The Intertemporal Evolution of Agriculture and Labor over a Rapid Structural Transformation: Lessons from Vietnam

We combine nationally representative household and labor force survey data from 1992 to 2016 to provide a detailed description of rural labor market evolution and how it relates to the structural transformation of rural Vietnam, including the agricultural sector. Our study adds to the emerging literature on structural transformation in low-income countries using micro-level data and helps to answer several policy-related questions. We find limited employment creation potential of agriculture, especially for youth. Real wage convergence across sectors and rural-urban regions has gone hand-in-hand with increased diversification of the rural economy into the non-farm sector nationwide and rapid advances in educational attainment in all sectors’ and regions’ workforce. Minimum wage laws seem to have played no significant role in increasing agricultural wages. This enhanced integration also manifests in steady attenuation of the longstanding inverse farm sizeyield relationship. Farming has remained securely household-based and the farmland distribution has remained largely unchanged. Small farm sizes have not obstructed mechanization nor the uptake of labor-saving pesticides, consistent with factor substitution induced by rising real wage rates. Rice yields increased rapidly in the earlier years, but more slowly over the past decade or so. As rural households rely more heavily on the labor market, human capital accumulation (rather than land endowments) have become the key driver of improvements in rural household wellbeing.


The Intertemporal Evolution of Agriculture and Labor over a Rapid Structural
Transformation: Lessons from Vietnam

Introduction
Over the past three decades, Vietnam has undergone one of the most rapid structural transformations of any low-income agrarian nation in history. Since the Doi Moi reforms initiated in late 1986 with the objective of creating a robust, socialist-oriented, market economy out of what had been a fully centrally-planned one, Vietnam has consistently experienced well-above-average real GDP growth of 4-8 percent annually. 5 From 1992the earliest year for which nationallyrepresentative survey data are availableto 2016, per capita real gross national income (GNI) more than tripled, turning Vietnam from one of the world's poorest countries, at less than US $500/year (in constant 2010 dollars) to roughly $1700, while the share of GDP in agriculture fell from 34% to 16% and the share of the workforce employed in agriculture fell from 68% to 42% (World Bank 2019). Put differently, in 1992 Vietnam's income level and economic structure was strikingly similar to present day Liberia. Today it is a solidly (lower) middle-income country enjoying comparatively rapid economic growth and poverty reduction. Starting as one of the poorest, most agriculture-intensive economies with still-above-average population growth, Vietnam provides a valuable case study for understanding the structural transformation process among low-income countries and its implications for agriculture and labor markets. Since the global poor disproportionately reside in rural areas and work (at least part-time) in agriculture, understanding this process is essential to developing viable poverty reduction strategies in today's low-income agrarian economies.
Our study is motivated by several current policy concerns in developing countries. One, does structural transformation lead to regional specialization, with rural areas concentrating in farming and losing workers who leave for cities? Do we see rapid transition into the non-farm sector within rural areas? To what extent does an increasingly productive farming sector absorb young workers in low-income countries (Losch 2016, Mueller andThurlow 2019)? Two, do real wages for farmworkers converge to those earned by non-farm workers, and do rural wage rates converge to urban rates? Do minimum wage laws enacted primarily for urban non-farm workers seem to bind in the agricultural sector (Belman andWolfson 2015, Bhorat et al. 2017)? Three, does structural transformation lead to family farm consolidation, mechanization, and the displacement of workers as farms grow larger and more mechanized (Mrema et al. 2008, OECD 2016? Does the commonly observed inverse farm size-productivity relationship in developing countries attenuate, suggesting more competitive and integrated rural factor and output markets and less rationale for agricultural policies favoring smaller farms? Four, is structural transformation associated with rapid increases in well-being for households that remain in rural areas (Barrett et al. 2017, Christiaensen et al. 2018)?
This paper offers descriptive evidence on the structural transformation process that Vietnam has undergone. We combine nationally representative household data (Vietnam Household Living Standards Surveys, VHLSS) from 1992 to 2016 and labor force survey (LFS) data from 2007 to 2016 in order to provide a detailed description of rural labor market evolution and how it relates to the transformation of Vietnamese agriculture. This paper thereby adds to the emerging literature on structural transformation in low-income countries using micro-level data.
