Analysis of vulnerability to poverty and food insecurity among productive social safety net program participants in Tanzania

Abstract

In the quest to eradicate poverty, Tanzania has been implementing the productive social safety nets (PSSN), among other efforts. However, despite these well-intentioned efforts, there are valid concerns in the literature highlighting the risk that participants may, in the future, plunge back into poverty and food insecurity. This calls for a nuanced understanding of the vulnerable contexts of social safety net program participants. We draw on the 2017-18 Household Budget Survey data to analyze vulnerability to poverty (VP) and food insecurity (VFI) and their determinants among households enrolled in the PSSN program in Tanzania. We follow the vulnerability as expected poverty approach and estimate the determinants using the Tobit model. We found that 13.9 and 20.6 percent of the PSSN participants were highly vulnerable to poverty and food insecurity, respectively, whereas vulnerability was likely to be lower in male-headed households. Residing in rural areas, ageing, household size, food assistance, credit, and occupation were the most important determinants of VP and VFI. Therefore, there is a need for policy efforts focused on enhancing the effectiveness of SSNs to be cognizant of the vulnerability contexts of participants, as well as the differential implications of safety net programs on various household groups.

Keywords:

• Social safety net
• vulnerability
• vulnerability to poverty
• vulnerability to food insecurity
• vulnerability as expected poverty
• Tobit regression

REVIEWING EDITOR:

• George Mudimu, Development Sciences, Marondera University of Agricultural Sciences and Technology, Marondera, Zimbabwe

SUBJECTS:

• African Studies; Development Policy; Rural Development; Economics and Development; Economics

1. Introduction

Poverty and food insecurity continue to affect a sizable proportion of the population in most developing countries, particularly African countries. For example, approximately 460 million people (approximately 30 percent) across the continent live below the extreme poverty line of 1.90 in 2022 (Galal, 2023). Similarly, the Global Report on Food Crises (GRFC) (2022) reported that 140 million people in Africa are facing acute food insecurity, and subsequently, at least one in five Africans will go to bed hungry. Subsequently, in Sub-Saharan Africa (SSA), approximately 123 million people (12% of SSA’s population) were acutely food insecure by 2022 (Baptista et al., 2022). In Tanzania’s mainland, 26.4% and 8.0% of the population live below the basic needs and food poverty lines, respectively (National Bureau of Statistics (NBS), 2019). Despite efforts to fight poverty and food insecurity, the number of people living in poverty and extreme poverty remains high across the country, with over four million living in extreme poverty (World Bank, 2019b).

To address poverty and enhance social well-being, there have been several efforts championed by governments and development partners within the broader framework of Tanzania’s five-year development planning. The productive social safety net (PSSN) program is at the center of these efforts. The PSSN, through instruments such as public works (PW) and conditional cash transfers (CCTs), as is the case for other social safety net programs, was adopted under the premise that inter alia it would support poor households to move out of the chronic poverty and the risk of falling or remaining in poverty or food insecurity in the future (Mahanta & Das, 2021; World Bank, 2018). However, despite these well-intentioned efforts, there are valid concerns in the literature highlighting the risk that program participants may, in the future, plunge back into poverty and food insecurity (Banda, 2021). Although a number of studies in Tanzania have analyzed household vulnerability and subsequently identified diverse factors associated with vulnerability, they do not shed light on sub-populations who are chronically poor and enrolled in social safety net programs (Aikaeli et al., 2021; Innocent, 2020; Pantaleo & Ngasamiaku, 2021; Mutabazi et al., 2015; Pietrelli & Scaramozzino, 2019; Sarris & Karfakis, 2006). In view of the foregoing, a nuanced understanding of the vulnerability contexts of social safety net participants and their drivers is imperative to sustain these policy actions to alleviate poverty and prevent the chronically poor from falling into poverty and food insecurity.

Therefore, this study analyzes the levels of vulnerability to poverty and food insecurity and its determinants among productive social safety net (PSSN) households in Tanzania. This study makes two potential contributions to the literature. First, we identify the vulnerability status among social safety net participating households in the country and indicate those who are most likely to be poor and food insecure in the future – an ex-ante analysis (Celidoni, 2013; Chaudhuri et al., 2002b). Identifying households that are currently chronically poor but are more likely to be non-poor in the near future and those that are currently non-chronically poor but are more likely to be so in the near future is crucial for designing and implementing social safety net policies and programs. Second, we examined the determinants of household vulnerability. These are of great value to policymakers in designing effective SSN programs, both now and in the future.

The remainder of this paper is organized as follows. Section two discusses the concept of vulnerability, and provides more information on the productive social safety net program and how it is implemented in Tanzania. The data sources and methods of analysis are discussed in section three, while results and discussion are presented in section four and five, respectively. The final section (section six) presents conclusions and policy implications.

2. Vulnerability and social safety nets programs in Tanzania

Vulnerability is the likelihood that individual or household welfare will fall below a benchmark at a given time in the future (Hoddinott & Quisumbing, 2010). Individuals and households are considered vulnerable to poverty (VP) or food insecurity (VFI) if they are likely to be poor or food insecure in the near future (Mahanta & Das, 2021). This is because poverty and food security are dynamic rather than static concepts: between one year and the next, many households move into or out of poverty and food insecurity (Thang, 2018; Yaqub, 2000). To date, considerable efforts have been made by governments in Sub-Saharan Africa (SSA) to expand the coverage of social safety nets because these programs hold the premise of vulnerability reduction (Bastagli et al., 2019). For example, since 2000, Tanzania has been implementing more than 20 social safety net programs (Msuha et al., 2024). Hover, the existing studies do not shed light on vulnerability status among the social safety net participating households (Aikaeli et al., 2021; Innocent, 2020; Pantaleo & Ngasamiaku, 2021; Mutabazi et al., 2015; Pietrelli & Scaramozzino, 2019; Sarris & Karfakis, 2006).

The Tanzanian PSSN is one of the most extensive programs in Africa, enrolling more than 1.1 million poor households. The program comprises both conditional cash transfers (CCT) and public work (PWs). The CCT provides a transfer value of up to 16.3 per month per household using a combination of four targeting mechanisms to reach out to the targeted population: administrative targeting, proxy means testing, and geographical and community-based targeting (Msuha et al., 2024). The objectives of the CCTs in Tanzania are mainly to address poverty and vulnerability through enhanced consumption, food security, asset accumulation, and human capital development.

The PWs scheme provides a wage payment either in cash or in kind to a targeted household in return for the provision of labor (McCord, 2008). It is an additional source of income to smooth consumption and mitigate shocks (World Bank, 2019b). PWs hold the premise of smoothing consumption during the lean season and mitigating shocks (World Bank, 2019b), thus addressing poverty and vulnerability (Msuha et al., 2024). To date, three types of PWs schemes have been implemented in Tanzania: cash for work (CFW), food-for-work (FFW), and food for assets (FFA) (Msuha et al., 2024). The PWs schemes under the PSSN program focus on delivering CFW to poor and vulnerable households with at least one adult able to work. The eligible household receives a daily rate of 1.35 USD for 15 days of paid work per month for up to four months during the annual lean season. The scheme combines two targeting mechanisms: geographical and community-based.

