Persistent Poverty in Rural China: Where, Why, and How to Escape?

Glauben, Thomas; Herzfeld, Thomas; Rozelle, Scott; Wang, Xiaobing Article — Accepted Manuscript (Postprint) Persistent Poverty in Rural China: Where, Why, and How to Escape? World Development Provided in Cooperation with: Leibniz Institute of Agricultural Development in Transition Economies (IAMO), Halle (Saale) Suggested Citation: Glauben, Thomas; Herzfeld, Thomas; Rozelle, Scott; Wang, Xiaobing (2012) : Persistent Poverty in Rural China: Where, Why, and How to Escape?, World Development, ISSN 0305-750X, Elsevier, Amsterdam, Vol. 40, Iss. 4, pp. 784-795, https://doi.org/10.1016/j.worlddev.2011.09.023 This Version is available at: https://hdl.handle.net/10419/149557 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by-nc-nd/4.0/ Persistent poverty in rural China: Where, why, and how to escape? Thomas Glaubena, Thomas Herzfelda,b, Scott Rozellec, and Xiaobing Wangd aLeibniz-Institute of Agricultural Development in Central and Eastern Europe, Halle, Germany bDepartment of Social Sciences, Wageningen UR, Wageningen, Netherlands cCenter for Environmental Sciences and Policy, Freeman Spogli Institute for International Studies, Stanford University, USA dCenter for Chinese Agricultural Policy, Institute for Geographical Sciences and Natural Resources Research, Chinese Academy of Sciences, PR of China Corresponding author Thomas Herzfeld Leibniz-Institute of Agricultural Development in Central and Eastern Europe (IAMO), Theodor-Lieser-Str. 2, D-06120 Halle (Saale) Phone + 49 (0)345 2928 100 Fax + 49 (0)345 2928 199 Email [email protected] This is a “pre-submission” accepted manuscript, which has been published in “World Development”. Please cite this publication as follows: Glauben, T., Herzfeld, T., Rozelle, S., Wang, X. 2012. Persistent poverty in rural China: Where, why, and how to escape? World Development, forthcoming. You can download the published version at: doi:10.1016/j.worlddev.2011.09.023 © <2012>. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://creativecommons.org/licenses/by-nc-nd/4.0/ 3 Persistent poverty in rural China: Where, why and how to escape? Abstract Using rural household panel data from three Chinese provinces, this paper identifies determinants of long-term poverty and tests the duration dependence on the probability to leave poverty. Special emphasis is given to the selection of the poverty line and inter-regional differences across provinces. Results suggest that the majority of population seems to be only temporary poor. However, the probability to leave poverty for those who were poor is differently affected by poverty duration across provinces ranging from no duration dependence in Zhejiang to highly significant duration dependence in Yunnan. The number of non-working family members, education, and several village characteristics seem to be the most important covariates. Keywords: Poverty, Asia, China, rural population, hazard analysis, dynamics JEL classification: C14; I32; R29 Acknowledgments: Financial support from the German Research Foundation (DFG) and National Science Foundation of China (Grant: 71073155/G0305) are gratefully acknowledged. 4 Persistent poverty in rural China: Where, why and how to escape? 1. INTRODUCTION Although China accounts for three quarter of poverty reduction in the (developing) world during the last three decades, rural poverty is still a critical and highly debated issue in China. Understanding the nature of poverty, its persistence, and in particular the chance of moving out of poverty is a key to derive appropriate policies targeted to reduce it. Since the beginning of the economic reforms in the late 1970s, China has experienced rapid economic growth. This macroeconomic development has been accompanied by a dramatic reduction in absolute rural poverty at the individual level (Ravallion & Chen, 2007). Chinese statistics indicate a massive decline in the number of rural Chinese in absolute poverty from 250 million in 1978 to 34 million in 1999 based on a national poverty line. World Bank (2003), using the international US$ 1 per day poverty line, reports a slightly more modest decline in poverty from 260 to 97 million over the same period. According to the most recent figures, 40 million people still live in poverty (NBS, 2009). Much of the debate on poverty, both past and current, has focused on the possibility that poverty is a condition that only affects few households. However, for those affected it matters whether they remain in poverty for a remarkable portion of their lifetime or a transitory period only. Standard poverty measures such as the head count ratio ignore such underlying dynamics. As the effectiveness of different poverty reduction measures depends crucially on the nature of poverty - persistent or transitory, gaining insights on the flows into and out of poverty is essential from a policy perspective. The more temporary poverty, the more programs that aim at stabilization of short-term income fluctuations are appropriate. If 5 poverty is more persistent, the need for measures improving the long-term labor market outcomes or social security system is stronger. 