Dynamic value-added models of learning estimate that only a fraction of a student's achievement carries over from one grade to the next: in panel data from Pakistani public and private schools, Andrabi, Das, Khwaja and Zajonc (2011) estimate that only one-fifth to one-half of learning persists between grades, and parallel estimates of teacher- and school-effect persistence in other settings fall in a similar range. The economic and cognitive interpretation of this low persistence parameter remains unsettled: observed fade-out could reflect genuine forgetting of knowledge, failure of tests to register retained competence (including test-content misalignment and score-scale artifacts), substitution of household or school inputs away from previously boosted skills, or concavity of the education production function, and the source's own probes of these channels were inconclusive. The problem is to identify and quantify the mechanisms that generate low measured persistence of achievement — credibly separating genuine knowledge decay from measurement, input-substitution, and production-technology explanations — and thereby determine what the persistence parameter in dynamic value-added models actually measures. Each channel has been partially quantified in restricted settings, but no existing work decomposes the school-age persistence parameter itself, so the mechanism question remains open.

Background

Value-added (VA) models are the workhorse empirical framework in the economics of education for measuring the contribution of teachers, schools, and programs to student learning. In their simplest ("restricted") form, they assume that a student's lagged test score fully summarizes prior learning, so a school or teacher's contribution can be read off current scores conditional on last period's scores. The general dynamic version writes current achievement as yi,t=βyi,t−1+(inputs)+εi,ty_{i,t}=\beta y_{i,t-1}+\text{(inputs)}+\varepsilon_{i,t}, where the coefficient β\beta measures how much of past achievement persists into the current period. Estimating β\beta is statistically delicate — it is biased downward by measurement error in test scores and upward by omitted student heterogeneity, and the dynamic-panel estimators used to address these problems — Arellano–Bond generalized method of moments (GMM), which instruments the lagged dependent variable with earlier lags — require instrument validity that is often rejected in practice (Hansen's JJ test, a standard overidentification test of whether the instruments are jointly valid). Test scores themselves are typically constructed by Item Response Theory (IRT), a family of psychometric models that infer a latent ability from patterns of item responses, so the measured score is only a noisy, scale-dependent proxy for underlying knowledge.

A robust empirical regularity across many settings is that measured achievement gains fade rapidly: interventions that raise test scores by a quarter of a standard deviation often show a small fraction of that gain one or two years later, and estimates of year-to-year persistence of achievement cluster around 0.2–0.5. Jacob, Lefgren and Sims (2010) documented similarly rapid fade-out of teacher-induced learning in US data. Andrabi, Das, Khwaja and Zajonc (2011), using three rounds of the LEAPS panel (Learning and Educational Achievements in Punjab Schools, Pakistan) with dynamic panel methods that instrument for lagged scores, estimated β≈0.5\beta\approx 0.5 for English and Urdu and possibly lower for mathematics, showed that assuming perfect persistence severely biases estimated school effects, and explicitly posed the interpretation question in a section titled "Concavity, Smoothing or Forgetting?". Their candidate mechanisms were: (i) genuine depreciation of the knowledge stock (forgetting), consistent with the psychology of memory and with experimental fade-out of programs in India, Kenya, and the US; (ii) smoothing of achievement shocks — parents substitute inputs away from children who receive positive learning shocks (re-optimization toward a target), or teachers reallocate attention toward struggling children; (iii) concavity of the education production function — when you know more, additional learning costs more, so gains from a high base mechanically shrink; and (iv) test-side artifacts — skills that were genuinely retained are not displayed on later tests, or the content and scaling of tests shift across grades so that the same knowledge yields different measured scores.