The major policy concern is the reliance on households’ current status of poverty and food security as the basis of expanding coverage of social safety nets, which lacks evidence on ex-ante analysis (Lu et al., 2023). A household that is not poor or food insecure this year might be poor or food insecure the following year, and vice-versa (Mahanta & Das, 2021). Hence, expanding social safety net programs without ex ante measurement might leave out a substantial number of poor populations simply because they are non-poor today and subsequently enroll those who are poor today. One year later, those who were considered non-poor might became poor, and those who were considered poor goes above the poverty line, consequently committing type I or II errors in the design and implementation of anti-poverty policy interventions.

3. Materials and methods

3.1. Data

We draw on the 2017-18 Household Budget Survey (HBS) dataset, which is nationally representative data collected by the National Bureau of Statistics (NBS) in Tanzania. The HBS data series are a major source of information for estimating poverty and its associated characteristics in Tanzania. It is conducted separately for the Mainland (covering 26 regions) and Zanzibar (covering five regions). In Tanzania Mainland, the 2017-18 HBS is the latest dataset in the HBS series. The first round of HBS was conducted in 1991, and the NBS completed five rounds of HBS. In Mainland Tanzania, the PSSN module was specifically included in the 2017-18 HBS to measure the impact of the program. Our study covered mainland Tanzania.

The 2018 HBS adopted a two-stage cluster-sample design. The first stage involved the selection of 796 primary sampling units (PSUs) or enumeration areas from the 2012 population and housing censuses. This was followed by a listing exercise in which households residing in the selected PSUs were freshly listed before selecting households. A total of 51 sampling strata were created and a representative probability sample of 9,552 households was selected. This sample was designed to allow separate estimates for each of the 26 regions of mainland Tanzania, as well as urban and rural areas, separately at the national level. In our analysis, we merged and dropped unmatched mergers, resulting in a total sample size of 9,463. From this sample, we identified 844 households that participated in the PSSN program between 2013-2019.

3.2. Estimation strategy

This study follows a two-step analytical framework. In the first step, we estimate the vulnerability index of each household (vulnerability to poverty, VP, and vulnerability to food insecurity, VFI) using the vulnerability as expected poverty (VEP) approach. Following the VEP procedure suggested by Chaudhuri et al. (2002a), Gaiha and Imai (2008), and Günther and Harttgen (2009), the probability of a household falling below TZS 49,320 per adult equivalent per month (VP) and below TZS 33,748 per adult equivalent per month (VFI) was estimated as follows: (1) $${̂\mathrm{\text{VP}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}\left({\mathrm{\text{lnC}}}_{\mathrm{i}}<\mathrm{\text{TZS}}\mathrm{ }49,320\mathrm{\text{Xi}}\right)=\mathrm{Ф}\left(\frac{\mathrm{\text{ln}}\hspace{0.17em}\left(\mathrm{\text{TZS}}\mathrm{ }49,320\right)-{\mathrm{\text{Xi}}}^{̂\mathrm{\beta }}}{{\sqrt{\mathrm{X}\mathrm{i}}̂}^{\mathrm{\theta }}}\right)$$(1) (2) $$̂{\mathrm{\text{VFI}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}({\mathrm{\text{lnC}}}_{\mathrm{i}}<\text{TZS}33,748|\mathrm{\text{Xi}})=\text{Ф}(\frac{\mathrm{\text{ln}}(\mathrm{\text{TZS}}\mathrm{ }33,748)‐\mathrm{\text{Xi}}̂\mathrm{\beta }}{\sqrt{\mathrm{X}\mathrm{i}}̂\mathrm{\theta }})$$(2)

Where: VPht: Vulnerability to poverty of household ‘h’ at a time ‘t’.VFIht: Vulnerability to food insecurity of household ‘h’ at a time ‘t’.${\mathrm{\text{lnC}}}_{\mathrm{i}}$: Log of per capita consumption for the i-th householdTZS 49,320: Tanzanian official basic needs poverty line (NBS, 2019).TZS 33,748: Tanzanian official food poverty line (NBS, 2019).Xi: Set of observable household characteristicsФ(.): Cumulative density of the standard normal distribution functionXh: Current set of observable household characteristics for household h$̂\beta$: Predicted per capita mean consumption (predicted ‘yhat’)$̂\mathrm{\theta }$: Estimated variance of per capita (log) consumption

Equations (1)(1) $${̂\mathrm{\text{VP}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}\left({\mathrm{\text{lnC}}}_{\mathrm{i}}<\mathrm{\text{TZS}}\mathrm{ }49,320\mathrm{\text{Xi}}\right)=\mathrm{Ф}\left(\frac{\mathrm{\text{ln}}\hspace{0.17em}\left(\mathrm{\text{TZS}}\mathrm{ }49,320\right)-{\mathrm{\text{Xi}}}^{̂\mathrm{\beta }}}{{\sqrt{\mathrm{X}\mathrm{i}}̂}^{\mathrm{\theta }}}\right)$$(1) and (2)(2) $$̂{\mathrm{\text{VFI}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}({\mathrm{\text{lnC}}}_{\mathrm{i}}<\text{TZS}33,748|\mathrm{\text{Xi}})=\text{Ф}(\frac{\mathrm{\text{ln}}(\mathrm{\text{TZS}}\mathrm{ }33,748)‐\mathrm{\text{Xi}}̂\mathrm{\beta }}{\sqrt{\mathrm{X}\mathrm{i}}̂\mathrm{\theta }})$$(2) were estimated using a three-step feasible generalized least squares (FGLS) procedure (Amemiya, 1977). Finally, the household is considered vulnerable if the estimated VP/VFI is above or equal to 0.5 (Chaudhuri et al., 2002a). In the second step, we estimate the determinants of VP and VFI among productive social safety net program participants using Tobit regression model following Wooldridge (2002), the two-limit Tobit model as follows: (3) $$V_i^{*}={\mathrm{\beta }}_1{\mathrm{X}}_{\mathrm{i}}{\text{+e}}_{\mathrm{i}}{\mathrm{e}}_{\mathrm{i}}\text{|}{\mathrm{X}}_{\mathrm{i}}\sim \text{Normal}\mathrm{(}\text{0,}\delta 2\mathrm{)}$$(3) (3a) $$V_i=0\text{if}{V}_i^{\mathrm{*}}\le 0\mathrm{ }$$(3a) (3b) $$V_i=V_i^{\mathrm{*}}\text{if}0<V_i^{\mathrm{*}}<1\mathrm{ }$$(3b) (3c) $$V_i=1\text{if}{V}_i^{\mathrm{*}}\ge 1$$(3c)

Where $V_i^{\mathrm{*}}$ denotes the latent-value vulnerability index, $V_i$ is a vulnerability index used as a dependent variable, β is a parameter to be estimated, ${\mathrm{\epsilon }}_i$ is an independently distributed error term that is assumed to be normally distributed with zero mean and constant variance, σ2. The Tobit model in Equation 3(3) $$V_i^{*}={\mathrm{\beta }}_1{\mathrm{X}}_{\mathrm{i}}{\text{+e}}_{\mathrm{i}}{\mathrm{e}}_{\mathrm{i}}\text{|}{\mathrm{X}}_{\mathrm{i}}\sim \text{Normal}\mathrm{(}\text{0,}\delta 2\mathrm{)}$$(3) is estimated with a maximum likelihood estimation procedure, which is a general method for obtaining parameter estimates and performing statistical inference on the estimates (Amemiya, 1977). Appendix A presents the detailed estimation procedure for equations (1-3).