Most studies on structural transformation rely on macro-level data (Timmer 2002, 2009, Rodrik 2013, Dercon and Gollin 2014, Gollin 2014, Gollin et al. 2016, Rodrik 2016, Diao et al. 2017, while micro-level data shed light on transformation within one country and unmask potential heterogeneity in the growth process (McCullough 2016(McCullough , 2017. 6 To our knowledge, our study covers a longer period (24 years) than do prior studies using microdata.
For any of several reasons, the Vietnamese case might not provide an apt lens through which to view today's low-income agrarian economies. Vietnam has retained an extraordinarily strong state while shifting from a centrally planned to a market economy. Its proximity to the east Asian boom economies of China, Korea, and Taiwan, along with normalization of relations with the United States, has enabled buoyant export growth, 7 which is perhaps less accessible to other low-income agrarian nations. The country has a high population density, but rapidly slowing population growth, leading to an aging population. The central government invested heavily in education and health, leading to educational performance and health indicators more characteristic of upper-middle-and high-income countries. Nonetheless, the Vietnamese experience can be instructive for today's low-income agrarian economies, in part because its experience contrasts with some common narratives of what might be inevitable as such economies undergo structural transformation.
We have four main findings. First, Vietnamese households have diversified out of agriculture, manifest not only in decreasing shares of farming households, but also in a decline in the agricultural labor force within farming households, and in the agricultural income share of rural households. 8 We also observe uneven structural transformation across regions, with sharper reduction in the share of agricultural labor forces in more urbanized Red River Delta and Southeast 6 See also the papers in recent special journal issues on the topic of structural transformation, such as Agricultural Economics (48, S1; 2017), World Development (105;, or Journal of Development Studies (54, 5; 2018). 7 For example, McCaig and Pavcnik (2018) demonstrate that the US-Vietnam Bilateral Trade Agreement boosted labor productivity in manufacturing, inducing a reallocation of workers into formal manufacturing. 8 Throughout this paper, agricultural income is referred to agricultural revenue or gross income from agriculture with no deduction of input costs. regions. Second, real wages have increased rapidly in both the farm and nonfarm sectors, seemingly driven by rapid advances in educational attainment and not by changes in statutory minimum wage rates. The nonfarm sector has seen significantly higher real wage growth than the farm sector and the inter-sectoral wage gap has widened, although this may reflect selection on human capital (Coxhead, Vu and Nguyen 2016). Increasing employment in higher-productivity nonfarm sectors points to a successful structural transformation, which contributes to the overall economic growth (McMillan and Rodrik 2011). And real wage growth faster than overall GDP expansion indicates an inclusive growth process benefitting workers disproportionately. Third, the rapid structural transformation does not lead to family farm consolidation. 9 Family farm size remains small and the land distribution changed remarkably little over two-plus decades of rapid rural transformation. Nevertheless, mechanizationmainly through rental marketsand the use of labor-saving inputs like pesticides has grown steadily, likely driven by rising labor costs and farmers' improved access to finance as farm productivity grew. Rice yields continued increasing.
In line with more efficient factor and output markets, over time the commonly observed inverse farm size-productivity relationship has attenuated. Fourth, as rural households rely more heavily on the labor market rather than agriculture, human capital accumulation plays an increasingly more important role in household well-being. In contrast, land endowment becomes less strongly associated with per capita consumption expenditures over time, further underscoring the transition from an agrarian economy where sector-specific assets such as land are the main determinants of income to a more modern economy based more on the returns to accumulated human capital. McCaig and Pavcnik (2017) and Tarp (2017) also describe structural transformation in Vietnam. Using the 1990s-2000s nationally representative household survey and population 9 Our data do not provide information on commercial farms. The farm size in this paper refers to family farm size. census data, McCaig and Pavcnik (2017) document that, as the nonfarm sectors provided more job opportunities, people moved out of farming, driving up agricultural real wages and shrinking the rural-urban wage gap between agricultural and nonagricultural sectors. 10 Tarp (2017) also describes the rural transformation, relying on the 2006-2014 Viet Nam Access to Resources Household Survey (VARHS) data. VARHS focuses on rural areas and is not nationally representative, however.