The interpretation matters for policy: if fade-out is depreciation, interventions such as summer instruction or spaced schooling could counteract it; if it reflects household re-optimization or production-function concavity, there may be no remedy other than sustained environmental change; and if it is substantially a measurement or scale artifact, short-run test-score gains are far more informative about long-run human capital than they appear. The source paper's own tests of the smoothing channel were weak or insignificant (parental input responses to achievement shocks were limited; within- versus between-school comparisons did not support teacher targeting), so the authors leaned toward forgetting on external evidence while conceding that competing explanations were difficult to rule out. Later work has sharpened individual channels but not settled the decomposition: Cascio and Staiger (2012) showed that score renormalization across grades cannot fully explain fade-out; Kline and Walters (2016) and Feller, Grindal, Miratrix and Page (2016) showed that part of early-childhood fade-out reflects substitution between treatment and counterfactual care settings; Bailey, Duncan, Odgers and Yu (2017) argued that fade-out typically reflects control-group catch-up and reverting environments rather than decay of learned skill; and Kang, Duncan, Clements, Sarama and Bailey (2019), in the most direct decomposition to date, found that forgetting accounted for only about one-quarter of fade-out in early-childhood mathematics interventions. None of these, however, identifies the mechanisms behind the persistence parameter estimated for school-age learning in dynamic value-added models.

Problem Statement

Identify and quantify the mechanisms behind the low year-to-year persistence of measured achievement estimated in dynamic value-added models of learning (persistence parameters around 0.2–0.5, as in Andrabi, Das, Khwaja and Zajonc 2011). The object to be explained is the persistence coefficient β\beta on lagged achievement in the dynamic value-added equation of the Background, as estimated for school-age children in observational achievement panels; the fade-out of intervention-induced gains and the persistence of teacher or school effects are related objects documented in the literature and inform the mechanisms, but the required attribution concerns β\beta itself. An answer must credibly separate and attribute the low persistence of achievement across the candidate channels articulated in the source: (i) genuine depreciation of the knowledge stock (forgetting); (ii) test-side artifacts — knowledge that is retained but not displayed on later tests because of test-content misalignment across grades, score scaling and normalization, or noisy measurement; (iii) smoothing of achievement shocks through re-optimization by households (substitution of parental inputs away from previously boosted skills) or schools (reallocation of teacher attention); and (iv) concavity of the education production function. The answer should establish which of these channels quantitatively dominate the estimated persistence parameter in a credible data environment (with attention to the identification problems the source highlights: measurement error correlated across subjects, instrument validity, and heterogeneity in persistence across subjects and student ability), and should state what the lagged-achievement coefficient consequently measures — a depreciation rate of knowledge, a reduced-form blend of household and school behavior with a production technology, or substantially a measurement artifact — including the implications for whether short-run test-score gains are valid proxies for long-run human-capital impacts. The dominance conclusion and the resulting interpretation of the coefficient are to be stated for the data environment in which they are established; any general claim about what the persistence parameter measures, or about short-run gains as proxies for long-run human capital, must be qualified to the settings the evidence covers.

