Are Female-Headed Households More Food Insecure? Evidence from Bangladesh

Transcription

1 World Development Vol. 38, No. 4, pp , 2010 Ó 2009 Elsevier Ltd. All rights reserved X/$ - see front matter doi: /j.worlddev Are Female-Headed Households More Food Insecure? Evidence from Bangladesh DEBDULAL MALLICK Deakin University, Victoria, Australia and MOHAMMAD RAFI * Research and Evaluation Division, BRAC, Dhaka, Bangladesh Summary. This paper uses household and village-level survey data to investigate the food security of male- and female-headed households in Bangladesh with particular attention to indigenous ethnic groups, and finds no significant differences in the food security between these two types of households. The absence of social and cultural restrictions among the indigenous groups permitting their females greater freedom to participate in the labor force coupled with informal redistributive mechanism is attributed to their less food insecurity. This result indicates that noneconomic institutions can significantly impact economic outcomes such as food security. Ó 2009 Elsevier Ltd. All rights reserved. Key words food security, female-headed households, generalized threshold model, indigenous ethnic group, South Asia, Bangladesh 1. INTRODUCTION It has generally been observed that female-headed households are more food insecure than male-headed households. The former are said to be more vulnerable to food insecurity due to the triple burden: (i) the female head, who is the main income earner, faces various disadvantages in the labor market and many productive activities, (ii) she is also responsible for maintaining the household including household chores and child care in addition to working outside, and thus she is activity burdened, and (iii) she faces a higher dependency ratio for being the single income earner (Fuwa, 2000). In this paper, we investigate the food security of male- and femaleheaded households using household and village-level survey data for the Bengali and four indigenous ethnic groups living in the Chittagong Hill Tracts (CHTs) in Bangladesh. 1 This paper departs from the existing literature in two respects. First, we use the perception of the respondents about their own food security status rather than consumption (converted into calories) or expenditure data. Both subjective and consumption-based measures of food security have potential advantages and disadvantages, but our choice is driven by the peculiarities of the CHT and the nature of the study. Given the large seasonal volatility of consumption in Bangladesh, especially among the indigenous people (Bleie, 2005), and since data were collected in a single round of survey, consumption data must systematically under- or over-report the true food security. Based on food production, availability, purchasing power, and access to common resources, the respondents defined the food security status of their households in one of the following four categories severe (chronic) food shortage, occasional (transitory) food shortage, break-even, and food surplus. Second, we pay particular attention to the food security of the indigenous ethnic groups. It has been observed that the indigenous ethnic groups have historically been concentrated in certain geographical regions which are usually characterized by low productivity and high incidence of poverty 593 (Fuwa, 2000). These areas often lack adequate public services such as health facilities, schools, and proper sanitation. Indigenous people also face discrimination in the labor market, such as lower earnings after controlling for education and other work characteristics (World Bank, 1999). The CHT is no exception, justifying an independent investigation of the food security of the indigenous groups. Given that the responses on food security are ordered, we estimate the generalized threshold model introduced by Maddala (1983) and Terza (1985) that overcomes several limitations of the ordered probit model. Since the coefficients estimated have no meaningful interpretation without some additional calculation, we calculate the elasticities. Our main finding is that there is no significant difference in food security between male- and female-headed households among the indigenous ethnic groups, in contrast to the conventional view about the vulnerability of the female-headed households. However, this result does not hold if Bengalis are included in the sample and is robust to alternative specifications. We explore several possible explanations from the development literature and suggest socio-cultural and religious practices as the reasons for this result. The absence of social and cultural restrictions among the indigenous groups in the CHT permits females greater freedom to participate in the labor force. The indigenous ethnic groups practice Buddhism, Hinduism, and animism, while the Bengalis are predominantly * The authors would like to acknowledge the helpful comments and assistance offered by Prasad Bhattacharya, Matthew Clarke, Shyamal Chowdhury, Rajeev Dehejia, Chris Doucouliagos, Chris Geller, Mike Kidd, Atonu Rabbani, John Strauss, Mehmet Ulubasoglu and three anonymous referees. The Research and Evaluation Division of BRAC, Bangladesh has generously provided the dataset. All errors and omissions are the sole responsibility of the authors, and the conclusions do not necessarily reflect the views of the organizations with which they are associated. Final revision accepted: September 17, 2009.