Our study builds on these two studies in at least four aspects. First, we use a longer duration of nationally representative data to document the evolution of rapid structural transformation over 24 years in Vietnam. Second, we provide a more detailed description of rural, and especially agricultural transformation. Third, we complement the nationally representative household survey data with LFS data to provide a more detailed description of rural labor markets and a cross-check on the household survey data. Fourth, we show important, policy-relevant findings beyond those from the existing studies, including the increasing role of human capital (and declining role of land holdings) in determining welfare of rural households, the attenuation of farm size-productivity relationship over time, as well as increased role of machinery in farming.
The paper proceeds as follows. Section 2 describes the data. Section 3 describes the evolution of rural and agricultural labor markets. Section 4 describes the evolution of the agricultural sector. Section 5 explores the evolution of well-being among rural households. Section 6 concludes.

Data
The data that we use come from two nationally representative data sets, the ten-round VHLSS from 1992 to 2016 and the six-round LFS from 2007 to 2016. Such rich, nationally representative descriptive analysis is uncommon during an extended period of rapid growth. For some analyses, we merge the data from household and commune surveys to construct five rural household panels: VHLSS 1992/1998, VHLSS 2002/2004, VHLSS 2006/2008 Although the GSOV claims the data are representative of the country, households in surveyed communes are on average better off than households in other communes (Hansen and Le, 2013). This has created some concern that the sample selection deviated from a fully random approach at some point(s), although these differences are not substantial. are considered workforce, whereas the latter two are considered out of the workforce. Appendix Table A2 summarizes the number of individuals aged 15 and above surveyed in each round. 14 A household is considered as an agricultural/farming household or a household engaged in agriculture if any member had been worked in agriculture (including forestry and aquaculture) during past 12 months before each survey time.

Evolution of agricultural and non-farm employment patterns
Consistent with household-level results from VHLSS, individual-level LFS data reveal similar patterns (Appendix Table A3). The share of individual workers employed in agriculture declined from 48.4% to 39.4% from 2007 to 2016. The fact that these shares are far less than the proportion of agricultural households signals that even agricultural households have long diversified their earnings portfolios across sectors (on which, more below), as is true in sub-Saharan African low-income agrarian nations as well (Barrett et al. 2001). Such a pattern was also observed in today's high-income countries during their structural transformations.
This increasing diversification into non-farm employment is perhaps best seen by looking at individual household members' employment in farming (as farmers or farm workers) within farming households. Panel A of Another way to grasp the sharp, nationwide transition of rural households towards nonfarm, and especially wage labor, is through the shares of total household income arising from agricultural versus wage earnings. As shown in Figure  best represents the dramatic structural transformation of the rural Vietnamese economy over this period, as agriculture has become less important as an employer and as a source of income for households even as its productivity has increased sharply and the use of modern inputs that boost labor productivitye.g., fertilizers, improved seeds, machinery, pesticideshas increased rapidly (see section 4).
Appendix Table A4 summarizes the evolution of the median shares of agricultural income and wage income, by region. All regions share the common trend that the income share from farming continuously reduces while the share from wages continuously increases. In the South East and the Red River regions, a median rural household only had 2.3% and 7.8% income from farming in 2016, respectively. Rural households' lower dependence on farming is most pronounced in the Red River region, which saw a sharp reduction of median income share from farming, from 41.1% to 7.8%, during the merely 14-year period. This is consistent with the rapid urbanization surrounding the metropolitan area of Hanoi. This resembles patterns in high income countries, where even within the farm sector, most households earn more net income from wages than from agriculture.
Vietnamese agriculture is traditionally dominated by farmers who cultivate their own land.
This pattern remained largely unchanged during the structural transformation. Among the population employed in agriculture, only 5.1% were hired workers in 1992. That share increased very slowly, to just 8.0% by 2016. 15 Farming has remained concentrated among households with access to land who employ predominantly family labor on the farm. The corollary to this is that Vietnam has not seen the emergence of a farmworker class. 16 Table 3 reports the agricultural labor force composition by age, gender, and education levels from 2007 to 2016. Panel A shows that the shares of the agricultural labor force in age groups below 50 years old (15-20, 20-30, 30-40, 40-50) all declined over this period while older age groups (50-60, 60-70, and 70+) accounted for increasing shares of the agricultural labor force.