References

  1. Andrabi, T., Das, J., Khwaja, A. I., & Zajonc, T. (2011). Do Value-Added Estimates Add Value? Accounting for Learning Dynamics. American Economic Journal: Applied Economics, 3(3), 29–54. DOI: 10.1257/app.3.3.29. https://doi.org/10.1257/app.3.3.29 (full text also available as Harvard CID Working Paper No. 158)
  2. Jacob, B. A., Lefgren, L., & Sims, D. P. (2010). The Persistence of Teacher-Induced Learning. Journal of Human Resources, 45(4), 915–943. DOI: 10.3368/jhr.45.4.915. https://doi.org/10.3368/jhr.45.4.915
  3. Cascio, E. U., & Staiger, D. O. (2012). Knowledge, Tests, and Fadeout in Educational Interventions. NBER Working Paper No. 18038. DOI: 10.3386/w18038. https://www.nber.org/papers/w18038
  4. Chetty, R., Friedman, J. N., & Rockoff, J. E. (2014). Measuring the Impacts of Teachers II: Teacher Value-Added and Student Outcomes in Adulthood. American Economic Review, 104(9), 2633–2679. DOI: 10.1257/aer.104.9.2633. https://doi.org/10.1257/aer.104.9.2633
  5. Kline, P., & Walters, C. R. (2016). Evaluating Public Programs with Close Substitutes: The Case of Head Start. Quarterly Journal of Economics, 131(4), 1795–1848. DOI: 10.1093/qje/qjw027. https://doi.org/10.1093/qje/qjw027
  6. Feller, A., Grindal, T., Miratrix, L., & Page, L. C. (2016). Compared to What? Variation in the Impacts of Early Childhood Education by Alternative Care Type. The Annals of Applied Statistics, 10(3). DOI: 10.1214/16-aoas910. https://doi.org/10.1214/16-aoas910
  7. Protzko, J. (2015). The Environment in Raising Early Intelligence: A Meta-Analysis of the Fadeout Effect. Intelligence, 53, 202–210. DOI: 10.1016/j.intell.2015.10.006. https://doi.org/10.1016/j.intell.2015.10.006
  8. Bailey, D. H., Duncan, G. J., Odgers, C. L., & Yu, W. (2017). Persistence and Fadeout in the Impacts of Child and Adolescent Interventions. Journal of Research on Educational Effectiveness, 10(1), 7–39. DOI: 10.1080/19345747.2016.1232459. https://doi.org/10.1080/19345747.2016.1232459
  9. Kang, C. Y., Duncan, G. J., Clements, D. H., Sarama, J., & Bailey, D. H. (2019). The Roles of Transfer of Learning and Forgetting in the Persistence and Fadeout of Early Childhood Mathematics Interventions. Journal of Educational Psychology, 111(4), 590–603. DOI: 10.1037/edu0000297. https://doi.org/10.1037/edu0000297
  10. Singh, A. (2020). Learning More with Every Year: School Year Productivity and International Learning Divergence. Journal of the European Economic Association, 18(4), 1770–1813. DOI: 10.1093/jeea/jvz033. https://doi.org/10.1093/jeea/jvz033
  11. Bau, N., & Das, J. (2020). Teacher Value Added in a Low-Income Country. American Economic Journal: Economic Policy, 12(1), 62–96. DOI: 10.1257/pol.20170243. https://doi.org/10.1257/pol.20170243
  12. Jones, S. (2020). Testing the Technology of Human Capital Production: A General-to-Restricted Framework. Oxford Bulletin of Economics and Statistics, 82(6), 1429–1455. DOI: 10.1111/obes.12374. https://doi.org/10.1111/obes.12374
  13. Brandt, K. (2023). When Private Beats Public: A Flexible Value-Added Model with Tanzanian School Switchers. Economic Development and Cultural Change, 72(1), 159–206. DOI: 10.1086/718893. https://doi.org/10.1086/718893
  14. Das, J., Singh, A., & Yi Chang, A. (2022). Test Scores and Educational Opportunities: Panel Evidence from Five Low- and Middle-Income Countries. Journal of Public Economics, 206, 104570. DOI: 10.1016/j.jpubeco.2021.104570. https://doi.org/10.1016/j.jpubeco.2021.104570
  15. Kuhfeld, M., Soland, J., Register, B., & McEachin, A. (2024). Testing an Explanation for Summer Learning Loss: Differential Examinee Effort Between Spring and Fall. Educational Evaluation and Policy Analysis, 46(3), 581–590. DOI: 10.3102/01623737231165027. https://doi.org/10.3102/01623737231165027
  16. Hart, E. R., Bailey, D. H., Luo, S., Sengupta, P., & Watts, T. W. (2024). Fadeout and Persistence of Intervention Impacts on Social–Emotional and Cognitive Skills in Children and Adolescents: A Meta-Analytic Review of Randomized Controlled Trials. Psychological Bulletin, 150(10), 1207–1236. DOI: 10.1037/bul0000450. https://doi.org/10.1037/bul0000450
  17. Staiger, D. O., Kane, T. J., & Johnson, B. D. (2025). Why Does Value-Added Work? Implications of a Dynamic Model of Student Achievement. NBER Working Paper No. 34119. DOI: 10.3386/w34119. https://www.nber.org/papers/w34119
  18. Andrabi, T., Bau, N., Das, J., & Khwaja, A. I. (2025). Heterogeneity in School Value Added and the Private Premium. American Economic Review, 115(1), 147–182. DOI: 10.1257/aer.20221422. https://doi.org/10.1257/aer.20221422