our findings indicate that this rather small modification in the encoding process leads to an improved and more robust linkage quality despite the potentially adverse effects arising from using inconsistent encoding parameters at the data owners. Fine-tuning record linkage methods for high-quality outcomes requires training data that contains information about confirmed matches and non-matches. To address this, we propose a linkage protocol that allows labeling of record pairs in a multi-layer clerical review procedure, in which more information is gradually disclosed if needed. Only a fraction of the uncertain pairs has to be reviewed manually by a (human) oracle using masked display of selected attributes. The majority of labeling is done automatically using hardened attribute-level encodings based on a model trained in active learning cycles built upon the manual reviews. While adding complexity to the otherwise straightforward linkage protocol, the approach not only leads to improved linkage quality but also allows to control the linkage outcome with respect to the balance of false and missed matches and to increase the reliability of the otherwise unverified results.