4. Results

4.1. Descriptive results

Table 1 provides a summary of the descriptive statistics. The average age of the heads of PSSN households was 55.2 years. Female-headed households were older (58.3 years) than their male counterparts (52.9 years). It was also noted that on average, those who were older lived in urban areas. Most PSSN households were headed by men (57.8 percent), with 42.2 percent being headed by females. The average household size in adult equivalents among PSSN households was 5.2, and male-headed households had more families than female-headed households did (Table 3).

Table 1. Descriptive statistics.

Variable	Sex / Location		t statistic /χ2 statistic	P-value
	Female	Male
Panel A:
Age of HH head (Years)	58.33	52.85	−52.92	0.000
Family size (adult equivalent)	4.57	5.62	68.28	0.000
Dependency ratio (ratio)	1.67	1.55	−16.12	0.000
Received education subsidy (1 = yes)	2.65	3.00	4.67	0.000
Received health subsidy (1 = yes)	2.17	1.82	23.22	0.000
Received credit (1 = yes)	0.32	2.37	98.55	0.000
Having occupation (1 = yes)	0.41	0.77	2.96	0.000
Panel B:	Urban	Rural
Age of HH head (Years)	56.36	54.83	−13.40	0.000
Family size (adult equivalent)	5.02	5.22	9.12	0.000
Dependency ratio (ratio)	1.34	1.67	37.76	0.000
Received education subsidy (1 = yes)	1.82	3.83	31.32	0.000
Received health subsidy (1 = yes)	0.28	3.71	53.02	0.000
Received credit (1 = yes)	0.32	2.38	17.17	0.000
Having occupation (1 = yes)	0.57	0.61	43.66	0.000
Panel C:	PSSN	Non-PSSN
Vulnerability to poverty (VP Index)	0.34	0.39	−31.47	0.000
Vulnerability to food insecurity (VFI Index)	0.41	0.46	−32.13	0.000

On average, the dependency ratio was 1.6, suggesting that there were approximately 160 dependents (0-14 and 65+) for every 100 working-age adults among PSSN households. Every PSSN household member aged between 15 and 64 years had to generate income for themselves, and an additional 1.6 persons. Further analysis in Table 3 indicates a significantly higher dependency ratio in female-headed households than in their male-headed counterparts, whereas the dependency ratio was higher in rural families.

The descriptive results also indicated that the majority of PSSN households resided in rural areas, with only 22.4 per cent living in urban areas. Since PSSN targets the chronically poor population, the result is in line with the National Bureau of Statistics (2020) on the distribution of the poor population on Tanzania’s mainland, where 81 percent of them reside in rural areas. Most PSSN programme participants (51.5 percent) owned livestock, which could be attributed to the investment undertaken after receiving cash transfers. With PSSN, it is hypothesized that when beneficiary households receive cash transfers apart from increasing consumption, they are expected to be used in ways that have immediate effects on increasing savings and investment in productive assets (World Bank, 2019a). We conducted further analysis to compare vulnerability to poverty (VP) and food insecurity (VFI) between PSSN participants and non-participants within the poor and vulnerable sub-populations. The results indicated that the VP and VFI were lower among PSSN participants than their non-participant counterparts. This is probably an indication that the PSSN program reduces the probability of households remaining poor or food insecure in the future.

4.2. Vulnerability to poverty and food insecurity status among PSSN participants

In this section, we present the results of the analysis of household VP and VFI status among the PSSN participants. Following the VEP approach, we estimated the probability of a household falling below TZS 49,320 per adult equivalent per month (VP) and below TZS 33,748 per adult equivalent per month (VFI), as described in Equations (1)(1) $${̂\mathrm{\text{VP}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}\left({\mathrm{\text{lnC}}}_{\mathrm{i}}<\mathrm{\text{TZS}}\mathrm{ }49,320\mathrm{\text{Xi}}\right)=\mathrm{Ф}\left(\frac{\mathrm{\text{ln}}\hspace{0.17em}\left(\mathrm{\text{TZS}}\mathrm{ }49,320\right)-{\mathrm{\text{Xi}}}^{̂\mathrm{\beta }}}{{\sqrt{\mathrm{X}\mathrm{i}}̂}^{\mathrm{\theta }}}\right)$$(1) and (2)(2) $$̂{\mathrm{\text{VFI}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}({\mathrm{\text{lnC}}}_{\mathrm{i}}<\text{TZS}33,748|\mathrm{\text{Xi}})=\text{Ф}(\frac{\mathrm{\text{ln}}(\mathrm{\text{TZS}}\mathrm{ }33,748)‐\mathrm{\text{Xi}}̂\mathrm{\beta }}{\sqrt{\mathrm{X}\mathrm{i}}̂\mathrm{\theta }})$$(2) , respectively. The VP and VFI status of each household was determined using a score of 0.5 as the threshold level (Pritchett et al., 2000). Based on this classification, a household was referred to as having low vulnerability to poverty (VP) or vulnerability to food insecurity (VFI) if the vulnerability score (VP/VFI) was less than 0.5 and classified as highly vulnerable when the score (VFI) was greater than or equal to 0.5. Vulnerability status was disaggregated by rural-urban location and sex, and the results are presented in Panels A, B, and C (Table 2). Panel A presents VP and VFI for PSSN households that received either CCTs or CCTs + PWs. Panel B shows households that received only CCTs, and Panel C shows households that received CCTs + PWs.

Table 2. Classification of households’ vulnerability status.