2 594 WORLD DEVELOPMENT Muslims and are socially and culturally conservative (Khan, 1997). Females from the indigenous groups engage in income-generating activities outside the household at a higher rate. Religious and cultural practices of the indigenous groups also favor the female-headed households in informal redistribution. However, we find that female-headed households are activity burdened as they maintain household chores in addition to working outside the home, thus corroborating one of the tenets of the triple burden. The finding is indicative in the sense that noneconomic institutions can significantly impact economic outcomes such as improving the food security of a household, especially the female-headed one. It has important policy implications as well. The design of developmental assistance programs should take into consideration the social and cultural heterogeneity even within a region in a country. This is more relevant to the policy makers who design and implement anti-poverty policies or microfinance schemes for culturally diverse ethnic groups, not only in Bangladesh but also in similar regions such as North- East India, Nepal, and Sri Lanka. The rest of the paper is organized as follows. Section 2 reviews the literature on the incidence of and reasons for differential food security among male- and female-headed households. Section 3 describes the CHT and different ethnic groups living therein. Section 4 discusses sampling procedures, data collection methods and variables considered. Section 5 reports the descriptive statistics of food security for the ethnic groups. Section 6 explains our estimation model, and Section 7 discusses the results. Finally, Section 8 concludes. 2. FOOD SECURITY AND GENDER OF THE HOUSEHOLD HEAD (a) Concept of food security Food security is a multifaceted concept. No single measure can encompass all of its aspects (Chung, Haddad, Ramakrishna, & Riely, 1997, p. 8). However, briefly it can be defined as the peoples security against risks of not having access to required food. At the household level, food security is defined as sustainable access to sufficient quantity and quality of food to ensure adequate dietary intake and a healthy life for all household members (FAO, 1992). However, even when food supplies are adequate at the aggregate level, a number of factors may prevent poor households or individuals from accessing food. They may lack purchasing power, may not have access to land for own cultivation, may lack the necessary assets or access to credit that helps them smooth consumption, or may find themselves outside any public assistance or other programs (Sen, 1981, 1995). Food (in)security has also a temporal dimension. It is defined as transitory when a population suffers from a temporary decline in food consumption and as chronic when a population is continuously unable to acquire sufficient food (Chung et al., 1997, p. 5). Depending on its asset endowments, a household adopts different strategies to cope with transitory food insecurity that often involves borrowing money and the sale or consumption of productive assets. These practices may undermine the long-term productive potential of poor households and may eventually lead to chronic food insecurity. Households most vulnerable to food insecurity generally include those which lack productive assets and depend on irregular income from daily wage labor. Groups such as day laborers, casual fishermen, and beggars fall into this category (Sen, 1981). Households living in areas that are susceptible to natural disasters, or are inaccessible or unsuitable for agricultural production are also vulnerable to food insecurity. Within households, women, girls, elderly, and disabled members are often the victims of discrimination and therefore, more vulnerable. 2 Access to proper sanitation and health care, general nutritional awareness, and caring practices are important determinants of an individual s capacity to absorb and utilize the nutrients in the diet and ultimately of one s food security status. (b) Food security and female-headed households It has generally been considered that female-headed households are more vulnerable to food insecurity. In Section 1, we mentioned the triple burden that results in greater vulnerability for the female-headed households. However, there is no empirical regularity of the simultaneous presence of all three factors. For example, Barros, Fox, and Mendoncßa (1997, p. 232) observe in the context of urban Brazil that the main reason for female-headed households being poor is not a lower number of earners relative to family size but the lower earning power of these earners. Since males earn more than females in the same job, a household lacking male-earned income simply has a much higher probability of being poor. In addition, socio-cultural factors can prohibit women s participation in the labor force. In some of the poorest areas of South Asia, cultural restrictions on women s ability to participate fully in food production activities have left them particularly vulnerable in times of economic crisis (Kabeer, 1990, p. 1). Female-headed households are more vulnerable to nonincome aspects of poverty as well. Being activity burdened, female-headed households employ additional household members including school-going children in income-generating activities. This is reflected in the low attainment in schooling for children in the female-headed households (Buvinić & Rao Gupta, 1997, p. 268). McLanahan (1985, p. 883) also finds that, after controlling for income, children in the female-headed households have a lower rate of socio-economic attainment than children in the male-headed households. If female-headed households utilize all available resources including existing human capital to survive, then they cannot invest in future human capital formation, thus the likelihood of transmitting poverty to the next generation is higher. There is also a heterogeneity among the female-headed households with different reasons for the female being the household head. For example, Drèze and Srinivasan (1997) find evidence that widow-headed households are more disadvantaged than other types of female-headed households. There are also counter examples that female-headed households are no less food insecure than male-headed households. Quisumbing, Haddad, and Peňa (2001), using household survey data set for 10 developing countries, find no statistically significantly higher incidence of poverty among the femaleheaded households in two-thirds of the countries. Among the exceptions is Bangladesh, where the female-headed households have consistently higher poverty among the bottom third of the population; however, this survey excludes the CHT region. The higher incidence of poverty among the female-headed households is also subject to the poverty indices used for measurement. For example, Drèze and Srinivasan (1997, pp ), in the context of rural India, find no evidence of higher incidence of poverty among the female-headed households in terms of standard poverty indices based on household