1 Previous studies provide mixed results of the pervasiveness of persistent poverty in China. Recent estimates reveal that between 20 and 25 percent of the country’s absolute poor live in persistent poverty (Chronic Poverty Research Centre, 2005, p. 87). Further, McCulloch & Calandrino (2003) use data from 1991 till 1995 collected in rural Sichuan and show that 44 percent were poor in at least one year. However, only 6 percent of the households were consumption poor in all five years. On the basis of panel data from four southern provinces, Jalan & Ravallion (1998) find that almost 60 percent of rural poverty in the three poorest provinces, Yunnan, Guangxi, and Guizhou, can be classified as persistent, but less than 20 percent in the better-off province Guangdong. Finally, a very recent analysis by Gustafsson & Ding (2009, p. 597) highlights that 6.2 percent of ethnic minorities experienced one three-year spell of poverty compared to 3.3 percent of the ethnic majority in rural China. Surprisingly, the difference between the two ethnic groups disappears when looking at the share of long-term poor on all poor households, 20 percent of minority households compared to 23 percent of the majority households. The present study goes beyond the analysis of the nature of poverty and aims to analyze determinants of poverty persistence of Chinese rural households. More specifically, we analyze factors which explain a household’s move out of poverty including the impact of time spent in poverty. Using rural household panel data over the period 1995-2004 from the three provinces Zhejiang, Hubei, and Yunnan we first apply an ordered probit model to empirically examine household, farm and regional characteristics affecting the probability that households are long-term poor. Since the probit analysis is a static concept, we additionally apply a hazard approach to empirically examine how much of the preceding time spent in poverty increases or decreases the probability to leave poverty. 6 This paper contributes to the literature on poverty in rural China in several dimensions. First, it represents an attempt to analyze the transient and persistent aspect of poverty using different poverty lines representing two different philosophies of poverty measurement. Secondly, the results obtained from the ordered probit model and the hazard model provide important complementary insights into differences in the nature of poverty across rural Chinese regions. Moreover, the econometric analysis uses quite recent household-level data, capturing a period (1995-2004) of continued economic liberalization (Wang, Herzfeld, & Glauben, 2007). Finally, the results allow to argue for regionally differentiating policies against poverty. The paper proceeds as follows. The next section discusses the choice of the appropriate poverty thresholds. The data underlying this analysis will be introduced in Section 3. Section 4 discusses briefly the incidence of poverty in the three provinces and compares the results according to the poverty threshold. The econometric analysis of longterm poverty is presented in Section 5 and Section 6 provides evidence of duration dependence on the probability to exit poverty. Finally, the paper ends with a conclusion. 2. CHOICE OF POVERTY THRESHOLD The economic literature suggests two different philosophies to measure poverty (Green & Hulme, 2005). Whereas the first concept captures a person’s economic inability to meet very basic needs, the second measure expresses a certain distance of individual income from the community norm. Usually, absolute poverty lines consist of a food component which bases on a minimum caloric requirement of between 2100 and 2400 calories per person and day and a nonfood component (Chen & Ravallion, 1996; Khan & Riskin, 2001). In essence, an absolute poverty measure neglects any interpersonal comparisons of utilities which has been named ‘welfarism’ in the literature (Sen, 2008). Nevertheless, as argued by Khan & Riskin 7 (2001), any absolute poverty threshold carries some society’s judgment what constitutes a minimum living standard, thus, could never be value free. However, most Western countries use a concept of relative poverty due to the view that a measure of poverty should express whether a member of the society is excluded from the standards of living broadly available to others in the same society. Any individual should be able to participate fully in the social life of a community. This is nicely illustrated by a quotation from Adam Smith: “A linen shirt, for example, is, strictly speaking, not a necessary of life. The Greeks and Romans lived, I suppose, very comfortably though they had no linen. But in the present times, through the greater part of Europe, a creditable day-labourer would be ashamed to appear in public without a linen shirt.” (cited after Sen, 1984; p. 79). Finally, in most countries a normative consensus is established that members of a community should benefit fairly equal from a general increase in prosperity. The disadvantage of a relative poverty line is its development over time which is highly correlated with welfare measure’s development within the sample. A widening of the income distribution, for instance, will lead to an increase of the poverty headcount rate. A relative poverty line, thus, provokes criticism of measuring inequality instead of poverty (Townsend, 1985; Sen, 1985). Whereas an absolute poverty line gives poverty estimates which are independent of the welfare measure’s distribution, it needs a continuous update to