		Vulnerability to poverty			Vulnerability to food insecurity
		High	L ow	χ2-value	High	Low	χ2-value
Panel A: PSSN
Sex of household head	Female	4.7	37.6		5.5	36.8
	Male	9.1	48.5		14.7	43.0
	Total	13.9	86.1	42.1***	20.2	79.8	221.3***
Location	Urban	0.6	21.8		1.7	20.7
	Rural	13.3	64.3		18.5	59.1
	Total	13.9	86.1	279.4***	20.2	79.8	267.6***
Panel B: CCTs Only
Sex of household head	Female	5.0	38.1		5.8	37.3
	Male	9.7	47.2		14.8	42.1
	Total	14.7	85.3	53.8***	20.6	79.4	218.7***
Location	Urban	0.7	22.3		1.8	21.1
	Rural	14.1	63.0		18.8	58.3
	Total	14.7	85.3	311.0***	20.6	79.4	277.3***
Panel C: CCTs + PWs
Sex of household head	Female	1.7	35.3		1.7	35.2
	Male	9.0	54.0		15.2	47.9
	Total	10.7	89.3	202.7***	16.9	83.1	532.8***
Location	Urban	1.2	16.3		2.1	15.4
	Rural	9.5	73.0		14.8	67.7
	Total	10.7	89.3	27.0***	16.9	83.1	31.0***

Note. *** p < 0.01, ** p < 0.05, * p < 0.1. Figures are in percent.

Table 3. Determinants of VP and VFI among PSSN program participants.

Explanatory Variable	Dependent Variable: VP			Dependent Variable: VFI
	(1)	(2)	(3)	(1)	(2)	(3)
Age of HH head	0.0010*** (0.0000)	0.0010*** (0.0000)	0.0010*** (0.0000)	0.0001*** (0.0000)	0.0001*** (0.0000)	0.0001*** (0.0000)
Sex of HH head	−0.0224*** (0.0002)	−0.0235*** (0.0002)	−0.0264*** (0.0006)	−0.0189*** (0.0003)	−0.0192*** (0.0004)	−0.0261*** (0.0004)
Marital status	−0.0219*** (0.0003)	−0.0223*** (0.0002)	−0.0208*** (0.0006)	−0.0452*** (0.0003)	−0.0451*** (0.0003)	−0.0466*** (0.0004)
Family size	0.0564*** (0.0001)	0.0568*** (0.0001)	0.0548*** (0.0002)	0.0647*** (0.0001)	0.0647*** (0.0001)	0.0683*** (0.0001)
Dependency ratio	0.0494*** (0.0001)	0.0503*** (0.0001)	0.0464*** (0.0002)	0.0424*** (0.0001)	0.0426*** (0.0001)	0.0408*** (0.0001)
Location	0.1524*** (0.0004)	0.1532*** (0.0004)	0.1420*** (0.0007)	0.1138*** (0.0002)	0.1146*** (0.0002)	0.1067*** (0.0005)
Credit	−0.1310*** (0.0009)	−0.1374*** (0.0008)	−0.1110*** (0.0015)	−0.0915*** (0.0004)	−0.0980*** (0.0003)	−0.0798*** (0.0009)
Business ownership	−0.0032*** (0.0003)	−0.0027*** (0.0003)	−0.0059*** (0.0005)	−0.0001 (0.0002)	−0.0002 (0.0002)	0.0003 (0.0003)
Livestock ownership	−0.0812*** (0.0002)	−0.0822*** (0.0002)	−0.0794*** (0.0003)	−0.0766*** (0.0002)	−0.0772*** (0.0002)	−0.0742*** (0.0003)
Occupation	−0.1565*** (0.0022)	−0.2033*** (0.0022)	−0.1561*** (0.0030)	−0.1561*** (0.0011)	−0.1746*** (0.0008)	−0.1620*** (0.0017)
Education subsidy	−0.0207*** (0.0003)	−0.0191*** (0.0003)	−0.0192*** (0.0005)	0.0072*** (0.0002)	0.0077*** (0.0002)	0.0044*** (0.0003)
Health Subsidy	−0.0171*** (0.0003)	−0.0167*** (0.0003)	−0.0117*** (0.0005)	0.0104*** (0.0002)	0.0106*** (0.0002)	0.0202*** (0.0004)
Food assistance	−0.2126*** (0.0019)	−0.2277*** (0.0022)	−0.1724*** (0.0005)	−0.1394*** (0.0008)	−0.1313*** (0.0011)	−0.1576*** (0.0004)

Note. **, **, and * denote the statistically significant variables at the 1%, 5%, and 10% levels, respectively. The figures in parentheses are the standard errors. Columns (1), (2), and (3) are Average Marginal Effects (AME). Columns (2) and (3) are robustness checks. They show that the effect of explanatory variables on dependent variables in the regression is the same, regardless of whether households are enrolled in CCTs only (Model 2) or in PWs and CCTs (Model 3). The interpretation of the results and discussion are based on Model (1).

The results in Table 2 show that 13.9 and 20.6 percent of PSSN households were highly vulnerable to poverty (VP) and food insecurity (VFI) respectively, whereas 86.1 and 79.8 per cent of PSSN households had low vulnerability to poverty and food insecurity respectively. We also found that the prevalence of VP and VFI was higher for households enrolled in CCT only than for those enrolled in both CCTs and PWs. This could be an indication that CCT + PW has a greater impact than CCT alone. The findings further revealed that the prevalence of VP and VFI is significantly related to the location and sex of the household. For example, most households in rural areas are more vulnerable than those in urban areas.

4.3. Econometric results on the drivers of vulnerability to poverty and food insecurity

To gain insight into the determinants of vulnerability to poverty and food insecurity among PSSN Program participants, the Tobit model estimation procedure described in Equation (8) was employed. Table 3 presents the results of Tobit regression analysis. The overall model fit was significant at the one percent (Prob > F = 0.000), implying that the independent variables were jointly important determinants of VP and VFI (Tables B1–B3 in Appendix B). Thus, the hypothesis that none of the explanatory variables were related to VP and VFI was rejected, permitting the interpretation of the results. The average marginal effect (AME) of the explanatory variables on VP and VFI are presented in column (1), whereas columns (2) and (3) are robustness checks, which are described in detail in Section (4.4). The AME can be interpreted as the average change in probability when the explanatory variable increases by one unit. The complete regression results are presented in Tables B1–B3 in Appendix B.

The results in Column (1) of Table 3 show eight (8) variables namely. Sex of the household head, marital status, credit, livestock ownership, occupation, education, health subsidies, and food assistance were identified as determinants with an inverse relationship with vulnerability to poverty (VP) and food insecurity (VFI). For example, the gender of the household head negatively affected households’ VP and VFI at the 1 percent level of significance. This result can be interpreted as male-headed PSSN households being 2.2 and 1.9 percent less likely to fall below or stay in basic need poverty and food poverty lines, respectively, in the future compared to their female counterparts, Ceteris Paribus. With respect to marital status, the Tobit result suggests that VP and VFI among PSSN households who were separated, divorced, widowed or never married were more likely to be less vulnerable by 2.2 and 4.5 percent respectively compared to those who were monogamous and polygamous married or living together, other factors held constant. Access to credit decreases households’ VP and VFI by 13.1 and 9.2%, respectively, compared to their counterparts with no access to credit (Ceteris Paribus. The results also indicated that livestock ownership, occupation, provision of education and health subsidies, and food assistance had significant effects on reducing VP and VFI at the one percent probability level. Among the eight variables identified to have a negative effect, the provision of food assistance seems to have greater VP and VFI reduction effects (21.3 percent), followed by occupation (15.7 percent) and credit (13.1 per cent).

