Indonesia Breaks the Usual Multigenerational Persistence Pattern

Parent-child education correlations are often used as a shorthand for long-run mobility. That shortcut is risky if advantages persist through grandparents, marriage markets, or other family channels that do not show up in a two-generation estimate. This paper asks whether the high multigenerational persistence found in many high-income settings also appears in Indonesia, a setting where parent-child mobility is low and financial constraints are more binding.

The answer is more subtle than the usual story. Using the Indonesian Family Life Survey and Census validation, the authors find high two-generation persistence in education but less three-generation persistence than a mechanical extrapolation from parent-child correlations would predict.

Core Contribution

The paper’s main contribution is to connect a theoretical result to a country case where the sign of the grandparent coefficient reverses the pattern seen in much of the developed-country literature. In a standard three-generation regression, the coefficient on grandparents, conditional on parents, is positive only when the observed grandparent-child correlation exceeds the product of the parent-child correlations. The paper uses this mapping to show why a positive coefficient is not a universal feature of multigenerational transmission.

In the Indonesian sample, the coefficient is negative. That means higher grandparental education is associated with lower grandchild education once parental education is held fixed. The authors interpret this not as grandparents harming outcomes, but as evidence that the dynamic structure differs: direct parental resources and credit constraints can raise parent-child persistence while making longer-run persistence decay faster.

Technical Approach

The empirical core uses five waves of the IFLS from 1993, 1997, 2000, 2007, and 2014. The restricted sample contains 8,278 third-generation individuals born between 1975 and 1988, with education observed for both parents and all four grandparents. Education is measured as completed years of schooling by age 25, with consistency checks across survey waves and Census comparisons for cohort trends.

The theory compares a simplified Becker-Tomes model, a latent factor model, and a more general Becker-Tomes formulation. The key object is the grandparent coefficient from a regression of child education on parent and grandparent education. Under stationarity, the paper uses the relationship

to clarify what the sign means. A positive value implies more three-generation persistence than the parent-child correlation alone predicts; a negative value implies faster decay.

The paper then tests two mechanisms. For credit constraints, it uses geographic exposure to the 1997-98 Asian financial crisis, measured by changes in DMSP-OLS nighttime lights between 1996 and 1998 at the pre-crisis kabupaten of residence. For marital sorting, it uses IFLS questions on whether spouses were selected by the family, and compares individual, province, and ethnicity definitions of family-based spouse selection.

Results and Analysis

The descriptive results establish the tension. Average years of education rose from roughly 1.3-2.6 among grandparents to 4.8-6.1 among parents and 9.7 among the child generation. Yet intergenerational persistence remains high: Table 2 reports a G1-G2 education correlation of 0.516 and a G2-G3 correlation of 0.638 using average parental and grandparental education. The observed G1-G3 correlation is 0.312, close to but below the predicted 0.329 from multiplying the two parent-child links.

The regression evidence is sharper. Table 3 estimates a parental education coefficient of 0.502 and a grandparental education coefficient of -0.0426 in the baseline three-generation regression. The negative sign remains under ethnicity controls, kabupaten controls, controls for zero education among parents and grandparents, and a high-school completion outcome. This is the paper’s central empirical result: Indonesia has low parent-child mobility, but the three-generation process does not imply unusually low long-run mobility.

The crisis evidence is more targeted and less uniform. In the pooled crisis specification, the coefficient of interest is negative but small and statistically imprecise. Once the authors split treated cohorts by school stage and gender, the effect is concentrated among boys: for males, the interaction between grandparent education, crisis exposure, and being born in 1979-82 is -0.1190, and the corresponding estimate for 1983-85 is -0.1490. Both are statistically significant at the 5% level. The interpretation is plausible but narrow: the crisis results support the credit-constraint channel mainly for boys at secondary-school ages.

The marital-sorting results point in the opposite direction. About 28% of women and 19% of men in the parent generation report that their parents chose their spouse. In Table 6, family choice lowers the spousal education association by -0.0465, but increases links to parents-in-law by 0.067 and 0.068. Table 7 then shows a positive interaction between family choice and grandparental education, from 0.0956 to 0.136 across definitions. Family-based matching therefore seems to add multigenerational persistence, but it cannot explain the negative baseline grandparent coefficient.

Caveats

The evidence is strongest as a careful Indonesia case study, not as a general law for developing countries. Grandparental education is often reported by household heads rather than by grandparents themselves, which may attenuate estimates. The crisis design also depends on nighttime lights as a local shock measure and yields its cleanest results only for male secondary-school cohorts. The payoff is still substantial: the paper shows why two-generation mobility rankings can mislead when the underlying transmission mechanisms differ across countries.