express the real costs to reach a minimum living standard. Additionally, the calculation of such an absolute line introduces possible sources of measurement errors, like inter-provincial differences in food basket composition if a concept of minimum food intake is used or inter-provincial differences in price developments. More specifically, the Chinese national absolute poverty line has been criticized as too low and subject to several sources of bias among them the minimum level of caloric requirement, food bundle’s composition, the use of planned prices, and the valuation of non-food expenditures (Park & Wang, 2001). Furthermore, the heterogeneous economic development of Chinese provinces might create the need for a regionally differentiated 8 poverty measure. Callan & Nolan (1991) conclude in their review of various ways to determine a poverty line that not one single measure of poverty is likely to dominate. Table 1 presents examples of different poverty lines applied in studies on China. For a more detailed comparison of the development of poverty within China related to differently defined poverty lines we refer to Park & Wang (2001). Table 1 around here Finally, the choice of the appropriate welfare indicator needs some attention. Previous studies use mainly income or consumption expenditure based poverty lines. One main difference between both is the expected smoothing of consumption from year to year, and therefore, a lower variability than income measures. Alternatively, Carter & May (2001) as well as Barrett & Carter (2005) suggest an asset based poverty line. However, its computation is far from being generalized. Gustafsson, Li, & Sato (2004) apply a subjective poverty line to estimate the extent of poverty among urban residents in China. The subjective poverty line is derived from the assessment of an adequate income for a specific type of household to make means end through specific survey questions. Reflecting the advantages and disadvantages of the various definitions of poverty thresholds, the present study compares three different poverty lines: a province-specific relative poverty line and two absolute poverty lines, the national and an international one. To be more specific, the relative rural poverty line is defined at the half-median of the adjusted net income per capita per province in our data. We assume that individuals evaluate their well-being relative to others in same village or province. Income growth in one province should have no impact on the poverty incidence in another province. The national rural poverty line is nominally constant across provinces and the international poverty line bases on US$ 1 a day converted at Purchasing Power Parities. 2 The national rural poverty line is 15 than according to the other two poverty lines. The direction of the estimated coefficients, however, is the same in explaining relative as well as absolute poverty except for one variable. Out of the household characteristics, household size and share of non-working household members seem to increase the probability to live more years in poverty. A higher share of educated household members at each of the three educational levels and the existence of any household member with a non-rural household registration (Hukou) each reduces the probability to be long-term poor. Our results are in line with findings by Gustafsson & Ding (2009) who find a higher probability to be permanently poor linked to larger households. Similarly to our results, they find household head’s education related to a lower probability to be poor. The quantitatively different estimates for age of household head and non-working household members using the relative poverty line compared to the two absolute lines points to lagging opportunities of these households to participate in the general income growth. Conversely, any household member holding a non-rural registration (Hukou) enables the respective household to easier raise household income above absolute thresholds, however, seems not to affect its relative position in the income distribution. Somewhat unexpected, households operating on relatively larger farms show a higher probability to be persistently poor. More in line with expectation, households with a lower diversification of income sources (HHI) experience a higher probability to be persistently poor in absolute terms. 10 The quantitative magnitude of both variables suggests the highest impact on poverty defined by the national poverty line, that is, for households with very low incomes. Finally, the estimated coefficients representing village characteristics reveal a high influence of spatial characteristics on an individual’s probability to be poor. Whereas villages with a higher population density and located close to cities (Suburb) are unanimously associated with a lower probability to be long-term poor, the local unutilized labor capacity has an increasing effect on this probability. Similarly, villages with a higher share of 16 migrating inhabitants are home of more long-term poor households. The different behavior of relative and absolute poverty lines is illustrated by the variable village average income. Whereas the probability to lag behind the relative poverty line is higher for households in richer villages, the probability to remain below the absolute poverty line is significantly lower. This somewhat counterintuitive result might be explained by the fact that relatively richer villages are signified by a broader income distribution and tend to be located in the richest province Zhejiang. 