Data linkage is used increasingly across large administrative databases to enhance data resources to allow addressing complex research and policy questions [1]. Accurate data linkage is important as linkage errors can undermine the quality of the linked data by introducing selection bias, and limiting researchers’ ability to accurately represent the targeted populations [2]. In data linkage, names are often used as key identifiers to decide if records from multiple data sources belong to the same person.

Romanisation refers to the transformation of names in local languages to a commonly operable language, which is English in most developed countries in the Global North. Romanisation can be viewed as a process of encoding that selectively retains information operable in (Western) databases (e.g. only containing Latin-based alphabets), and drops information deemed irrelevant or too costly to retain, such as using American Standard Code for Information Interchange (ASCII) for encoding. Unicode is an improvement over ASCII with greater flexibility on type of characters it could encode, but still does not provide tonal representations. Other examples relate to data system designs, people with multiple surnames (e.g. common in Spain and Latin-American countries) might only be allowed to provide one or use hyphenated versions, or letters not in the English alphabet and diacritical marks might be ignored or misrepresented. Romanisation is not always standardised, e.g. there are over 10 different systems for Arabic [3].

When romanised names are not consistently represented in databases, linkage errors are more likely to occur (i.e. false matches, where records belonging to different people are linked together, or missed matches, where records belonging to the same person remain unlinked). As large-scale linked administrative data are used more readily by health professionals and policy makers to generate evidence and drive decisions, this disparity in linkage rates perpetuates health and social inequities for under-represented groups, such as migrants and people from ethnic minoritised communities [4, 5]. For example, a historical linkage of census data from the United States matched only 3.6% of male Chinese migrants between 1880-1900 compared to 16.3% of English migrants [6]. A recent linkage of asylum and resettled refugees with census data in the United Kingdom found a substantial difference in linkage rates by language and country of origin [7]. There is a growing proportion of births in England (37%) and London (66%) to families where one or both parents were born outside the UK; between 2013-2017, 53% of singleton child births in New York City were born to non-US-born mothers [8]. Use of poorly romanised and processed names in data systems will selectively and continually under-represent subpopulations in results, if current data linkage ‘blind spots’ or structural barriers in data systems to inclusion in linked datasets are not actively addressed.

Building on Postel’s work on Chinese-specific pre-processing for historical linkage [9], this data note describes three problems faced by, but not unique to representing Chinese characters for data linkage. We compare standardised romanisation of Cantonese (Jyutping) [10], Mandarin (Pinyin) [11], with the non-standardised Hong Kong Government Cantonese Romanisation system (HKG-romanisation) [12], as a case example. We then propose a solution model for alternative name encodings, by demonstrating the advantages of challenging current standard practices in global information systems to improve linkage and promoting data equity [13].

We scraped online student class lists from schools in Hong Kong that provided both Chinese and English Names (n = 774). We only included names that had a Mandarin or Cantonese origin, based on the provided Chinese and (romanised) English names (n = 771). Records providing English names with no space within forename characters, or use of non-accented Pinyin as English names reflect a Mandarin origin.

We derived Jyutping and Pinyin using the pinyin_jyutping package [16] in Python 3.8 [17], which used an online open-source Cantonese dictionary CC-Canto [18], for each character of the Chinese name. Jyutping and Pinyin tones are represented using numbers: 1-6 for Cantonese and 1-5 for Mandarin. We provided a short introduction to Chinese names and romanisation systems in Appendix 1. We manually entered the records that the package failed to translate, and stored pronunciations and tones in separate columns. Raw name list, cleaned data and codes are available from a University College London Research Data Repository [19].

Of the 771 included names, a large majority (97.7%) had a 3-character full name, most (83.9%) had a romanised English forename, and most names were given based on Cantonese (97.4%) (Table 1).

We compared how closely three different systems of romanisation (HKG-romanisation, Jyutping, Pinyin) represented the original Chinese characters, in terms of uniqueness, which provided some information on the utility of these systems for balancing sensitivity and specificity of linkages.

Our sample of 771 individuals all had unique Chinese full names and shared 123 unique surnames (Table 2). The top five most frequent surnames accounted for over 30% of surnames. HKG-romanisation resulted in 152 unique surnames with 29 extra surnames than Chinese. Both Jyutping and Pinyin reduced the number of unique representations of names. The HKG-romanisation system represented different Chinese characters using the same codes, for example, “Chiu” is used to represent (Jyutping: Ziu6, Pinyin: Zhao4) and (Jyutping: Jau1, Pinyin: Qiu1). The HKG-romanisation also represented the same Chinese characters using different codes, for example, is represented as “Chow”, “Chau”, or “Chiau”, where Jyutping would consistently represent it as “Zau1”, and Pinyin “Zhou1”. Both Jyutping and Pinyin represented the same characters consistently, but both represented different characters using the same codes, for example, “Wong4” for both and in Jyutping, and “Yan2” for both and in Pinyin. Pinyin without tones had all the problems of Pinyin, plus the inability to differentiate tones, hence was less specific.