3 ARE FEMALE-HEADED HOUSEHOLDS MORE FOOD INSECURE? EVIDENCE FROM BANGLADESH 595 per capita expenditure. This finding is robust to using adult equivalence consumption for reasonable choice of consumption equivalence scales. But the female-headed households are poorer after controlling for economies of scale that accounts for their smaller size. In other words, after controlling for household size and child adult ratio, female-headed households are poorer than the male-headed households. Ravallion and Lanjouw (1995, p. 1413) find that the relationship between household size and consumption (and therefore, poverty) depends on the size elasticity of the cost of living. Female-controlled incomes have a more favorable impact on intra-household allocation than male-controlled incomes (Duflo, 2005). Unearned income in the hands of a mother contributes greater improvement to children s health and nutrition (Thomas, 1990), and larger expenditure shares in household nutrients, health, and housing (Thomas, 1992). Buvinić and Rao Gupta (1997, p. 269), reviewing 65 studies carried out in the last decade, conclude that female-headed households prefer to invest scare resources in children, which translates to increased child welfare relative to income. Duflo and Udry (2004) find that a bigger share of the women s contribution to the household income is spent on food and on private goods for women, while a bigger share of the men s contribution to the household income is spent on alcohol and tobacco and on private goods for men. Table 1. Ethnic groups in the CHT Source: Bangladesh Bureau of Statistics (1992) Ethnic groups Population in the CHT (%) With Bengali Without Bengali (1) (2) (3) Bengali 48.6 Bawm Chak Chakma Khumi Kheyang Lushei Marma Mro Pankhua/Pankko Tanchangya Tripura Others a Total (974,445) (501,144) Figures in parentheses are total populations. a Includes Rakhain, Saotal, Asamis, etc., each with less than 0.02% of total population. 3. THE CHITTAGONG HILL TRACTS AND THE ETHNIC GROUPS The purpose of this section is to describe the overall economic (under)development in the region and the ethnic and socio-cultural differences between the indigenous groups and Bengalis. These differences, as we argue later, are important in explaining food security in the female-headed households. The Chittagong Hill Tracts (CHT), with an area of 13,181 square kilometers representing 9.15% of Bangladesh, is located in the south-eastern tip of the country. Geographically the CHT is a part of Hill Tripura and Arakan Yoma branching off from the Himalayan ranges and continuing to the south through Assam and Hill Tripura of India to Arakan of Myanmar. The topography of the region is featured by hills, ravines, and cliffs, originally covered by dense bamboo, tree, and creeper jungle but is presently bare in many places. Altogether 11 small indigenous groups, Mongoloid by ethnic origin, live in the CHT. The groups are Bawm, Chak, Chakma, Khumi, Kheyang, Lushei, Marma, Mro, Pankhua (or Pankko), Tanchangya, and Tripura. The groups bear a closer resemblance to hill peoples of the vast strip of land extending from Tibet to Indo-China (Khan, 1995, p. 3). The ethnic origin, socio-cultural background, and geographical features of the region make the indigenous peoples different from Bengalis. However, there is also heterogeneity among the indigenous groups in terms of language, cultural practices, religion, and even food habits. Table 1 shows that the population size varies widely among the ethnic groups. At one extreme, Bengalis constitute about 49% of the total population. Among the indigenous ethnic groups, Chakma, Marma, Tripura, and Mro are the most numerous each with more than 20,000 people and together constitute 92.8% of the indigenous people in the region. As part of the policy of the Government of Bangladesh (GOB), Bengalis were systematically settled in the CHT in three phases in By the end of this period 300, ,000 Bengalis had been settled (The Guardian, 1984). Their settlement during the period was so heavy that it made the Bengalis the dominant socio-cultural group in the region. Although the indigenous population in the CHT increased by 2.4% annually on average during , their proportion to the total population decreased from 93% to 51.4% during the same period. 3 One may ask, from where was the land made available for such a large number of Bengalis? A few thousand acres of reserve forest containing both mixed and hilly lands were released for settlement, but these lands hardly amounted to one-tenth of what was required for the settlement (May, 1984; Rafi & Chowdhury, 2001, p. 31). The rest were settled by evicting the indigenous people from their lands by grossly violating their traditional rights and affecting their lifestyles. The settlement policy of the GOB made 100,000 indigenous people homeless. About half of them became refugees in the Indian states of Tripura and Mizoram. The rest scattered themselves within the region living either as dependents on their relatives or moved into the forest (Mohsin & Ahmed, 1996, p. 271). The economy of the CHT is predominantly agricultural. The jhume (or swidden) 4 cultivation is the traditional and predominant mode of cultivation practiced by the indigenous people. A Chakma Raja (king) in the later part of 19th century, after realizing that plough cultivation is more productive than jhuming, decided to introduce it in his territory. Thereafter it gradually penetrated into the agricultural production system in the CHT (Shelley, 1992, p. 65). The cultivation is done at the plains located in the river valleys and the stretches of plains between hills where irrigation is possible. In recent decades, the indigenous groups have been considerably deprived of these lands by the construction of a dam on the Karnofully river to generate hydroelectricity inundating most of these lands, and by eviction from their lands for the settlement of Bengalis. Very recently national NGOs such BRAC and Grameen Bank have extended microfinance programs in the region. International development agencies (such as UNDP and USAID) have also undertaken several development programs both directly and indirectly through the government.