11 Table 7 around here 6. POVERTY EXITS The ordered probit analysis does not take into account, first, that the time a household spent in poverty could be longer than the observed duration. That is, for poor households in the first or the last year of the sample we might underestimate the true length of poverty spells. Second, the probability to leave poverty might be influenced by the time spent in poverty. In order to properly control for these two aspects, a hazard model has to be used. Bane & Ellwood (1986) and Stevens (1994) are among the first to apply this method to poverty analysis. The advantage of the hazard model is that it controls for the influence of the time a household spent in poverty on his instantaneous probability to raise household income above the poverty line conditioning on covariates. Additionally, it does not impose any linearity restriction between the states poor and non-poor on the one side and the covariates on the other side. Define Tas the variable measuring duration of spells in poverty and the matrix of covariates X. As poverty status is derived from annual data and exact time of exit is unknown, we are forced to analyze the probability of exit within the jth year: Pr(tj-1<T≤tj). 12 17 The interval hazard rate, h(j), or discrete hazard rate, expresses the probability of exiting poverty within a given year, conditional upon poverty to year tj:h(j) = Pr(tj-1<T≤tj|T > tj1,X). After controlling for the impact of covariates, the hazard function provides a convenient definition of duration dependence. Negative duration dependence and a decreasing hazard to leave poverty exists at tif the probability to exit poverty decreases with the number of years a household remains poor. The hazard function can be represented as a product of the baseline hazard h0(j) and the explanatory variables. By depending on duration tonly, the baseline hazard neglects any heterogeneity among households. However, behavioral heterogeneity among individuals might change the individual hazard. Part of such variation can be accounted for by controlling for household’s observed individual characteristics. Accordingly, we specify a complementary log-log hazard rate (Jenkins, 1995): h(j,X) = 1 –exp[-exp(’X+j)], where , the baseline hazard, is modeled as a piecewise-constant function by using dummy variables for each year in poverty. The matrix of covariates Xcontains the household, farm and village characteristics and βis a vector of parameters to be estimated. In order to reflect the heterogeneity across the three provinces, additionally, each econometric model is estimated for each province separately. 13 All time-varying covariates are observed independently of the poverty status. Therefore, standard asymptotic estimation techniques provide viable means of estimates of the relative risk parameters β(Kalbfleisch & Prentice, 2002, p. 196). We apply the grouped data approach suggested by Prentice & Gloeckler (1978). 14 Before specifying the semi-parametric hazard model, we compare the shape of the unconditional hazard and survival functions across provinces and poverty lines. Figure 2 presents Kaplan-Meier survival functions. The length of the poverty spell is shown on the horizontal axis and the vertical axis represents the probability to leave poverty. Almost all survival functions suggest a decreasing probability to leave poverty after one additional year 18 in poverty. Furthermore, the estimates across all three poverty lines indicate the highest probability to remain poor for households in Yunnan. The following analysis will provide evidence of the baseline hazard function’s shape after controlling for covariates. Figure 2 around here Turning to the results of the semi-parametric hazard model, in each of the following models the null hypothesis that all coefficients are zero is clearly rejected by likelihood ratio tests. Except for the national poverty line, likelihood ratio tests justify the explicit modeling of duration dependence. The results of the semi-parametric hazard models are presented in two tables: Table 8 reports the estimated coefficients of the baseline hazard and Table 9 displays the coefficients of the covariates. Larger estimates of the duration-specific dummies are associated with a higher hazard of exiting poverty. Whereas an estimate of zero corresponds to a hazard rate of 0.63, an estimated coefficient of one corresponds to a hazard rate of 0.93. Therefore, the estimates in Table 8 suggest hazard rates close to one for Yunnan and lower hazard rates for the other two provinces. However, as most of the coefficients are not statistically significantly different from zero, the estimates suggest a piecewise-constant effect of duration in poverty on the probability to leave poverty. Especially, for the subsample from Yunnan, the poorest province, the probability to leave poverty does not change significantly with time spent in poverty. Such a result suggests that the unconditional estimates in Figure 2 lead to a misleading conclusion if neglecting the impact of covariates. Furthermore, in the case of Hubei, the estimated coefficients point to an increasing probability to pass the national poverty line. Estimates for Zhejiang suggest a constant impact of the spell length on the probability to leave poverty. Summing up, the results point to a different nature of duration dependence across provinces. 