Language barrier, among other social determinants, is known as an important factor predicting poorer health and barrier to care [4-10]. However, the impact of language barrier on diabetes and its ocular consequences have not previously been documented. Language barrier can be easily measured by a participant’s English proficiency during survey interview [11,12]. English proficiency during the interview is a functional measure determined by interviewers and therefore it is not subject to self-assessment bias [10,11].

In traditional logistic regression model, after controlling for important covariates and risk factors, Tamil-speaking Indians were still significantly more likely to have T2DM (OR=1.25; 95% CI: 1.04 to 1.52); and among those with diabetes, DR (OR=1.20; 95% CI: 1.05 to 1.70), VTDR (OR=1.70; 95% CI: 1.06 to 3.01) and VI (OR=1.56; 95% CI: 1.03 to 2.34) compared to English-speaking Indians. There was no significant interaction between English proficiency and socioeconomic measures (P for interaction >0.05, data not shown) and between English-proficiency and age (P=0.42). We also carried out stratified analyses by examining the associations of English proficiency (Tamil versus English) with T2DM, DR, VTDR, and VI, stratified by country of birth, education or income. The relationships between English proficiency and T2DM were slightly strengthened among the Singapore-born Indians, those with secondary education level or higher, and those with an income level<S$1000 (all with P<0.001). The relationships of English proficiency with DR, VTDR and VI were slightly strengthened among the Singapore-born Indians, those with primary education level or less, and those with an income level≥S$1000 (all with P<0.001).

Table ​Table22 shows the findings of our Oaxaca decomposition analyses for T2DM, DR, VTDR and VI. Tamil-speaking Indians had a higher prevalence of T2DM than English-speaking Indians, by 11.6 percentage points. In the analyses stratified by age groups, Tamil-speaking Indians consistently had a higher prevalence of T2DM (data not shown), Two thirds of the difference (8.4/11.6) was attributed to the differences in the groups’ individual characteristics (“explained” component) and the rest could not be explained by the difference in individual characteristics (“unexplained” component). Age had the biggest contribution to the “explained” component: if age distributions in the two groups were similar, the difference in prevalence of T2DM would have been predicted to reduce by 3.4 percentage points. By contrast, if gender distributions in the two groups were similar, the difference in prevalence would have been predicted to even increase by 1.9 percentage points.

Among the patients with T2DM, Tamil-speaking Indians were more likely to have DR (by 6.1 percentage points) and VTDR (by 4.9 percentage points) than English-speaking Indians (Table ​(Table2).2). 50.8% (3.1/6.1) of the difference in DR prevalence and 24.5% (1.2/4.9) of the difference in VTDR were attributed to the differences in the groups’ individual characteristics (“explained” component) and the rest could not be explained by the difference in individual characteristics (“unexplained” component). Duration of diabetes and socioeconomic status (including income and housing type) had substantial contribution to the “explained” component for both DR and VTDR prevalence.

Among the patients with T2DM, Tamil-speaking Indians were twice as likely as English-speaking Indians to have VI, giving a gap of 17.0 percentage points (Table ​(Table2).2). Around 50% (8.3/17.0) this difference was attributed to the differences in the groups’ individual characteristics (“explained” component). Age and socioeconomic factors (including reading literacy and income) had substantial contribution to the “explained” component.

The origins of the “unexplained” language-related disparities are multi-factorial, and as suggested by Marmot and others, the disparities could be broadly due to material deprivation and/or the lack of capability to control life and fully participate in the society (psychosocial disadvantage) [24]. We propose two possible explanations. First, English proficiency can be perceived as a proxy measure of acculturation and reflects immigrants’ culture, social identity and political ideology [9], given that most of our participants are first or second generation of the immigrants from Indian subcontinent. In this regard, Asian Indians who speak English during interview are presumably the ones who are more adaptive to local culture and are more likely to be absorbed into the dominant society – a community that have an advantage in obtaining occupation opportunity, receiving social support, avoiding psychological stressors, and maintaining a healthy lifestyle. As a result, they may be less likely to have diabetes and its complications compared to Tamil-speaking Indians. Second, the “unexplained” disparities may be due to a lack of diabetes knowledge, medical information, patient-physician communication, and treatment adherence among those with poor language skill [6-8,25]. This view is supported by the findings from the United States that language ability can directly influence access to health care and has impact on health among the Hispanic populations [6-8]. Finally, our findings may be attributable to a “healthy migrant effect” (i.e., the new immigrants were generally healthier than the local residents), but our stratified analyses showed that this language-related disparity was also seen in Singapore born Indians. Further research is needed to evaluate and identify ways in which language barriers affect diabetes management and DR care, and to assess the cost effectiveness of language-specific health improvement programs and linguistic service among this heterogeneous population. Geographic condition is unlikely an explanation, given that the two communities were living in the same areas (totaling 42.6 sq mile) and there were no transformational barriers across different districts [16].