The econometric results in Table 3 indicate that four variables, namely, aging, household size, dependency ratio, and residing in rural areas, had a positive influence on households’ VP and VFI. For example, a one-year increase in the age of household heads among PSSN households increased VP and VFI by 0.1 percent, Ceteris Paribus. The results suggest that older household heads among PSSN households are more likely to be vulnerable to poverty and food insecurity than younger heads. With respect to household size, the result indicates that one extra person in the PSSN household increased VP and VFI by 5.6 and 6.5%, respectively, while other factors held constant. Similarly, our results suggest that a unit increase in the proportion of economically inactive labor force to the active labor force among PSSN households increases VP and VFI by 4.5 and 4.2%, respectively. This increase is highly significant at the one percent level. On the other hand, the marginal effect suggests that vulnerability to poverty and food insecurity among PSSN households located in rural areas was 15.2 and 11.4% higher, respectively, compared to their counterparts located in urban areas. This difference was statistically significant at the one percent level. Among the four variables, the rural-urban dichotomy was found to have a significantly higher positive effect on VP and VFI among PSSN households.

4.4. Robustness checks

The econometric results indicated that VP and VFI were influenced by the age of the household head, sex of the household head, marital status, household size, dependency ratio, location, credit, livestock ownership, occupation, education and health subsidies, and food assistance (Table 3). We conduct several robustness checks to test the validity of our results. First, we checked whether these determinants were driven by confounding factors (treatment levels). For example, the determinants of VP and VFI among PSSN households enrolled in CCTs might be different from PSSN households enrolled in CCTs + PWs. Therefore, we confirmed the robustness of our results by running a secondary analysis restricting the sample households that received CCTs only and repeating a similar analysis to households that received CCTs + PWs. The results for this subset of treatments were consistent with those of our main analysis, indicating that our results were robust to this concern. Second, we conducted a two-sample t-test analysis as a mean-comparison test for categorical variables to determine whether there was a significant difference in VP and VFI as a result of their influence. The results went a long way confirm that the VP and VFI of households are significantly influenced by similar factors identified in the Tobit model estimates. This result confirms the validity of our results. Appendix B provides the results of the secondary analysis.

5. Discussion

This study found evidence that some PSSN households in Tanzania are highly vulnerable to both poverty (VP) and food insecurity (VFI). The results indicate that 13.9 and 20.6 percent of households enrolled in PSSN programs were highly vulnerable to VP and VFI, respectively. These results suggest that these groups of households have little chance of escaping poverty and food insecurity in the near future (Sileshi et al., 2019). Contrary to this subgroup, the results also show that 86.1 and 79.8 per cent of PSSN households had low vulnerability to poverty and food insecurity, respectively, suggesting that they were suffering from transient poverty and food insecurity. For this category, even if their current expenditure value was below the poverty line or food poverty line they were ‘less likely’ to continue being poor or food insecure-implying that they can entirely escape from poverty and food insecurity in the near future (Sileshi et al., 2019). Generally

These results suggest that, although the PSSN holds the premise of vulnerability reduction (Bastagli et al., 2019), not every household enrolled in the PSSN is affected by the program in the same manner. While other PSSN participants seemed to get rid of poverty, others did not. We note that this variation is mostly due to differences in the gender of the household head, rural-urban dichotomy, ageing, and household size (Pantaleo & Ngasamiaku, 2021). Similar results have been found by Sabates‐Wheeler (2021) in Ethiopia, where some PSNP participants were unable to escape poverty after years of program implementation. Other scholars have recently found similar results related to key drivers of vulnerability differentials in anti-poverty reduction policies (Eshetu & Guye, 2021; Pantaleo & Ngasamiaku, 2021; Peng, 2022; Wang et al., 2024). Variations in PSSN impacts on poor households have important implications for the effective design and implementation of social safety nets (Msuha et al., 2024) and subsequently facilitate the efficient use of limited resources (Schochet et al., 2014).

With respect to gender, we found evidence that male-headed households were less likely to fall below or stay in the basic needs poverty and food poverty lines in the future compared to their female counterparts. While this cross-sectional analysis did not capture causal effects, we aimed to determine whether key correlations were consistent between male and female PSSN recipients. This pattern is of interest to policymakers who want to understand whether the gender of the household influences the likelihood of falling or staying in poverty and food insecurity. Our findings are in line with Aikaeli et al. (2021) in Tanzania, Azeem et al.,(2017) in Pakistan, Ojo (2019), Addai et al.,(2022) in Ghana, Agidew and Singh (2018), Eshetu and Guye (2021) in Southern Ethiopia, and Babatunde et al.,(2008) in Nigeria, although they did not focus on households enrolled in SSN programs. The reasons for this phenomenon are not straightforward, but we critically examined some of the characteristic features of male-headed PSSN households in our study. We found that the majority of them had occupation and access to credit; thus, it is logical to see why they were less likely to be VP and VFI, suggesting that PSSN efforts are required to generate additional household economic activities and invest in rural services, where most seem highly vulnerable (Sabates‐Wheeler et al., 2021).

Other factors that significantly contributed to variations in VP and VFI included residing in rural areas, ageing, and household size. Specifically, the majority of households residing in rural areas are more vulnerable to poverty and food insecurity than their counterparts in urban areas are. These results support the view that poverty is predominantly a rural phenomenon in Tanzania (Aikaeli et al., 2021; NBS, 2019). Other studies have consistently reported the prevalence of VFI in rural areas of the Global South (Kate et al., 2019; Mthethwa & Wale, 2021). The age of the household head was more likely to have higher vulnerability to poverty and food insecurity among PSSN households. This is consistent with the widely accepted view that as age increases, households become less productive, and their ability to accumulate assets tends to decrease (Bukenya, 2017; Sileshi et al., 2019).

These results are surprising, especially regarding the positive influence of household size on the VP and VFI. This is unexpected, as we might expect PSSN households with more members to have more transfers, and hence, the possibility of increased welfare. This is because the PSSN household receives a fixed transfer for each child aged 0-18 years old (child grant), a variable transfer paid depending on the number of children apart from those aged 0-18 years, and a disability grant paid for a household with a disability. Subsequently, those with labor capacity receive PWs transfers (World Bank, 2019b). While we are, unable to unpack this mystery, most PSSN households could probably be composed of large numbers of non-children and non-productive members who are also not eligible for PWs transfer. Thus, this imposes pressure on limited income to purchase food and meet basic needs. This is consistent with the resource dilution hypothesis, which postulates that parental resources are limited and that, as the number of children in the family increases, the resources accrued by any one child decline (Jæger, 2009).