19 Tables 8 & 9 around here Besides the time spent in poverty, various household, household business and village characteristics influence the probability to leave poverty. The estimated coefficients of the semi-parametric hazard models besides the baseline function are presented in Table 9. Similar to determinants of long-term poverty, household covariates facilitating significantly the move out of poverty are related to household’s educational attainment and position in village administration. Whereas primary education shows quantitatively the highest influence to leave relative poverty, secondary education has the highest impact on the hazard to pass the international poverty line. With respect to the national poverty line, both levels of education show quantitatively a similar influence. A look on the descriptive statistics of the population at risk might explain this difference. The share of household members with elementary education is very similar in the subgroups below the relative and the international poverty line. But the share of household members with secondary education in families below the national poverty line is more than half that of families experiencing relative poverty spells. Similarly, the probability to leave poverty is between 1.9 and 2.8 times higher for households with a member working in the village administration. The fact that the predicted effect is higher with respect to the relative poverty line is interpreted as showing the existence of networks which help in running emerging economic activities successfully. Similar results have been reported by Knight & Yueh (2008) for the role of social capital in urban labor markets and Knight & Song (2003) regarding the access to employment in local non-farm enterprises. Conversely, the probability to leave poverty is lower for larger households and a higher number of (officially) non-working family members. 15 Moving on to the household business characteristics reveals a lower probability to leave poverty for households with more land per working family member. Comparing the size of the coefficients points to a stronger impact for households below the national poverty 20 line than for relatively poor households. However, the reliance on cropping activities as the single household agricultural activity is predicted to result in a 26 percent lower probability to pass the international poverty line. More productive assets increase the hazard to climb out of poverty. Comparing the estimates across the three poverty lines points to a quantitatively largest impact of productive assets on the probability to raise income above the national poverty line. Looking at the relative poverty line, the quantitatively lower estimate suggests that household business activities do not belong to the fastest growing income sources. Among the local characteristics, population density and village’s location in a mountainous area are main drivers to end spells of poverty. With respect to the relative and the national poverty line households in mountainous villages experience a two time higher probability to leave poverty. Looking at the international poverty line the hazard is much lower and yields only a 1.4 higher probability. Households in villages with a higher share of migrants show a lower probability to leave poverty. This result does not allow any conclusion concerning a direct negative relationship between migration and poverty. Rather we argue that migration originates from the poorer and economically disadvantaged villages. Furthermore, our current specification of the model excludes any potential future flows of money from migrants back to their village. Surprisingly and in line with results of the ordered probit analysis, the average income of the village acts as inhibitor to leave relative poverty. Comparing the hazard models across provinces yields only few differences. 16 Most strikingly, the household size increases the probability to leave relative poverty in Zhejiang but reduces the probability to leave relative poverty in Yunnan. In both regions, larger families seem to face different opportunities to participate in economic activities. At the same time, the downward shift of the baseline hazard to leave relative poverty for households with a higher share of non-working members for the sample from Zhejiang is more than five times as large as for Hubei and almost ten times as large as for Yunnan. Finally, the negative 21 impact of land entitlements per capita on the probability to leave poverty is predicted to be the absolutely highest for Yunnan and with respect to the relative poverty line. A final qualifying remark seems appropriate. The share of censored observations, that is, poor households where either entry or exit into poverty is unobserved, ranges from 30 to 80 percent. Especially, for the sub-sample from Yunnan this group of households forms the majority of the sample at risk (52-80 percent). Based on this observation an extended sample is required to obtain more sound conclusions of duration dependence for the long-term poor. 7. CONCLUSIONS This paper provides new evidence on the poverty