There were 743 unique combinations of forenames in Chinese, corresponding to 641 unique Jyutping representations, and 679 Pinyin representations (Table 2). In our data of predominantly Cantonese-based names, we observed more unique Pinyin forename combinations than Jyutping. Similar to naming clusters in other languages, Hong Kong people tend to use variations of similarly sounding names. For example, forename “Paak3Hei1” occurred six times, but each represented a unique Chinese forename. Pinyin would represent these six names in four different ways. In this case, Pinyin becomes more specific than Jyutping in differentiating people in our sample. We expect the specificity of Pinyin to be lower in a majority Mandarin-speaking sample, or in a larger population with more variations in names. Both Jyutping and Pinyin seem to be able to increase linkage quality in dealing with pronunciation-based errors and rare names.

We compared how different name encoding strategies impact on blocking. Since forename and surname combination almost uniquely identifies every individual in our dataset, we only used blocking on surnames as an example. We evaluated whether different encodings of the same surname would be put into the same block and produced recall and precision statistics based on the ground truth (Chinese surnames). Blocking based on full HKG-romanisation performed the worst in terms of recall at 68.8%, whilst both Jyutping and Pinyin (with tonal information) achieved over 95% recall, at comparable or superior precision (Figure 2). Full comparison of other blocking rules is described in Appendix Table 1 and 2 (Appendix 2). Block rules 5,6 and 8,9 compares blocking using first two characters of Jyutping and Pinyin respectively with and without tonal information. Incorporating tonal information improved precision in both cases.

A further promise of tonal representations of Chinese names is the potential of using tones to impute missing characters. Frequencies of tonal combinations of Jyutping and Pinyin names likely follow a Zipf’s distribution [20], where a few combinations represent most names and a long tail of combinations have very few counts. In our sample, the top 10 tonal combinations represented 38.3% names in Jyutping and 45.0% in Pinyin (Figure 3). By further incorporating vowel information, statistically estimating missing tonal information should help calibrate non-tonal romanisation systems. For example, in a 3-character name where the first 2 characters have the Pinyin tone “2-3” and the last character is missing, we can expect it is at least twice as likely for the third character to have a “2” tone than a “4” tone. Combining tonal information with frequency of vowel combinations for each character, we could be potentially re-encoding romanised Chinese names. Further work using a large-scale Chinese names database would contribute to this regard.

We demonstrate that both Jyutping and Pinyin are promising alternatives to representing Chinese characters, accounting for pronunciation-based errors, rare names and tonal information, compared to HKG-romanisation or non-tonal Pinyin.

For data linkage researchers with access to Chinese names along with HKG-romanised named, using alternative encodings such as Jyutping and Pinyin would substantially improve blocking performance. Contextual knowledge on data sources can help in choosing appropriate blocking and linkage strategies. For example, in the 1880-1990 US linkage, Chinese names in the US Census are more likely to be based on Cantonese than Mandarin. Blocking strategies based on Jyutping surnames may therefore increase efficiency of linkage by identifying more true matches, but will increase the number of missed matches, compared to Pinyin or other romanisations. As romanised Chinese names are ascribed by immigration officers in this particular setting, Jyutping will represent forenames more precisely and sensitively than Pinyin for this linkage.

Phonetic encodings can be used independently to deal with missing data or used to calculate probabilistic weights. We recommend including each term of the character and phonetic encoding separately to provide more flexibility in weight adjustment. The above recommendations are limited to datasets not already encoded. We are not aware of any approaches that could consistently re-encode tonal languages, especially when tonality of characters is not captured in romanised encoding. Using frequency-based methods to probabilistically assign re-encoded characters could be possible, by incorporating character position, vowels, and tonality. However, this approach may require strong assumptions on the distribution of the tones in any established dictionary to be the same as the sample. Future work should explore the feasibility and contexts in which re-encoding is suitable.

Developing a language-specific pre-processing and romanisation approach requires leadership and skills from people who speak these languages. We ask database owners in developing countries or former colonial regions (such as Hong Kong), where English remains the dominant language in which these databases are operating, to consider how names can be best represented and how tonal information can be retained in their databases. As for developed countries, tone-sensitive romanisation systems provide more flexibility in developing linkage strategies and could improve linkage quality for minoritised ethnic populations and migrants. Operationally, with vast advancement in voice-to-text transcription, asking individuals to say their name in their mother tongue may be a simple way to record extra information that is conducive to data linkage. This is relevant for other tonal languages, as well as character-based non-tonal languages.

Collecting, preserving and utilising people’s names in their original languages, or alternatively standardised romanisation systems, is ethically and socially pertinent and may support the development of language-specific pre-processing and linkage strategies that result in more inclusive research data that better represents the targeted populations.