6. Conclusion and policy implications

This study on vulnerability to poverty, food insecurity, and their determinants represents the first vulnerability analysis to focus on chronically poor households enrolled in a social safety net program (PSSN) to track changes attributable to the program. Below, we summarize several key takeaways of our analysis. Using vulnerability as expected poverty in the 2017-18 HBS dataset, we estimated the probability of a household falling below TZS 49,320 per adult equivalent per month as vulnerability to poverty (VP) and below TZS 33,748 per adult equivalent per month as vulnerability to food insecurity (VFI). A household with a probability score (VP/VFI) greater than or equal to 0.5 was classified as being highly vulnerable, and less than 0.5 was classified as being less vulnerable. Based on these findings, we found that 13.9 and 20.6 percent of households enrolled in the PSSN program were highly vulnerable to VP and VFI, respectively, whereas the rest were less vulnerable. Using a Tobit regression model, we found that the prevalence of vulnerability to poverty and food insecurity is likely to be lower in PSSN male-headed households than in their female-headed counterparts. Generally, the Tobit regression results showed that VP and VFI were likely to be higher among PSSN households residing in rural areas and significantly lower among those who received food assistance, credit, and having occupations.

The policy implications of these findings are as follows: Although poor households have been enrolled in the PSSN program for more than five years (World Bank, 2019a), not every household is affected by the program in the same manner. While some participants are still highly VP and VFI, others seem to be less vulnerable – variation exists. Considering this variation and its causes is critical in informing decisions on how to best target PSSN interventions and improving the design and/or implementation of PSSN in Tanzania. Consequently, PSSN intervention efforts are likely to be impactful for participants if they target mostly female-headed households, those residing in rural areas, and households with larger family sizes. Further, the provision of public work schemes in terms of food for work (FFW) and other productive livelihood support packages, including credit, can be important instruments to address vulnerability to poverty and food insecurity.

Authors’ contributions

Basil Msuha: Conceptualization; writing-original draft; methodology; formal analysis; discussion and conclusion: Luitfred D. Kissoly: Supervision, review, and editing

Disclosure statement

The authors declare that they have no affiliations with or involvement in any organization or entity with any financial interest in the subject matter or materials discussed in this manuscript.

Data access statement

Research data supporting this publication are available from the National Bureau of Statistics repository located at: https://www.nbs.go.tz/index.php/en/census-surveys/poverty-indicators-statistics/household-budget-survey-hbs/477-the-2017-18-household-budget-survey-dataset

Notes

1 Spatial and temporal price deflation are needed to make the welfare aggregates comparable across time and space.

Appendices

Appendix A.

Methods of Analysis

A| Analytical Framework

This study follows a two-step analytical framework. In the first step, we estimate the vulnerability index of each household (vulnerability to poverty, VP, and vulnerability to food insecurity, VFI) using the vulnerability as expected poverty (VEP) approach. In the second step, we estimated the Determinants of VP and VFI among Productive Social Safety Net Program Participants using a Tobit regression model. Therefore, we first describe the VEP method followed by the Tobit model.

A1| Measuring Households’ Vulnerability: The VEP Approach

In economic literature, there are three principal methods for measuring vulnerability - vulnerability to poverty (VP) and vulnerability to food insecurity (VFI). These are vulnerability as expected poverty (VEP), vulnerability as low expected utility (VEU), and vulnerability as uninsured exposure to risk (VER) (Chaudhuri et al., 2002a; Hoddinott & Quisumbing, 2010). In both approaches ‘vulnerability’ can only be quantified by making some of the assumptions which are: estimation of expected consumption per capita (E(c_t+1)), its variance (σ²), and the poverty line – basic needs or food poverty line (z), and assuming that the consumption per capita (or its log) is normally distributed (Jonathan & Shahidur, 2009). With these parameters, one can estimate the probability that a household will be poor (vht), and thereafter determine whether the household may be considered vulnerable. VP and VFI are due to either low expected consumption or high variability in consumption (Jonathan & Shahidur, 2009). With regard to data requirements, VEP can be evaluated using both cross-sectional and panel data, whereas VEU and VER require panel data.

Because the current study used a household budget survey dataset, which is cross-sectional data, the VEP approach was used to estimate the VP and VFI for each household. Following the VEP approach, VP or VFI is theoretically expressed as (AP-1) $${\mathit{\text{VI}}}_{\mathit{\text{ht}}}=\mathrm{\text{Pr}}(C_{h,t+1}<\text{z}|\mathrm{\text{Xh}})$$(AP-1)

Where: VI_ht denote vulnerability index (VP or VFI) of household ‘h’ at a time ‘t’. C_h,t+1 denote The future total or food consumption per adult equivalent (at a time ‘t + 1’); Z denote poverty line (TZS 49,320 per adult per month) or food poverty line (TZS 33,748 per adult per month); and X_h denote current set of observable household characteristics for household h.

Equation (AP-1) implies that: The vulnerability (VP or VFI) in a household (${\mathrm{\text{VI}}}_{\mathrm{\text{ht}}}$) is determined by the likelihood that the future household per capita consumption (C_h,t+1) will be less than the established basic needs or food poverty line (Z). Estimating the VI involves determining the probability distribution of future consumption (total consumption or food consumption). The Tanzanian basic needs a poverty line or a food poverty line as a measure of household welfare. We assume that household consumption (total or food consumption) is log-normally distributed. Then, following the VEP approach, the probability of a household falling below TZS 49,320 per adult equivalent per month (VP) and below TZS 33,748 per adult equivalent per month (VFI) was estimated using equations (AP-2a) and (AP-2b), respectively: (AP-2a) $$̂{\mathrm{\text{VP}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}\left({\mathrm{\text{lnC}}}_{\mathrm{i}}<\mathrm{\text{TZS}}\mathrm{ }49,320\mathrm{\text{Xi}}\right)=\mathrm{Ф}\left(\frac{\mathrm{\text{ln}}(\mathrm{\text{TZS}}\mathrm{ }49,320)‐\mathrm{\text{Xi}}̂\mathrm{\beta }}{\sqrt{\mathrm{X}\mathrm{i}}̂\mathrm{\theta }}\right)$$(AP-2a) $̂{\mathrm{\text{VFI}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}\mathrm{ }({\mathrm{\text{lnC}}}_{\mathrm{i}}<\text{TZS}33,748|\mathrm{\text{Xi}})=\mathrm{Ф}(\frac{\mathrm{\text{ln}}\hspace{0.17em}\left(\mathrm{\text{TZS}}\mathrm{ }33,748\right)‐\mathrm{\text{Xi}}̂\mathrm{\beta }}{\sqrt{\mathrm{X}\mathrm{i}}̂\mathrm{\theta }})$ (AP-2b)