This observation may partly reflect the aging of the overall labor force in Vietnam. However, the aging in agricultural labor force is more severe than in the overall labor force. The share of labor force younger than 50 years old fell from 75 to 60 percentage points (a 20% reduction) for agriculture only, in comparison with a 10% reduction (from 81 to 73 percentage points) for the overall labor force. The aging of the national and agricultural labor force is in line with broader demographic patterns within the Vietnamese population, as shown in Appendix Figure A1 Table 3 indicates that the education level (i.e., highest level of education completed) of the agricultural labor force steadily increased during this period, consistent with improvements in education for the overall labor force in Vietnam. From 2007 to 2016, the share of the agricultural labor force that had never attended school fell from 36% to 22%, and the share of workers with lower secondary education or above dramatically increased from 15% to 49%.
This reflects the rapid rise in education nationwide. Moreover, with merely 10% possessing upper secondary education and 3% with college education or above, the agricultural labor force still had much lower educational levels than the overall labor force, within which 16% of workers had completed upper secondary school and another 16% college or above in 2016. The government's massive investments in education were clearly translating into a better educated workforce, with those gains accruing in all sectors, but disproportionately in the non-farm sectors. This is not surprising given that younger and better-educated populations have higher propensity to migrate from agriculture to nonfarm sectors (Hicks et al. 2017;Young 2013). Migration also contributes to aging of agricultural labor force which is commonly observed globally.
The gender composition of the agricultural labor force was virtually unchanged during this period (

Growth in real agricultural wages
The   Table A7. Women consistently received lower wages for the same tasks as men, pointing to gender inequality in returns to agricultural labor.
Besides gender differences, these results also show little evidence of systematic changes in the harvest season wage premiums and suggest that any changing task composition likely has little role in driving observed increases in real agricultural wages.
How do the rapid increases in wages relate to increases in wages in other sectors? Figure  help to explain increasing rates of urbanization and substitution away from agricultural towards nonagricultural employment. Comparing panels (a) and (b) of Figure 3, we observe similar trends in both urban and rural areas. This is a pattern of intersectoral differentiation in the returns to labor, not a rural-urban difference. Continuous transition of the labor force from agriculture to nonfarm sectors, combined with increasing industrial sector-agriculture wage ratio, features a successful structural transformation in which "high-productivity employment opportunities (in nonfarm sectors) have expanded and structural change has contributed to overall growth" (McMillan and Rodrik 2011). Table 4 examines the evolution of urban/rural real wage ratios overall as well as urban/rural real agricultural wages specifically. Panel A of Table 4 indicates a rapidly shrinking urban-torural wage ratio for both men and women, suggesting spatial convergence in labor markets. Panel B finds that agricultural wages evolved at similar rates in both urban and rural areas, with slight spatial convergence in female wage rates, but none for men. These results suggest agricultural labor market integration across regions, which is likely driven in part by the widespread seasonal rural-urban migration (de Brauw and Harigaya 2007). Migration from rural to urban areas during agricultural lean seasons tend to push up rural wages and thus lower urban/rural wage ratios. 19 To examine whether particular regions may be driving trends in agricultural wages, Appendix Table A8 reports median real agricultural wages by gender as well as across the six main geographic regions. While there is substantial heterogeneity in wage levels across regions, all regions experienced similarly large increases in real agricultural wages over this period, with parallel changes across genders within each region. Consistent with findings in Table 4, wages appear to converge across regions over time with the coefficient of variation across regions dropping from 0.39 in 1992 to 0.13 to 2010 for men and 0.22 to 0.17 for women over the same interval.
One factor driving these agricultural wage increases over this period may be increasing minimum wages imposed by the national government under the general Labor Code. Minimum wages vary by region and sector and are typically adjusted annually. Unlike in some countries, minimum wage requirements apply to farms, households, cooperatives, in short to any individuals or organizations who employ workers. But enforcement is widely understood to be spotty and thus it is unclear how much compliance there is. One might naturally suspect that agricultural wages fall below the required minima or that the minimum wage rates set by government bind for farms and firms. But especially if minimum wages constrain agricultural employers, then minimum wages might help boost real wages, both by directly inducing higher wages for agricultural workers as well as indirectly, by increasing reservation wages throughout the economy.