mobility of Chinese rural households. Using a 10-year panel over 2100 households from three provinces, the determinants of exiting poverty are investigated taking into account the duration of poverty. A hazard-rate multivariate-modeling framework is applied to three different poverty lines, one relative and two absolute ones. Obviously, rural households have been able to profit from general provincial economic development as shown in the example of Zhejiang. Here the goal should be to retain a broad participation in prosperous activities. Richer provinces such as Zhejiang might rely more on risk management institutions enabling households to continue to participate in economic growth. Policy measures should focus at institutions to manage price and income variation like market information or insurance mechanisms. Instead, absolute poverty is more prevalent and, independent of the poverty line, more persistent in the poorest province Yunnan. Here, policy measures should target at a general improvement of earning opportunities for the poorest households. Looking at the impact of duration dependence reveals significant differences across provinces. Whereas duration dependence does not matter for Zhejiang, it does so for Yunnan. However the chance of escaping poverty remains constant as the poverty 22 spell becomes longer. Therefore, our data do not provide evidence of a poverty trap where the longer a household is poor the lower its chance to escape poverty. Regarding the household related covariates, household size and the share of nonworking family members seem to increase poverty persistence. On the other hand the education of working family members at different levels and any household member having a position as village cadre increases the probability to leave poverty. Land endowment and reliance on cropping as single household business seem to increase poverty persistence. Particularly, the result with respect to land endowment allows two different conclusions. On the one hand, agricultural activities seem to yield only modest returns. On the other hand, non-functioning factor markets, e.g. for hiring labor, limit the full exploitation of the production potential. Which of the two alternatives is more appropriate, needs further analysis. Regarding the covariates controlling for geographical conditions, persistence of poverty is lower in more densely populated and, surprisingly, mountainous areas. Average income per capita at village level has an ambiguous role. It reduces the probability to climb above the relative poverty line. Although a relative poverty line captures more aspects of income inequality than pure needs of people, the determinants of persistent poverty point very frequently into the same direction as determinants of persistent absolute poverty. Despite the general conclusion that poverty is a transitory phenomenon, poor Chinese provinces show a higher relevance of persistent poverty across all poverty thresholds. Thus, different policy measures are needed in order to well address these issues. Province specific objectives and measures in combating poverty seem appropriate. In less wealthy provinces such as Yunnan very poor households should be targeted first and measures should aim at reducing absolute poverty. Here giving people access to prosperous economic activities might be the most relevant measure. Education of rural people seems to be the most promising starting point to enable households to overcome the duration dependence once they 23 experience poverty spells. Furthermore, helping farmers to set up additional household business activities besides cropping might reduce poverty persistence in rural China. Endnotes 1 For a survey of the literature on chronic poverty see McKay & Lawson (2003). 2 The exact definition of the national poverty line is described in detail in Yao (2005) and NBS (2004) 3 Due to lack of generally accepted equivalence scales for Chinese rural households, all household members are weighted equally. 4 Per capita Gross Regional Product in 2004 amounts to 23,942 RMB, 10,500 RMB and 6,733 RMB, respectively (NBS, 2006). 5 Extensive tests following the method suggested by Fitzgerald, Gottschalk, & Moffitt (1998) point to no influence of panel attrition on our results. Detailed results are available from the authors. 6 Unfortunately, the data set does not contain any information on health or nutrition status to account for nonmonetary dimensions of poverty as suggested by Baulch & Masset (2003). 7 Household characteristics bases mainly on the life cycle hypotheses and human capital theory. Village characteristics and village cadre status belong to the structural perspective (Callens, et al., 2004; McKernan & Ratcliffe, 2002; Iceland, 1997). 8 Unfortunately we cannot trace rural registration status to individuals in all years. Data of the last two years (2003 and 2004) show that household members with an non-rural registration enjoyed a significantly longer period of education and don’t work in agriculture (i.e. industry, commercial/other services). 9 The effect of an estimated coefficient on the middle categories of the dependent variable of an ordered probit model is ambiguous (Greene, 2000). 10 The Herfindahl-Hirschman-Index (HHI) expresses the degree of concentration of income from the different household business’ activities, wage work and capital income. It does not penalize any of these activities. 