Where:VP_ht: Vulnerability to poverty of household ‘h’ at a time ‘t’.VFI_ht: Vulnerability to food insecurity of household ‘h’ at a time ‘t’.${\mathrm{\text{lnC}}}_{\mathrm{i}}$: Log of per capita consumption for the i-th householdTZS 49,320: Tanzanian official basic needs poverty line (NBS, 2019).TZS 33,748: Tanzanian official food poverty line (NBS, 2019).X_i: Set of observable household characteristicsФ(.): Cumulative density of the standard normal distribution functionX_h: Current set of observable household characteristics for household _h$̂\beta$: Predicted per capita mean consumption (predicted ‘yhat’)$̂\mathrm{\theta }$: Estimated variance of per capita (log) consumptionequations (AP-2a) and (AP-2b) were estimated using a three-step feasible generalized least squares (FGLS) procedure (Amemiya, 1977). Finally, the household was considered vulnerable if the estimated VP/VFI was above or equal to 0.5 (Chaudhuri et al., 2002a). However, to operationalize equations (AP-2a) and (AP 2.2b), we followed the procedure suggested by Chaudhuri et al. (2002a), Gaiha and Imai (2008), and Günther and Harttgen (2009). The natural log of consumption expenditure per adult equivalent per month, which was spatially and temporally deflated,¹ was regressed using a set of observable household characteristics (independent variables). (AP-3) $$\mathrm{\text{ln}}\hspace{0.17em}{\mathrm{C}}_{\mathrm{i}}={\mathrm{\beta }}_0{{+\mathrm{\beta }}_1\mathrm{X}}_{\mathrm{i}}+{\mathrm{e}}_{\mathrm{i}}\mathrm{ }$$(AP-3)

Where:$\mathrm{\text{ln}}\hspace{0.17em}{\mathrm{C}}_{\mathrm{i}}$: Log of per capita consumption for the i-th householdXi: A set of observable household characteristics$\mathrm{\beta }$: A vector of parameters to be estimatede_i: A disturbance term with a mean zero and assuming heteroscedastic (non-homoscedastic)

The disturbance term in equation (AP-3) is heteroscedastic, referring to a condition in which the variance of the residual or error term varies widely (nonconstant). This implies that the error term variances vary across households, depending on Xi. Hence, creating technical difficulties with a typical multiple linear regression model, if estimated as is, may yield biased and inefficient estimates. Therefore, to correct this problem, the squared residuals from equation (AP-3) are regressed with a set of observable household characteristics (Xi), resulting in equation (AP-4). The resulting fitted values of this regression are known as estimates of σ²_e. (AP-4) $${\delta }_{e_i}^2={\theta }_0{{+\theta }_{1}X}_i+{\eta }_i$$(AP-4)

Where:${\delta }_{e_i}^2$: Squared residuals from equation (AP-3)Xi: A set of observable household characteristicsθ: A vector of parameters to be estimated${\eta }_i$: A disturbance term with a mean zero is assumed to be heteroscedastic (non-homoscedastic)

While the standard Ordinary Least Squares regression (OLS) techniques assume homoscedasticity, equations (AP-3) and (AP-4) assume unequal variance of the error term; that is, they are not equal across households (heteroscedastic rather than homoscedastic) but depend on $X_i$ and reflect the impact of shocks on household consumption (Günther & Harttgen, 2009; Sileshi et al., 2019). Since heteroscedasticity is assumed, using OLS to estimate $\mathrm{\beta }$ leads to unbiased, but inefficient coefficients (Günther & Harttgen, 2009). An econometric method that allows for a heteroscedastic standard error is then used to estimate Equations (AP-3) and (AP-4). Thus, the often-used method is the three-stage Feasible General Least Squares (FGLS) (Chaudhuri et al., 2002a; Christiaensen & Subbarao, 2005).

To conduct a three-stage FGLS, first, we first estimate equation (AP-3) using Ordinary Least Squares (OLS). Second, Equation (AP-4) was estimated using the squared residuals from Equation (AP-3) as the dependent variable. Third, to obtain consistent and asymptotically efficient estimators of θ and β, equation (AP-4) was re-estimated with OLS using predictions obtained from equation (AP-3) after weighting each residual by xiθ, and equation (AP-3) was re-estimated after using efficient θ and weighted least squares (Chaudhuri et al., 2002a; Sileshi et al., 2019).

Finally, Eq. (3)(3) $$V_i^{*}={\mathrm{\beta }}_1{\mathrm{X}}_{\mathrm{i}}{\text{+e}}_{\mathrm{i}}{\mathrm{e}}_{\mathrm{i}}\text{|}{\mathrm{X}}_{\mathrm{i}}\sim \text{Normal}\mathrm{(}\text{0,}\delta 2\mathrm{)}$$(3) was transformed to estimate the $\mathrm{\beta }$ as follows: (AP-5) $$\frac{\mathrm{\text{ln}}\hspace{0.17em}{\mathrm{C}}_{\text{i}}}{\text{ˆσe},\mathrm{i}\mathrm{ }}=\left(\frac{{\mathrm{X}}_{\mathrm{i}}}{\text{ˆσe},\mathrm{i}\mathrm{ }}\right)\mathrm{\beta }+\frac{{\mathrm{e}}_{\mathrm{i}}}{\mathrm{\sigma }\mathrm{e},\mathrm{i}}$$(AP-5)

Using the estimates β and θ, the next step is to compute the expected mean (equation AP-6) and variance (equation AP-7) of consumption for each household using consistent and asymptotically efficient estimators, as follows: (AP-6) $$\mathrm{E}[\mathrm{\text{ln}}\hspace{0.17em}{\mathrm{C}}_{\mathrm{i}}\left|X_i\right]={̂\beta }_0{{+̂\beta }_{1}X}_i\mathrm{ }$$(AP-6) (AP-7) $$\mathrm{V}\left[\mathrm{\text{ln}}\hspace{0.17em}{\mathrm{C}}_{\mathrm{i}}\right|X_i]={̂\delta }_{e_i}^2=̂{\theta }_0{{+̂\theta }_{1}X}_i$$(AP-7)

Finally, assuming ‘lnC_i’ is normally distributed, and using the estimated expected mean in equation (AP-6) and the variance of consumption in equation (AP-7). We can now estimate the probability of a household falling below the poverty line in the future using equations (AP-2a) and (AP-2b), as previously described.

A2| Evaluating determinants of vulnerability: Tobit model

After estimating the vulnerability index for each household, the determinants of vulnerability were evaluated using the two-limit Tobit model pioneered by Tobin (1958). The Tobit model, also called the censored regression model, is recommended when the dependent variable is censored from below, above, or both (Wooldridge, 2002). From Wooldridge’s (2002) perspectives, censoring from above happens when cases with a value ‘at or above’ some threshold, all take on the value of that threshold so that the true value might be equal to the threshold, but it might also be higher. In the case of censoring below, values that fall below a certain threshold are censored. This study uses a two-limit Tobit model because the dependent variable (vulnerability scores), which are derived from equations (AP-2a) and (AP-2(b), are bounded between 0 and 1 (censored from above and below simultaneously), thus rendering Ordinary Least Squares (OLS) inappropriate.

