Language and the Learning Crisis: Evidence of Transfer Threshold Mechanisms in Multilingual Reading in South India

Abstract

This paper provides evidence for a non-linear relationship between local language and English reading scores in multilingual South India. The mechanism suggests that being taught English prematurely may lead to struggles in English literacy acquisition; whereas being taught English after achieving a threshold reading level in the local language may lead to success in transitioning to English literacy acquisition. We base our findings on non-parametric and parametric regression analysis of data from assessments that were uniquely designed for South and Southeast Asian languages, and for multilingual children. Our findings help explain the global trend of increased school enrolment, with learning achievements lagging.

Acknowledgements

We are grateful to the All Children Reading partners (United States Agency for International Development, World Vision, and the Australian Government Department of Foreign Affairs and Trade) for supporting this research as part of the All Children Reading Grand Challenge for Development.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. We use the word ‘mother tongue’ for the language(s) of the home, ‘Lit 1’ for the primary literacy the child learns, and ‘Lit 2’ for the secondary literacy the child learns. In many contexts of India, Lit 1 is not the same as the ‘mother tongue’ but it is the language of the school, and of wider communication in the region. Lit 2, similarly, may be the second, third, fourth, or even later spoken language of the child, but it is the second literacy language they are acquiring. We use the term ‘L1’ for the first language when the first spoken language and the first literacy language are the same, and ‘L2’ for the second language when the second spoken language and second literacy language are the same – as is the case in most of the literature. It is important to distinguish these as the terms from the literature do not necessarily match the realities of many multilingual educational contexts of developing nations.

2. The second round of data collection did not measure oral vocabulary knowledge and included the measurement of a number of different skills, such as oral reading fluency and spelling. For this reason, the second round of data cannot be easily compared to the first and third round of data, and we rely on the latter datasets for our analyses.

3. We relied on two cross-sectional analyses rather than a single longitudinal analysis to determine the threshold mechanism because we did neither anticipate nor find sufficient variation in reading outcomes for the same individual children over the course of one year. Future research with larger sample sizes could benefit from the use of longitudinal growth models to assess the relationship between Lit1 and Lit 2 literacy skills.

4. The linguistic environment of the slums of India are complex, with wide ranging proficiency levels in multiple languages, extensive code-switching (Reddy, 2011), and ‘functional multilingualism’ (Bhatia & Ritchie, 2004).

5. We calculated Cronbach’s alpha reliability scores for all tests, and only used those tests with sufficient reliability (Cronbach’s alpha > 0.8) in our analyses.

6. Only students who scored 30 per cent or higher on the concept of print test were eligible for the decoding test and only students who scored 30 per cent or higher on the decoding test were eligible for the reading comprehension test. During the first round of data collection only 84 students out of 372 who took the decoding test were eligible for the Lit 2 reading comprehension and fluency tests, and during the third round of data collection, only 105 students out of 440 who took the decoding test were eligible to take the Lit 2 reading comprehension test.

7. We thank an anonymous reviewer for the suggestion to examine the possibility of a non-linear exponential association between Lit 1 and Lit2 skills after the threshold point.

8. For the locally weighted regression models, we used a bandwidth of 0.3. This ensured that only 30 per cent of the observations that are closest to the specific value of Lit 1 outcomes would be used to determine the correlation between Lit 1 outcomes and English outcomes. The smaller the bandwidth, the closer the estimate of the correlation will confirm the data. However, the use of bandwidths < 0.25 is not recommended because the regression function will capture random noise instead of the real relationship between Lit 1 and Lit 2 decoding outcomes (Jacoby, 2000).

9. We used the same nonlinear multivariate regression model below the threshold value, but in alignment with the conceptual model we did not find evidence for a non-linear relationship between Lit 1 and Lit 2 decoding skills below the threshold value. We did find evidence for a non-linear relationship between Lit 1 and Lit 2 decoding skills in the full sample, but this non-linear relationship is entirely driven by the data points above the threshold value. These results are available upon request.

10. We conducted an exploratory analysis to determine the robustness of the results to changes in the threshold value. Specifically, we examined the nonlinear quantitative relationship between Lit 1 decoding skills and Lit 2 decoding skills below and above the threshold for different values (ranging from 0.4 to 0.8) of the Threshold value. These results indicate that the association between Lit 1 decoding skills and Lit 2 decoding skills increases with the threshold value, which is consistent with an exponential increase in the relationship between Lit 1 decoding skills. and Lit 2 decoding skills after the threshold value. However, the sample size remains insufficient to determine the robustness of the threshold value of 0.6 to different specifications. Results of the exploratory analysis are available upon request.

11. Results are available upon request.