Perhaps surprisingly, median and mean wage rates in agricultural consistently exceed the minimum wage rate. 20 As shown in Panel A of Appendix Table A9, in 1992 average agricultural wages for both men and women were below the relevant minimum wage. But from 1998-2016 agricultural wages have consistently exceeded the minimum wage rates, by 17-119%, without any clear time trend. In Panel B, we see that the percentage of individual-specific wage rates that fell below the district-and-year-specific minimum wage has risen from 6.1% to 11.3% in the nonagriculture sectors and has risen significantly, from 14.5% to 28.4% nationwide, 2012-2016, in agriculture. So although mean and median agricultural wages steadily exceed minimum wages, it 20 There are four minimum wage "regions" in Vietnam into which each (sub-provincial) district is categorized. Before 2012, there were separate minimum wage schedules for domestic and foreign firms. In 2012, the two were merged into a single schedule. The minimum wage rates ('000 VND) for four regions for 2012, 2014, and 2016 ranged from (1400,1900,2400) to (2000,2700,3500), reflecting differential rates of minimum wage growth across space.
does appear that there has been increasing dispersion in agricultural wages, with the lowest wages not keeping pace with increases in region-specific minimum wages, especially among women (38.9% versus 22.9% among men in 2016) and in the more agricultural regions (those other than Red River Delta and South East). Minimum wage laws do not seem to be driving growth in real agricultural wages since noncompliance rates have been increasing in agriculture nationwide.

Evolution of the Agricultural Sector
Vietnam experienced a rapid transition over the 1992-2016 period, with large-scale movement of workers into non-farm employment, in rural as well as urban areas, and at sharply increasing real wage rates. What sort of transitions happened in the agricultural sector during this time, in particular do we see evidence of family farm consolidation due to labor exits, labor-saving factor substitution due to rising real agricultural wages in response to intersectoral labor market integration, and any erosion of small farms' competitiveness within Vietnamese agriculture?

Family farm size distribution
Does family farmland become inevitably consolidate during a rapid rural transformation? The answer in the Vietnamese case is clearly no. Vietnamese agriculture rests on very small farm units; that has remained unchanged throughout the structural transformation. As shown in Appendix  Small farms have not ignited a significant increase in land rental markets either. As shown in Appendix Figure A4, the proportion of farm households that rent land in or out has remained low and relatively stable, at 10% or less throughout the period. We do see some modest convergence between reports of renting in and renting out land, which might reflect survey respondents' increased willingness to report renting out land over time as the market-orientation 21 See Tarp (2017) for a more detailed description on land policy.
of the economy became more firmly established, or could reflect the sectoral outmigration of workers from farm families, leading to small-scale rentals.

Mechanization
With small farm sizes, one might naturally expect that mechanization rates would have remained

Agrochemicals use
Agricultural modernization commonly involves increased use not only of machinery, but also of agrochemicals, both chemical fertilizers and pesticides (which include fungicides, herbicides, and 23 Promoted by the strategy for investing in agricultural mechanization under the centrally planned economy, tractor adoption rates in Vietnam reached moderately high levels by 1980 (close to 30 percent of the entire country and even higher in southern Vietnam). Such exposures to mechanization likely have contributed to the resurgence of tractors in the 1990s. Most low-income countries have not been exposed to such high-level of machinery adoption as Vietnam and are likely faced with more demand constraints (Takeshima et al. 2021).
insecticides). Fertilizer boosts crop and weed growth, stimulating demand for labor, while pesticides typically reduce labor demand by substituting for labor-intensive methods of pest eradication. Unlike machinery, there are no economies of scale to fertilizer or pesticide use. So the dominant drivers of agrochemicals uptake will typically be the profitability of use, which is driven both by crop and input prices and by real wage rates.
As reflected in Figure 7, we see different patterns of use between fertilizers and pesticides.