4 596 WORLD DEVELOPMENT 4. DATA This study addresses the ethnic groups having more than 20,000 people. The selected ethnic groups are Bengali, Chakma, Marma, Mro, and Tripura together constituting more than 96% of the total population in the CHT. Altogether, 510 households were targeted for survey from each of the five ethnic groups observed, totaling 2,550 households. A two-step sampling procedure was followed in selecting the households. First, 30 villages were randomly selected from the total villages for each ethnic group under observation. A general tradition in the CHT is that people of the same ethnic group usually live in a particular village. Second, a list of all households in each village was prepared with the help of the karbari/headman 5 and other informed people such as school teachers, and then 17 households were randomly selected from the list. Next, 30 field workers, six from each ethnic group, with equal number of males and females, and equally fluent in Bengali and their native language, were hired. Field workers of a particular ethnic group interviewed the respondents of the same ethnic group because of language barriers. 6 The data collected include demography, land ownership and cultivation, production and exchange, savings, asset accumulation and food security at the household level, and physical infrastructure at the village level such as distance of the village from the Thana (lowest administrative unit) headquarters, nearest bazaar (or local market), all-weather pucca road, and bus stand. We use the perception of the respondents about their own food security status. Based on food production, availability, purchasing power, and access to common resources, the respondents defined the food security status of their households in one of the following four categories severe (chronic) food shortage, occasional (transitory) food shortage, break-even, and food surplus. 7 Information on intra-household distribution of food among the members was not sought because such a sensitive issue might induce systematic errors in the responses. It is important to mention that all adult household members present at home participated in the interview. In many occasions the household head was out for work during the interview. Although the household head or adult male has more information about some issues such as land holding and cultivation which s(he) replied, questions that require some judgment, such as food security, were discussed among household members before answering. Therefore, the perception about the food security status should be considered as responses at the household level rather than at the individual level. Consumption is sometimes used as a proxy for food security. These data were not collected because seasonal fluctuations in consumption are very large in the CHT and the study was designed for a single round of survey due to mountainous and inaccessible terrain and lack of transportation in the region. Bleie (2005, pp ) documents the seasonal fluctuations of food consumption among the indigenous people in Bangladesh. The traditional months of plentiful stocks ranging from Aghrayon to Magh (mid- December to mid-february) also coincided with the marriage and festival season. However, this season has now become the months of moderate food security. In Magh, some can afford three simple but fairly balanced meals daily with the majority having only two full meals. The periods of moderate food shortage are in the winter from Falgun to Biashakh (mid-january to mid-april). On the other hand, the periods of extreme food shortage, sometimes amounting to hunger, between Bhadro and Kartik (mid-august to mid- October) have been increasing. Although our subjective measure of food security overcomes some limitations by capturing the broadness of the concept and seasonal fluctuations in consumption, it may also be subject to other limitations. For example, perceptions of the male- and female-headed households about food security may systematically differ. 8 In such a case, consumption or the difference in the seasonal fluctuations in consumption between the male- and female-headed households would be a better measure. It is worth mentioning that both approaches have advantages and disadvantages and the jury is still out. Income has been found to be one of the important determinants of food security (Iram & Butt, 2004, p. 763). It is well known that income data are difficult to collect. The problem is more acute in this case given the level of underdevelopment in the CHT. Most households in the CHT, especially those among the indigenous ethnic groups, do not have any cash income. They produce for home consumption and in most instances borrow and repay in kind. There are three district headquarters and only a few bazaars in the region; these places are inhabited by Bengalis. Most of the indigenous people live in the hinterland, and there are virtually no mechanical means to transport to the market. Nonetheless, they sometimes travel to the market on foot to sell their agricultural products and to buy what they do not produce. Here, they are deprived of the fair price for their commodities because Bengalis control and manipulate the market against the indigenous people (Anti- Slavery Society, 1984; Mohsin, 1997; Rafi & Chowdhury, 2001). As a result, the market value of agricultural products cannot be imputed because of the lack of reliable price data. Since agriculture is the predominant economic activity in the region, the main source of income, whether cash or kind, is agricultural production. It is therefore, reasonable to assume that the amount of land cultivated or rice produced is a good proxy for income. Finally, it is imperative to clarify the definition of the household head. The household members were asked to determine the headship based on a combination of two criteria gender of the main income earner and of the ultimate decision maker in the family. 5. DESCRIPTIVE STATISTICS In this section, we provide descriptive statistics for food security status, demographic characteristics, cultivation, and agricultural production for the male- and female-headed households. Only 5.81% of all households in the CHT are female-headed (column 2 in Table 2) with little variation among the ethnic groups. About 10% of the female heads are married, compared to 94% of the male heads (columns 3 and 4 in Table 2). Most females are divorced, separated, or widowed, which is consistent with the findings by Fuwa (2000, pp ). Table 3 provides some stylized facts about the food security status for the male- and female-headed households and for different ethnic groups. Panel A classifies households only in two categories food insecure and secure (chronic and transitory food insecurities are combined into food insecurity and break-even and food surplus are combined into food security ). About 72% of all households in the CHT do not have food security. Higher percentages of the female- than the male-headed households are food insecure at 83% and 71%, respectively. This pattern is uniform across the ethnic groups. However, Pearson v 2 statistics show that the difference is significant only for the Bengali and the Mro at the 5% and 10% level, respectively.