11 The correlation coefficient between the variance of household incomes at village level and the village average income per capita is positive (0.66) and statistically significant. 12 For a more detailed description of the method the reader is referred to Kalbfleisch & Prentice (2002). 13 Formal tests of the equality of the survivor function across provinces support this hypothesis. Results are available upon request. 24 14 Iceland (1997) shows that dropping left-censored observations might cause serious selection bias. Therefore, both types of censoring, unobserved start of poverty spell as well unobserved end of poverty spell, are treated equally in this analysis. 15 Underlying data do not allow separating this effect into children and retired family members. 16 Detailed results are available upon request. 31 Table 3: Descriptive statistics of explanatory variables All provinces Zhejiang Hubei Yunnan Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Household characteristics Household size Number of permanent residents in household 4.1 1.4 3.6 1.2 4.1 1.5 4.4 1.5 Age of head Age of household head (in categories) 3.0 1.1 3.1 1.0 3.0 1.1 2.9 1.1 Dependents Share of non-working household members 0.3 0.2 0.3 0.2 0.3 0.2 0.3 0.2 Elementary education Share of household labor force with completed elementary education 0.4 0.3 0.4 0.3 0.4 0.3 0.4 0.4 Secondary education Share of household labor force with completed secondary education 0.3 0.3 0.4 0.3 0.4 0.3 0.2 0.3 Higher education Share of household labor force with completed higher education 0.1 0.2 0.1 0.2 0.1 0.2 0.0 0.1 Hukou Dummy for any household member with a non-rural registration 0.1 0.3 0.1 0.3 0.1 0.3 0.1 0.3 Cadre Dummy for any household member working as village cadre 0.1 0.2 0.1 0.3 0.1 0.3 0.0 0.2 Farm characteristics Land size Land size per capita (in mu) 1.3 1.4 0.6 0.6 1.3 1.4 1.6 1.7 Asset value Productive assets per capita (in 1000 yuan) 2.0 12.9 5.1 16.3 0.6 1.8 2.1 18.9 Cropping Cropping as single agricultural activity 0.1 0.3 0.0 0.2 0.1 0.3 0.0 0.2 HHI Hirschman-Herfindal-Index of household income sources 0.5 0.2 0.5 0.2 0.5 0.2 0.4 0.2 Village characteristics Unutilized labor force Share of person-working days less than 300/ year 0.1 0.2 0.2 0.1 0.1 0.2 0.1 0.1 Migration Share of out-migrating village residents 0.01 0.01 0.01 0.01 0.01 0.02 0.01 0.01 Population density Natural log of permanent residents over village area -1.3 1.3 -0.9 1.1 -1.0 1.0 -2.1 1.4 Suburb Village in suburban area 0.2 0.4 0.1 0.3 0.2 0.4 0.2 0.4 Plain Village on plain 0.3 0.4 0.5 0.5 0.3 0.4 0.2 0.4 Mountains Village in mountainous area 0.4 0.5 0.2 0.4 0.3 0.5 0.6 0.5 Average village income Annual net income per capita (in 100 yuan) 24.6 20.9 50.6 27.9 17.3 5.0 19.5 18.1 32 Source: Own computation based on RCRE data. Table 4: Rural poverty head-count, 1995-2004 Year Head-count ratio (relative poverty line) Head-count ratio (national poverty line) Head-count ratio (internat. poverty line) Average three provinces Zhejiang Hubei Yunnan Average three provinces Zhejiang Hubei Yunnan Average three provinces Zhejiang Hubei Yunnan [%] [%] [%] [%] [%] [%] [%] [%] [%] [%] [%] [%] 1995 9.17 11.78 7.13 10.41 1.19 0.21 0.33 3.67 19.20 2.57 14.49 43.67 1996 11.29 16.09 7.72 13.21 1.94 1.07 1.01 4.47 20.30 4.51 18.12 39.23 1997 11.35 18.16 6.95 12.83 1.84 1.50 1.01 3.67 18.42 4.27 17.71 33.20 1998 11.91 15.17 9.70 12.83 2.69 0.43 2.34 5.50 18.86 1.92 17.39 37.68 1999 11.81 11.56 9.38 16.46 3.77 0.86 1.67 10.37 19.95 3.21 17.41 40.45 2000 10.58 16.74 5.03 14.84 3.56 1.07 1.34 9.96 18.63 2.57 14.88 40.65 2001 11.66 10.90 7.92 19.11 3.87 0.95 1.56 10.57 16.91 1.66 15.40 32.72 2002 11.97 12.82 8.59 17.31 3.56 0.21 2.34 8.96 16.01 1.92 14.17 32.79 2003 11.85 13.25 7.47 18.53 2.75 0.85 1.00 7.74 13.58 1.71 8.36 34.42 2004 11.16 12.20 7.02 17.92 1.54 0.00 0.33 4.89 9.96 0.27 4.34 27.49 Source: Own computation based on RCRE data. 33 Table 5: Markov transition probability matrices Transition from … to … Relative poverty line National poverty line International poverty line Nonpoor Poor Nonpoor Poor Nonpoor Poor All three provinces Non-poor 0.93 (0.002) 0.07 (0.002) 0.98 (0.001) 0.02 (0.001) 0.92 (0.002) 0.08 (0.002) Poor 0.54 (0.012) 0.46 (0.012) 0.63 (0.023) 0.37 (0.023) 0.40 (0.009) 0.60 (0.009) Mean length of poverty 1.86 1.56 2.48 Prais mobility index 0.61 0.66 0.48 Zhejiang Non-poor 0.90 (0.005) 0.10 (0.005) 0.99 (0.001) 0.01 (0.001) 0.98 (0.002) 0.02 (0.002) Poor 0.60 (0.021) 0.40 (0.021) 0.77 (0.075) 0.23 (0.075) 0.75 (0.042) 0.25 (0.042) Mean length of poverty 1.68 1.29 1.33 Prais mobility index 0.70 0.78 0.77 Hubei Non-poor 0.95 (0.003) 0.05 (0.003) 0.99 (0.001) 0.01 (0.001) 0.91 (0.003) 0.09 (0.003) Poor 0.66 (0.019) 0.34 (0.019) 0.83 (0.036) 0.17 (0.036) 0.56 (0.014) 0.44 (0.014) Mean length of poverty 1.52 1.20 1.77 Prais mobility index 0.71 0.84 0.65 Yunnan Non-poor 0.92 (0.004) 0.08 (0.004) 0.96 (0.003) 0.04 (0.003) 0.87 (0.006) 0.13 (0.006) Poor 0.38 (0.019) 0.62 (0.019) 0.54 (0.028) 0.46 (0.028) 0.26 (0.011) 0.74 (0.011) Mean length of poverty 2.61 1.84 3.79 Prais mobility index 0.46 0.59 0.39 Note: Standard errors in parentheses; The probability to remain poor from one year to the next is denoted with pPP. Mean length of stay in poverty is 1/(1-pPP). Prais mobility index of the matrix P is calculated as