Labor-saving pesticide use has followed a pattern similar to that of machinery. In the earliest years of the VHLSS surveys, agrochemicals use was sharply increasing in farm size. The largest farms were more than twice as likely to use pesticides as the smallest farms. That relationship attenuated

Land productivity
24 Pesticides can also pose human health risks, especially to agricultural workers. Research in southeast Asian rice systems has previously established increased human health costs associated with expanded uptake of pesticides (Antle and Pingali 1994) and recent survey data from sub-Saharan Africa establish similar correlations as pesticide use increases in some parts of that region (Sheahan et al. 2017). Ecological and health costs of excessive pesticide use have been a growing concern in Vietnam ( Dasgupta et al. 2007). We leave exploration of the pesticide-health relationship in Vietnamese agriculture to future work.
Along with labor, the main input in agriculture is land. And in Vietnam, the main crop is rice.
Therefore, understanding the evolution of rice yields provides a useful indicator of the evolution of agricultural land productivity more broadly. It is possible that higher real wages lead farmers to apply less labor, leading to lower yields unless they compensate by using other inputs. We have just seen that Vietnamese farmers' use of machinery and agrochemicals increased significantly over the 1992-2016 period. Did this offset any adverse yield effects arising due to the higher costs of agricultural workers? It has long been observed in developing country agriculture that smaller farms are more productive per unit area cultivated than larger ones, on average (Chayanov 1926(Chayanov /1986Sen 1962;Berry and Cline 1979;Carter 1984;Barrett 1996;Benjamin and Brandt 2002;Barrett, Bellemare, and Hou 2010;Carletto, Savastano, and Zezza 2013). The dominant narrative behind the inverse relationship has historically been that multiple market failures can generate a size-productivity gradient even if the underlying technology of agricultural technology exhibits constant returns to scale (Feder 1985;Barrett 1996). The evidence of such an inverse farm size-productivity relationship has often justified land policies favoring small landholders and deterring farm size expansion, as well as agricultural credit policies to promote smallholder access to commercial inputs. 25 However, as a low-income agrarian economy undergoes rapid structural transformation, do factor markets for agricultural labor and machinery become more active, driving up real wages and attenuating the inverse relationship? Otsuka (2013) and Foster and Rosenzweig (2017)  To answer this question, we investigate the evolution of the inverse farm size-productivity relation using VHLSS panels from 1992/98 to 2014/16. We first estimate a rice yield equation using the five panels separately: ln = + 1 ln ℎ + 2 + 3 + , 25 A recent literature suggest that the inverse farm size-productivity relationship appears attributable to measurement error in crop output in multiple data sets from Africa (Gourlay et al. 2017, Desiere and Jolliffe 2018, Abay et al. 2019). Since there is no reason to expect an intertemporal trend in such measurement error in the VLSS data and yet we see a clear trend in the size-productivity relationship parameter, we suggest that the attenuation we observe is at least partly real, not merely an artefact of measurement error.
where ln is log rice yield (in kilogram per hectare) for farm/household i, and year t; is a household fixed effect which captures time invariant household and location-specific effects such as land quality and weather; ln ℎ is log rice planting area (in hectare); is a vector of householdspecific time-varying characteristics; is a year dummy which captures period-specific fixed effects (including interest rates, prices, and wages) that are common across communes; and is a random error term. The coefficient of interest, 1 , reflects the elasticity of rice yield with respect to planting area. A negative and statistically significant 1 estimate supports the presence of an inverse relationship. If such a relationship lessened over time, the absolute value of 1 will be smaller in a later panel than in an earlier panel. If such relationship is reversed, the 1 estimate will be positive. Table 5 reports the regression results of equation (1) Appendix Tables A10 and A11, are similar to those reported in Table 5, showing a significantly decreasing inverse size-productivity relationship for both spring and autumn rice over 1992/1998 and 2014/2016.
This change is associated with rising real wages and increasingly active machine rental and agricultural labor markets in rural Vietnam. As a result, the long-standing, labor-based productivity advantage assumed to exist among smaller farmers appears to have diminished altogether by the latter part of the period. Indeed, as real wages keep increasing, the inverse relationship may be reversed, leading to increased land concentration among farmers increasingly likely to employ machinery, without adverse effects on aggregate food production or prices.