5 ARE FEMALE-HEADED HOUSEHOLDS MORE FOOD INSECURE? EVIDENCE FROM BANGLADESH 597 Table 2. Percentage of female-headed households by ethnic groups and their marital status Ethnic group Female-headed households (%) The above picture does not reveal the intensity of food insecurity. Panel B of Table 3 reports both chronic food insecurity and transitory food insecurity. A higher percentage of the male- than the female-headed households suffer from transitory food insecurity while higher percentages of the femalethan the male-headed households suffer from chronic food insecurity. About 28% of the male-headed households suffer from chronic food insecurity, compared to 47% of the female-headed households. On the other hand, about 43% and 36% of the male- and female-headed households suffer from transitory food insecurity, respectively. But the Pearson v 2 statistics show that the difference is significant only for the Bengalis and the Chakma. The above discussions reveal that the Bengali is the only ethnic group in the region that has a significantly higher percentage of food insecure female-headed households. Table 4 reports descriptive statistics for other variables used in the multivariate analysis. 6. ANALYTICAL FRAMEWORK Percentage of household heads married Female Male (1) (2) (3) (4) Bengali 7.39 (37) Chakma 3.94 (20) Marma 9.94 (50) Mro 3.14 (16) Tripura 4.72 (24) All 5.81 (147) Figures in parentheses are number of households. Given the ordered nature of the responses on food security, which is the dependent variable, the natural choice is the ordered probit (or logit) model. The OLS regression is not appropriate for two reasons. First, it assumes, for example, that the difference between chronic food insecurity and transitory food insecurity is the same as that between break-even and food surplus. Second, it implicitly assumes that two respondents who give the same response have exactly the same attitude. The ordered probit model overcomes this problem because it estimates the parameters of the underlying distribution, rather than the response itself (Daykin & Moffatt, 2002, p. 159). However, the coefficient estimated by the ordered probit model does not have any direct interpretation; the marginal effect, i.e. how a change in one of the explanatory variables changes the predicted distribution of the dependent variable, is more convenient for interpretation. Two restrictive properties of the marginal effects in the ordered probit model limit its usefulness. First, relative marginal probabilities are constant across individuals, that is, they do not depend on individual i or h. Second, the marginal effects change sign exactly once when moving from the smallest to the largest outcome (single-crossing property). 9 To overcome the limitations, we also estimate the generalized threshold model introduced by Maddala (1983) and Terza (1985) that relaxes the single threshold assumption and allows for different indices across outcomes. The explanatory variables are the following: GEND represents gender of the household head (1 = male and 0 = female). AGE represents age of the household head (years). AGE2 represents square of AGE. EDU represents education of the household head (years of schooling). OCCU represents occupation of the household head (1 = nonfarm and 0 = farm). STAT represents status of the household in the village (1 = karbari/headman and 0 = general household) indicating its involvement in the indigenous power structure. DDR represents demographic dependency ratio. 10 PLND represents amount of plain land cultivated last year (acres). JHUM represents amount of jhume land cultivated last year (acres). VIBR represents economic infrastructure. There are reasons for excluding variables such as the value of savings and assets. Only households with higher food security are able to save and accumulate assets. Therefore, inclusion of these variables would induce endogeneity, for which we do not have instruments to correct the bias and inconsistencies of the estimated coefficients. The amount of rice produced is also excluded for the same reason as it is endogenous in a standard (static) farm household model (Bardhan & Udry, 2000, chapter 2). The status of the household in the village does not induce endogeneity. Unlike in the past, the positions of headman and karbari are hereditary now. The change took place during the British rule; prior to this the headman was elected by the people. The position of karbari was introduced by the British. Since both positions are hereditary, it cannot be said that food security gave people the opportunity to be involved in the power structure. The infrastructure variable (VIBR) has been constructed from four other variables distances of the village from the Thana headquarters, nearest bazaar, all-weather pucca road, and bus stand. Since a longer distance implies poorer infrastructure, we first calculate the reciprocal of each variable and then use principal component analysis to calculate a score. Table 3A. Food security status among male- and female-headed households by ethnic group (two broad food security categories) Ethnic group Food insecurity Food security Pearson v 2 (p-value) Male Female Total Male Female Total (1) (2) (3) (4) (5) (6) (7) (8) Bengali (345) (33) (378) (119) (4) (123) Chakma (427) (17) (444) (60) (3) (63) Marma (303) (40) (343) (150) (10) (160) Mro (223) (10) (233) (271) (6) (277) Tripura (398) (22) (420) (87) 8.33 (2) (89) All (1696) (122) (1818) (687) (25) (712) Figures in parentheses are number of households.