M(P) = 2tr(P) (Geweke, Marshall, & Zarkin, 1986). Source: Own computations based on RCRE data. 34 Table 6: Number of households in cohorts according to time spent in poverty Duration Relative poverty line All provinces Zhejiang Hubei Yunnan Never poor 807 115 438 254 1–4 years poor 452 125 214 113 ≥ 5 years poor 103 29 19 55 National poverty line All provinces Zhejiang Hubei Yunnan Never poor 1184 253 611 320 1–4 years poor 161 16 60 85 ≥ 5 years poor 17 0 0 17 International poverty line All provinces Zhejiang Hubei Yunnan Never poor 704 229 342 133 1–4 years poor 429 40 264 125 ≥ 5 years poor 229 0 65 164 Source: Own computation based on RCRE data. 35 Table 7: Determinants of long-term poverty (N = 1362) Relative poverty line National poverty line International poverty line Coef. Standard error Coef. Standard error Coef. Standard error Household characteristics Household size 0.06** 0.027 0.09*** 0.033 0.12*** 0.027 Age of head 0.09** 0.036 -0.03 0.052 0.03 0.037 Dependents 0.43** 0.181 0.29 0.246 0.22 0.174 Elementary education -0.69*** 0.131 -1.06*** 0.168 -0.84*** 0.134 Secondary education -0.54*** 0.146 -1.01*** 0.188 -1.13*** 0.147 Higher education -1.04*** 0.255 -0.97** 0.341 -1.48*** 0.263 Hukou -0.27 0.165 -0.55* 0.286 -0.55*** 0.167 Cadre -0.25 0.157 -0.21 0.221 -0.25 0.158 Farm characteristics Land size 0.12*** 0.035 0.17*** 0.034 -0.002 0.034 Asset value -0.01 0.012 -0.004 0.019 0.01 0.012 Cropping 0.24 0.209 -0.006 0.304 0.10 0.212 HHI 0.32 0.237 0.95** 0.375 0.48** 0.238 Village characteristics Unutilized labor force 3.56*** 0.421 1.75*** 0.624 0.92** 0.402 Migration 7.42*** 2.305 7.80** 3.269 5.02** 2.312 Population density -0.35*** 0.067 -0.27*** 0.099 -0.36*** 0.063 Suburb -0.76*** 0.187 -0.15 0.267 -0.44** 0.175 Plain 0.22* 0.120 0.57*** 0.168 0.40*** 0.108 Mountains -0.57*** 0.137 -0.32 0.200 -0.02 0.137 Average village income 0.02*** 0.003 -0.01 0.006 -0.03*** 0.005 1 1.81 0.268 2.13 0.350 0.27 0.270 2 3.36 0.279 3.83 0.367 1.64 0.273 Pseudo R2/ AIC 0.17 2001.2 0.24 930.6 0.26 2069.2 Note: ***, **, * denote 1%, 5%, and 10% significance level, respectively. Estimation bases on balanced panel. Source: Own computations based on RCRE data. 36 Table 8: Conditional estimates of duration dependence from hazard model Duration of poverty spell Relative poverty - Exit National poverty line - Exit International poverty line - Exit Pooled sample Zhejiang Hubei Yunnan Pooled sample Zhejiang Hubei Yunnan Pooled sample Zhejiang Hubei Yunnan 1 -0.07 0.85 -0.53 3.24 0.46 -4.69 4.13 0.72 0.93 0.08 2.73 2 -0.17 0.72 -0.36 3.07 0.16 -3.92a 3.76 0.83 1.09 0.05 3.03 3 -0.23 0.79 -0.75 3.49 0.64 4.35 0.70 1.10 a 0.01 2.80 4 -0.71 0.80a -1.06a 2.08a 0.29a 4.18 0.46 -0.22 2.62 5 -1.30a 0.06 0.02a 2.09 6 0.52 2.79 7 0.67 2.74a 8 -0.01a Note: Not enough observations for Zhejiang and national poverty line to estimate full model. Estimates in bold statistically significantly different from zero at least at 10 % significance level. a - Dummy controls for duration up to this spell length and higher. Source: Own computations based on RCRE data. 37 Table 9: Covariates of hazard model –exit from poverty Variable Relative poverty line National poverty line Internat. poverty line Coef. Standard error Coef. Standard error Coef. Standard error Household characteristics Household size -0.03 0.025 -0.13*** 0.043 -0.12*** 0.024 Age of head -0.05 0.034 -0.002 0.076 -0.01 0.031 Dependents -0.19 0.165 0.41 0.333 -0.41*** 0.160 Elementary education 0.39*** 0.127 0.62** 0.240 0.39*** 0.119 Secondary education 0.24* 0.142 0.61* 0.338 0.78*** 0.137 Higher education -0.03 0.204 -0.22 0.485 0.46* 0.253 Hukou 0.16 0.139 -0.86** 0.362 0.47*** 0.147 Cadre 1.02*** 0.319 a 0.66*** 0.242 Farm characteristics Land size -0.06** 0.031 -0.13*** 0.045 -0.05 0.032 Asset value 0.04*** 0.014 0.43** 0.196 0.14*** 0.041 Cropping -0.19 0.144 -0.41 0.324 -0.30** 0.151 HHI 0.48** 0.216 0.61 0.498 -0.30 0.250 Village characteristics Unutilized labor force -0.57 0.387 -0.62 1.010 0.72* 0.389 Migration -4.66 3.115 -5.89 10.123 -18.10*** 3.040 Population density 0.22*** 0.068 0.21 0.176 0.50*** 0.068 Suburb -0.004 0.151 0.07 0.454 -0.07 0.158 Plain 0.20* 0.112 -0.15 0.328 -0.19 0.126 Mountains 0.48*** 0.148 0.72* 0.409 0.22 0.138 Average village income -0.01*** 0.002 -0.01 0.010 -0.001 0.003 Model diagnostics Statistic p-value Statistic p-value Statistic p-value H0: all parameters except const = 0 406.65 <0.001 83.62 <0.001 965.80 <0.001 H0: no duration dependence 39.03 <0.001 3.06 0.55 31.05 <0.001 AIC 2354.8 501.6 3262.4 No. of spells/ No. of exits 1834/ 785 387/ 227 2922/ 976 Note: ***, **, * denote 1%, 5%, and 10% significance level, respectively. a –dropped due to perfect relation with dependent variable. Source: Own computations based on RCRE data. 38 0 .2 .4 .6 .8 Density 4 6 8 10 12 log of net household income per capita Zhejiang Hubei Yunnan Figure 1: Kernel Density Estimation of log(net income per capita) in 2004 Source: Own computation based on RCRE data. .00 .25 .50 .75 1.0 0 2 4 6 8 10 years spent in poverty Survival in relative poverty 1.0.75.5.250 0 2 4 6 8 10 years spent in poverty Survival in absolute poverty - national line .75 1.0.50.250 0 2 4 6 8 10 years spent in poverty Survival in absolute poverty - internat. line Zhejiang Hubei Yunnan Figure 2: Survival functions of poverty spells Source: Own computation based on RCRE data.