Diversification of agricultural production
Just as many observers expect structural transformation to lead to farm consolidation, so too might one naturally expect rising incomes and enhanced market access have naturally led to diversification of agricultural production over time. We can explore this hypothesis by constructing a Herfindahl-Hirschman Index (HHI) for each farm household as where i indexes the farm household and j indexes each of eight categories of agricultural outputs: ordinary rice, glutinous rice, high-quality rice, other food crops, industrial crops, fruits, aquaculture, and livestock.
is the value share of output j of the total output value for farm i.
HHI ranges from 0 to 1, with a higher value indicating lower diversification.
The top panel of Table 6 summarizes the HHI of agricultural output over the 2002-2016 period. Remarkably, the sector overall has exhibited decreased production diversity relative to the early 2000s. As seen in the bottom two rows, this effect is especially pronounced among the smallest farms. The largest quintile of farms, by land size, have seen some diversification.
Appendix Table A12 (2010) shows that migrant households tended to move out of labor-intensive rice production to more land-intensive crops due to lack of family labor for farming in 1990s. However, labor constraints may be partially relaxed as factor markets became more efficient in recent years.
Therefore, the migration-induced effects on production diversity may be lower over time. 26 26 We are not able to test this hypothesis due to data limitations.
While Vietnamese farms have not been diversifying their product mix appreciably over time, there has been a dramatic rise in farm households' reliance on markets. Appendix Table A13 reports the share of food consumption expenditure coming from own production, i.e., autoconsumption. The median share of autoconsumption of own food production dramatically decreased from 0.535 in 1992 to 0.197 in 2016. This reflects sharply increased dependence on markets to source food, even among increasingly productive farm households. Rising rural incomes lead to more diverse diets, but with more efficient food markets, smaller farms have opted to concentrate on specific crops. This may reflect market-driven specialization according to comparative advantage, or the need to specialize in order to benefit from labor-saving mechanization that exhibits economies of scale. In contrast, larger farms may resort to higher diversification as a hedge against greater price risk exposure (Bellemare et al. 2013).

The evolution of well-being among rural households
As rural households have diversified out of agriculture, how has their well-being changed over time? Figure  To understand what factors are associated with household wellbeing and income sources over time, we regress the logarithm of household per capita consumption expenditures, share of income from agriculture, and share on income from wages on land and human capital endowments, the latter measured by years of schooling of the highest educated member of a household. We control for household demographics and regional fixed effects and cluster standard errors at the commune level. Since we do not control for several key relevant unobserved variables, our regression results should be interpreted as association rather than causality. Tables 14-16. In Table 7 Columns (1)-(6) and Appendix Tables 14-15, we look at how landholding and human capital are associated with income shares from agriculture and from wages, respectively. Not surprisingly, landholding is positively associated with the household income share from agriculture and negatively associated with the share from labor earnings. Education is negatively associated with the income share from agriculture and positively associated with that from wages. The coefficient estimate relating education (years of schooling) to income share from wage increased steadily, from 0.0086 in 1992 to 0.0211 in 2016, suggesting that better educated rural households rely more on labor markets as an income source over time. The younger labor force is also better educated and more likely to migrate from agriculture to nonfarm sectors (largely through seasonal migration from rural to urban), leaving a more severely aging agricultural labor force than the nonfarm sectors. Not surprisingly, as we observed earlier, rural households now earn more from labor markets than from agriculture. Columns (7)-(9) of Table 7 and Appendix Table 16 present results on consumption expenditure. Both landholdings (owned land) and education are positively associated with consumption in all rounds. However, the coefficient of landholding becomes smaller over time, even becoming statistically insignificantly different from zero in the 2016 round, while the coefficient of education becomes larger and more significant over time. These results reinforce the earlier findings that although rural households in Vietnam have remained engaged in farming, they are increasingly dependent on the returns to human capital in labor markets and depend less today than previously on landholdings to support their well-being. Although agricultural productivity has increased sharply over time, the large improvements observed in rural well-being ( Figure 9) appear most strongly associated with improvements in human capital remunerated in labor markets increasingly integrated across sectors and space. Indeed, as shown in Appendix Figure A6, rural household expenditure is positively correlated with income share from wages and negatively correlated with income share from agriculture for all rounds. Our finding also suggests that, for rural households, human capital accumulation (rather than land endowment) is an essential means of successful transformation.