6 598 WORLD DEVELOPMENT 7. RESULTS Table 3B. Food security status among male- and female-headed households by ethnic group (four food security categories) Ethnic group Chronic food insecurity Transitory food insecurity Break-even Surplus Pearson v 2 (p-value) Male Female Total Male Female Total Male Female Total Male Female Total (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) Bengali (93) (20) (113) (252) (13) (265) (99) 5.41 (2) (101) 4.31 (20) 5.41 (2) 4.39 (22) Chakma (150) (9) (159) (277) (8) (285) (51) 5.00 (1) (52) 1.85 (9) (2) 2.17 (11) Marma (111) (19) (130) (192) (21) (213) (126) (9) (135) 5.30 (24) 2.00 (1) 4.97 (25) Mro 8.50 (42) (2) 8.63 (44) (181) (8) (189) (214) (6) (220) (57) 0.00 (0) (57) Tripura (283) (19) (302) (115) (3) (118) (74) 8.33 (2) (76) 2.68 (13) 0.00 (0) 2.55 (13) All (679) (69) (748) (1017) (53) (1070) (564) (20) (584) 5.16 (123) 3.40 (5) 5.06 (128) Figures in parentheses are number of households. We report the results for both the ordered probit and generalized threshold models. Although the former model suffers from some limitations, it is frequently estimated and serves as a useful benchmark for comparison. All models include ethnicity and district dummies to account for the ethnic and geographical heterogeneity. (a) Ordered probit model We begin with the results for the ordered probit model. Table 5 reports the results for all ethnic groups. Both the coefficients and elasticities 11 are reported. The results show a positive and significant coefficient of the education of the household head, the involvement in the indigenous power structure, and the amount of both jhume and plain land cultivation. The coefficient of the gender of the household head is also positive and significant at the 1% level. Due to the singlecrossing property, elasticities with respect to these variables are negative for j = 1 (chronic food insecurity) and positive for j = 4 (food surplus) (Boes & Winkelmann, 2006, p. 170; Greene, 2000, p. 877). Conversely, the coefficient of the demographic dependency ratio is negative and significant indicating that elasticity is positive for j = 1 and negative for j = 4. Elasticities with respect to age of the household head and infrastructure are not significant. For the break-even category (j = 3), the male-headed households have about 37% higher probability of food security than the female-headed households. A percentage point increase in educational attainment increases food security by 0.12%. Elasticities with respect to the plain and jhume land cultivation are low at 0.12 and 0.07, respectively. For the households suffering transitory food insecurity (j = 2) none of the elasticities are significant. Next we estimate the same model excluding the Bengali. The results, reported in Table 6, show that the sign and significance of the coefficients and elasticities for different food security categories do not change from those estimated in the full sample. The only major change in the results is that the coefficient of and the elasticity with respect to the gender of the household head are now weakly significant at the 10% level. (b) Generalized threshold model Given the limitations of the ordered probit model, which is our benchmark, we now estimate the generalized threshold model. But first we test which of the two models optimally uses information in our case. The null hypothesis is that there are no category-specific parameters. If this hypothesis is rejected, then it can be considered that the ordered probit model does not optimally use the information contained in the data, but the generalized threshold model does. We test this hypothesis by the likelihood ratio (LR) test, where the test statistic is LR = 2(Log L ordered probit Log L generalized threshold ), which is asymptotically v 2 distributed with (J 2)k degrees of freedom (Boes & Winkelmann, 2006, p. 10). The value of the statistic for the full sample is 176 with 32 degrees of freedom. We therefore, reject the null hypothesis in favor of the generalized threshold model. The results for all ethnic groups are reported in Table 7. There are some important changes in the results from the ordered probit model. The elasticity with respect to the gender of the household head is now insignificant for the food surplus category (j = 4). For the transitory food insecure category (j = 2), for which none of the predictors were signif-

7 ARE FEMALE-HEADED HOUSEHOLDS MORE FOOD INSECURE? EVIDENCE FROM BANGLADESH 599 Table 4. Descriptive statistics of demographic characteristics, cultivation and production Variables All ethnic groups Excluding Bengalis Male Female Male Female (1) (2) (3) (4) (5) Age of the household head (years) (13.312) *** (13.800) (13.201) (12.879) Education of the household head (years of schooling) (2.888) *** (1.948) (2.647) *** (2.030) Household head employed in farm sector (%) a * ** Family size (2.250) ** (1.770) (2.212) (1.834) Ratio of number of family members engaged in (0.235) *** (0.308) (0.231) *** (0.293) income-generating activities to family size Demographic dependency ratio (47.045) ** (55.871) (46.349) ** (51.415) Economic dependency ratio ( ) *** ( ) (127.65) ** (102.11) Amount of plain land cultivated last year (acres) 0.65 (0.21) *** 0.25 (0.59) 0.72 (2.22) ** 0.31 (0.65) Amount of jhume land cultivated last year (acres) 0.98 (2.06) *** 0.28 (0.60) 1.21 (2.23) *** 0.38 (0.67) Sample size Figures in parentheses are standard deviations. a Significance of Pearson v 2 ; otherwise, significance of the t-test for equality of means. * Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level. Table 5. Results for the ordered probit model (all ethnic groups) Independent Coefficient Elasticity variables j = 1 (chronic j = 2 (transitory j = 3 (break-even) j = 4 (food surplus) food insecurity) food insecurity) (1) (2) (3) (4) (5) (6) GEND *** (0.111) *** (0.132) (0.007) *** (0.124) *** (0.248) AGE (0.011) (0.573) (0.014) (0.532) (1.068) AGE (0.000) (0.284) (0.005) (0.263) (0.528) EDU *** (0.009) *** (0.015) (0.002) *** (0.015) *** (0.029) OCCU (0.067) (0.020) (0.000) (0.018) (0.037) STAT *** (0.130) *** (0.005) (0.000) *** (0.004) *** (0.009) DDR *** (0.048) *** (0.032) (0.001) *** (0.030) *** (0.059) PLND *** (0.058) *** (0.048) (0.002) *** (0.044) *** (0.089) JHUM *** (0.019) *** (0.023) (0.001) *** (0.021) *** (0.043) VIBR (0.025) (0.000) (0.000) (0.000) Cut point (0.283) Cut point (0.285) Cut point (0.292) Log likelihood Predicted probabilities N Dependent variable: food security status (1 = chronic food insecurity, 2 = transitory food insecurity, 3 = break-even, and 4 = food surplus). Figures in parentheses are White (1980)-corrected robust standard errors. The model includes ethnicity and district dummies but they are not reported. *** Significant at the 1% level. icant in the ordered probit model, elasticities with respect to the amount of jhume land cultivated and occupation of the household head turn out to be significant. A percentage point increase in the amount of jhume land cultivated increases the likelihood of food security by 0.05%. For the food surplus category, elasticity with respect to the amount of jhume land cultivated becomes insignificant. The household head being employed in the farm rather than nonfarm sector increases the probability of food security by 4% for the transitory food insecure category. Contrastingly, the household head being employed in the nonfarm rather than the farm sector increases the likelihood of food security by 5% and 15% for the break-even and food surplus categories, respectively. Lastly, we estimate the generalized threshold model for only the indigenous ethnic groups. For this reduced sample as well, the likelihood ratio test rejects the null hypothesis of no category-specific parameters, thus suggesting the generalized threshold model as the appropriate one. The value of the LR test statistic is 166 with 30 degrees of freedom (one ethnicity dummy is excluded). The results are reported in Table 8. The most important change in the result is that elasticity with respect to the gender of the household head turns out to be insignificant for all food security categories; it becomes negative but not significant for j =2. We now discuss some general trends in the results. First, elasticities with respect to the age of the household head and infrastructure are insignificant in all models. The CHT is still so underdeveloped that a marginal improvement in infrastructure has no impact on adoption and diffusion of technology or marketing of agricultural products. Second,

8 600 WORLD DEVELOPMENT Table 6. Results for the ordered probit model (Indigenous ethnic groups only) Independent Coefficient Elasticity variables j = 1 (chronic j = 2 (transitory j = 3 (break-even) j = 4 (food surplus) food insecurity) food insecurity) (1) (2) (3) (4) (5) (6) GEND * (0.122) * (0.142) (0.006) * (0.135) * (0.274) AGE (0.013) (0.673) (0.019) (0.637) (1.297) AGE (0.000) (0.332) (0.007) (0.314) (0.639) EDU *** (0.011) *** (0.015) (0.002) *** (0.015) *** (0.030) OCCU (0.085) (0.016) (0.000) (0.015) (0.030) STAT *** (0.130) *** (0.006) (0.001) *** (0.005) *** (0.011) DDR ** (0.055) ** (0.035) (0.002) ** (0.033) ** (0.067) PLND ** (0.055) ** (0.049) (0.002) ** (0.046) ** (0.095) JHUM *** (0.020) *** (0.029) (0.002) *** (0.028) *** (0.056) VIBR (0.025) (0.000) (0.000) (0.000) (0.001) Cut point (0.343) Cut point (0.346) Cut point (0.352) Log likelihood Predicted probabilities N Dependent variable: Food security status (1 = chronic food insecurity, 2 = transitory food insecurity, 3 = break-even, and 4 = food surplus). Figures in parentheses are White (1980)-corrected robust standard errors. The model includes ethnicity and district dummies but they are not reported. * Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level. Independent variables Table 7. Results for the generalized threshold model (all ethnic groups) Elasticity j = 1 (chronic food insecurity) j = 2 (transitory food insecurity) j = 3 (break-even) j = 4 (food surplus) (1) (2) (3) (4) (5) GEND *** (0.144) (0.082) ** (0.179) (0.517) AGE (0.774) (0.421) (0.834) (2.367) AGE (0.391) (0.210) (0.410) (1.126) EDU *** (0.022) (0.011) *** (0.021) *** (0.051) OCCU (0.025) *** (0.013) * (0.026) ** (0.069) STAT *** (0.010) (0.005) *** (0.007) * (0.014) DDR * (0.041) (0.023) ** (0.047) (0.149) PLND *** (0.093) (0.040) (0.057) *** (0.067) JHUM *** (0.060) ** (0.026) *** (0.037) (0.035) VIBR (0.000) (0.000) (0.000) (0.000) Log likelihood Predicted probabilities N Dependent variable: food security status (1 = chronic food insecurity, 2 = transitory food insecurity, 3 = break-even, and 4 = food surplus). Figures in parentheses are White (1980) corrected robust standard errors. The model includes ethnicity and district dummies but they are not reported. * Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level. significance of elasticities with respect to the education and occupation of the household head, the demographic dependency ratio, involvement in the indigenous power structure, and the amount of land cultivated vary across models and food security categories. Finally, the predicted probabilities do not change much across models. Predicted probability is the largest for the transitory food insecure category at around 0.50 and the lowest for the food surplus category at around (c) Robustness checks The result that the gender of the household head is not significant for food security contrasts the conventional idea and therefore, requires further investigation. In the following, we perform several robustness checks. Our attention is on the elasticity with respect to the gender of the household head. When food security is disaggregated into four categories, the number of observations for some categories is very small rela-

9 ARE FEMALE-HEADED HOUSEHOLDS MORE FOOD INSECURE? EVIDENCE FROM BANGLADESH 601 Independent variables Table 8. Results for the generalized threshold model (Indigenous ethnic groups only) Elasticity j = 1 (chronic food insecurity) j = 2 (transitory food insecurity) j = 3 (break-even) j = 4 (food surplus) (1) (2) (3) (4) (5) GEND (0.167) (0.097) (0.212) (0.677) AGE (0.892) (0.507) (0.980) (2.929) AGE (0.450) (0.253) (0.481) (1.366) EDU *** (0.022) (0.011) *** (0.021) *** (0.055) OCCU (0.019) *** (0.010) ** (0.019) ** (0.050) STAT *** (0.013) (0.006) *** (0.008) * (0.018) DDR (0.045) (0.026) ** (0.052) (0.181) PLND *** (0.101) * (0.044) (0.056) *** (0.072) JHUM *** (0.077) ** (0.034) *** (0.046) (0.045) VIBR (0.000) (0.000) (0.000) (0.001) Log likelihood Predicted probabilities N Dependent variable: food security status (1 = chronic food insecurity, 2 = transitory food insecurity, 3 = break-even, and 4 = food surplus). Figures in parentheses are White (1980) corrected robust standard errors. The model includes ethnicity and district dummies but they are not reported. * Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level. Table 9. Results for the binary probit model Independent variables All ethnic groups Indigenous ethnic groups only Coefficient Elasticity Coefficient Elasticity (1) (2) (3) (4) (5) GEND * (0.143) * (0.173) (0.163) (0.197) AGE (0.013) (0.719) (0.016) (0.870) AGE (0.000) (0.351) (0.000) (0.427) EDU *** (0.011) *** (0.020) *** (0.013) *** (0.020) OCCU *** (0.082) *** (0.024) *** (0.095) *** (0.018) STAT *** (0.163) *** (0.006) *** (0.165) *** (0.007) DDR ** (0.065) *** (0.044) *** (0.075) *** (0.049) PLND * (0.074) * (0.060) * (0.068) * (0.062) JHUM *** (0.026) *** (0.031) *** (0.026) *** (0.40) VIBR (0.027) (0.000) (0.028) (0.000) Constant *** (0.352) Log likelihood Predicted probability N Dependent variable: food security status (1 = food security and 0 = no food security). Figures in parentheses are White (1980) corrected robust standard errors. The model includes ethnicity and district dummies but they are not reported. * Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level. tive to others. We estimate a binary probit model by combining four categories into two food secure and insecure. The results, reported in Table 9, support the results of the generalized threshold model. The coefficient of and the elasticity with respect to the gender of the household head are significant for the full sample but become insignificant when only the indigenous groups are retained. In another robustness check, we bootstrap the generalized threshold model by redrawing the sample 100 times. The results (not reported) confirm the robustness of our previous results. 12 It is sometimes argued that expenditure data may fail to reveal the relative poverty of the female-headed households unless economies of scale that accounts for their smaller size are controlled for (Drèze & Srinivasan, 1997). Our estimates are not affected by economies of scale for two reasons. First, we do not use consumption or expenditure data. Second, respondents implicitly incorporate this issue in their perceptions about their own food security. Nonetheless, we re-estimate the model adding the household size. The elasticity with respect to the gender of the household head does not meaningfully change and that with respect to the household size is found to be insignificant except for the break-even category of the indigenous groups (results not reported). (d) Toward an explanation We examine several possible reasons from the literature and investigate those as possible explanations of the results. In the absence of well-developed credit market, poor households smooth income rather than consumption (Morduch,