Conclusions
Vietnam's dramatic structural transformation over the past generation offers an uncommon glimpse into the path followed as a low-income agrarian economy grows rapidly. In 1992, Vietnam looked remarkably comparable to current day Liberia in terms of per capita income, share of output and employment in agriculture, reliance on rice and cassava as staple crops, etc. Today it continues to grow at a rapid rate (6-7% annually), diversifying and creating jobs quickly, and transforming into an increasingly urban and non-farm lower middle-income economy. Several key patterns of Vietnam's structural transformation merit comment as they relate to prospective futures for today's low-income agrarian economies.
First, the direct employment creation potential of agriculture, especially for youth, is limited. The agricultural labor force is slowly shrinking and aging more rapidly than is the labor force as a whole. Even farming families are diversifying out of agriculture, increasingly earning more of their total household income from the non-farm sector. Youth are increasingly well educated, enjoying a wider array of remunerative non-farm job options than their parents did.
Meanwhile, the endogenous changes in agriculture, especially mechanization and uptake of laborsavings inputs such as pesticides, relax farm households' labor constraints, freeing young people to seize non-farm opportunities.
Second, real wage convergence between rural-urban regions has gone hand-in-hand with increased diversification of the rural economy into the non-farm sector nationwide and rapid advances in educational attainment in all sectors' and regions' workforce. This enhanced integration also manifests in steady attenuation of the longstanding inverse farm size-yield relationship, which only exists when there exist multiple rural market failures. Minimum wage restrictions do not seem to explain growth in real agricultural wages. Indeed, while compliance with minimum wage laws appears quite high in the non-farm sector, noncompliance in the agriculture sector has been increasing this decade, especially in the most agriculturally dependent regions. Minimum wage laws have not prevented a widening in the intersectoral wage differential, which likely reflects differing returns to human capital, particularly educational attainment.
Third, there is no indication of significant disinvestment of households from farmland nor of significant growth in agricultural labor demand nor the growth of a farmworker population.
Indeed, the family farmland distribution has remained largely unchanged over these 24 years, as has the share of workforce earning wages in agriculture. There has been no farm consolidation and no appreciable diversification out of rice production. Although this precludes seizing economies of scale, thanks to the emergence of robust machinery rental markets it has not obstructed mechanization, nor the uptake of labor-saving pesticides. Rice yields increased rapidly in the earlier years, more slowly over the past decade. But farm households have clearly become better integrated into commercial marketing channels, as reflected in the sharp decrease in the share of food autoconsumed from home production.
Fourth, nonfarm sectors have been providing high-productivity employment opportunities, which is a driving force contributing to wellbeing improvement among rural households. As rural households rely more heavily on the labor market, human capital accumulation (rather than land endowment) is an essential means for rural households to benefit from successful transformation.
Will today's low-income agrarian economies necessarily follow the path Vietnam has taken over this past quarter century? That seems unlikely, given the many context-specific features that have guided Vietnamese development over the past generation. Nonetheless, there are important lessons to be learned from the experience of one of the world's most rapidly transforming rural economies. Barrett, C. B., Christiaensen, L., Sheahan, M., & Shimeles, A. (2017). "On the structural transformation of rural Africa". Journal of African Economies 26(S1): i11-i35. Barrett             parentheses, clustered at the commune level. The variable "Log total area of rice" centered around their sample means. * p<0.10, ** p<0.05, *** p<0.01.   The LFS sample is randomly selected in a two-stage stratification design. In the first stage, each centrally governed city or province constitutes a main stratum, which is divided into two substratums representing Population and Housing Census enumeration areas in rural and urban areas. Enumeration areas are then randomly selected using the Kish method. In the second stage, 15-20 households were selected from each sub-stratum enumeration area, yielding a sample that is statistically representative at the national, urban/rural, and six regional levels.                 . Standard errors in parentheses, clustered at the commune level. Regional dummies are included in all regressions. * p<0.10, ** p<0.05, *** p<0.01  . Standard errors in parentheses, clustered at the commune level. Regional dummies are included in all regressions. * p<0.10, ** p<0.05, *** p<0.01 Notes: The sample includes households from VHLSS 1992-2016. Standard errors in parentheses, clustered at the commune level. Regional dummies are included in all regressions. * p<0.10, ** p<0.05, *** p<0.01