Following Wooldridge (2002), the two-limit Tobit model is specified as (AP-8) $$V_i^{*}={{\mathrm{ }\mathrm{\beta }}_1\mathrm{X}}_{\mathrm{i}}+{\mathrm{e}}_{\mathrm{i}}\mathrm{ }{\mathrm{e}}_{\mathrm{i}}\left|{ X}_i \text{˜ Normal}\right(0,{\mathrm{\delta }}^2)$$(AP-8) (AP-8a) $$V_i=0\text{if}{V}_i^{\mathrm{*}}\le 0\mathrm{ }$$(AP-8a) (AP-8b) $$V_i=V_i^{\mathrm{*}}\text{if}0<V_i^{\mathrm{*}}<1$$(AP-8b) (AP-8c) $$V_i=1\text{if}{ V}_i^{\mathrm{*}}\ge 1\mathrm{ }$$(AP-8c) where $V_i^{*}$ denotes the latent-value vulnerability index, $V_i$ is a vulnerability index used as a dependent variable, β is a parameter to be estimated, ${\mathrm{\epsilon }}_i$ is an independently distributed error term that is assumed to be normally distributed with zero mean and constant variance, σ². The Tobit model in equation 2.8(2) $$̂{\mathrm{\text{VFI}}}_{\mathrm{i}}=̂\mathrm{\text{Pr}}({\mathrm{\text{lnC}}}_{\mathrm{i}}<\text{TZS}33,748|\mathrm{\text{Xi}})=\text{Ф}(\frac{\mathrm{\text{ln}}(\mathrm{\text{TZS}}\mathrm{ }33,748)‐\mathrm{\text{Xi}}̂\mathrm{\beta }}{\sqrt{\mathrm{X}\mathrm{i}}̂\mathrm{\theta }})$$(2) is estimated using a maximum likelihood estimation procedure, which is a general method for obtaining parameter estimates and performing statistical inferences on the estimates (Amemiya, 1977).

Appendix B.

Complete Tobit regression results

Table B1. Determinants of Vulnerability to Poverty and Food Insecurity among PSSN Program Beneficiaries.

Explanatory Variable	Vulnerability to Poverty				Vulnerability to food insecurity
	AME	SE	t	P>|t|	AME	SE	t	P>|t|
Age of HH HEAD	0.0010	0.0000	117.84	0.000	0.0001	0.0000	25.17	0.000
Sex of HH head	−0.0224	0.0002	−102.24	0.000	−0.0189	0.0003	−54.24	0.000
Marital status	−0.0219	0.0003	−86.29	0.000	−0.0452	0.0003	167.54	0.000
Family size	0.0564	0.0001	922.68	0.000	0.0647	0.0001	492.61	0.000
Dependency ratio	0.0494	0.0001	410.88	0.000	0.0424	0.0001	576.4	0.000
Location	0.1524	0.0004	407.98	0.000	0.1138	0.0002	513.58	0.000
Credit	−0.1310	0.0009	−153.49	0.000	−0.0915	0.0004	225.78	0.000
Business ownership	−0.0032	0.0003	−11.01	0.000	−0.0001	0.0002	−0.27	0.789
Livestock ownership	−0.0812	0.0002	−404.98	0.000	−0.0766	0.0002	499.78	0.000
Occupation	−0.1565	0.0022	−69.61	0.000	−0.1561	0.0011	145.61	0.000
Education subsidy	−0.0207	0.0003	−68.08	0.000	0.0072	0.0002	36.01	0.000
Health Subsidy	−0.0171	0.0003	−55.72	0.000	0.0104	0.0002	48.49	0.000
Food assistance	−0.2126	0.0019	−112.69	0.000	−0.1394	0.0008	177.64	0.000
Number of obs.		844			Number of obs.			844
F(13, 733)		113722			F(13, 733)			84965
Prob > F		0.000			Prob > F			0.000
Number of strata		51			Number of strata			51
Number of PSUs		796			Number of PSUs			796
Design df		745			Design df			745

Table B2. Determinants of VP and VFI - Robustness check with sample restricted to household enrolled to CCTs only.

Explanatory variable	Vulnerability to poverty				Vulnerability to food insecurity
	AME	SE	t	P>|t|	AME	SE	t	P>|t|
Age of HH head	0.0010	0.0000	111.83	0.000	0.0001	0.0000	21.55	0.000
Sex of HH head	−0.0235	0.0002	105.37	0.000	−0.0192	0.0004	−52.13	0.000
Marital status	−0.0223	0.0002	−92.67	0.000	−0.0451	0.0003	161.05	0.000
Family size	0.0568	0.0001	914.71	0.000	0.0647	0.0001	482.61	0.000
Dependency ratio	0.0503	0.0001	406.23	0.000	0.0426	0.0001	548.78	0.000
Location	0.1532	0.0004	409.40	0.000	0.1146	0.0002	521.28	0.000
Credit	−0.1374	0.0008	182.61	0.000	−0.0980	0.0003	302.95	0.000
Business ownership	−0.0027	0.0003	−9.07	0.000	−0.0002	0.0002	−0.95	0.344
Livestock ownership	−0.0822	0.0002	403.45	0.000	−0.0772	0.0002	484.33	0.000
Occupation	−0.2033	0.0022	−92.27	0.000	−0.1746	0.0008	213.55	0.000
Education subsidy	−0.0191	0.0003	−59.77	0.000	0.0077	0.0002	40.23	0.000
Health Subsidy	−0.0167	0.0003	−56.16	0.000	0.0106	0.0002	48.79	0.000
Food assistance	−0.2277	0.0022	105.92	0.000	−0.1313	0.0011	116.17	0.000
Number of obs.		802			Number of obs.			802
F(13, 733)		109226			F(13, 733)			87085
Prob > F		0.000			Prob > F			0.000
Number of strata		51			Number of strata			51
Number of PSUs		796			Number of PSUs			796
Design df		745			Design df			745

Table B3. Determinants of VP and VFI - robustness check with sample restricted to household enrolled to CCTs + PWs.

Explanatory variable	Vulnerability to poverty				Vulnerability to food insecurity
	AME	SE	t	P>|t|	AME	SE	t	P>|t|
Age of HH head	0.0010	0.0000	94.7	0.000	0.0001	0.0000	15.5	0.000
Sex of HH head	−0.0264	0.0006	−42.5	0.000	−0.0261	0.0004	−68.6	0.000
Marital status	−0.0208	0.0006	−37.6	0.000	−0.0466	0.0004	126.7	0.000
Family size	0.0548	0.0002	300.1	0.000	0.0683	0.0001	916.6	0.000
Dependency ratio	0.0464	0.0002	202.8	0.000	0.0408	0.0001	287.5	0.000
Location	0.1420	0.0007	195.1	0.000	0.1067	0.0005	215.3	0.000
Credit	−0.1110	0.0015	−76.2	0.000	−0.0798	0.0009	−88.1	0.000
Business ownership	−0.0059	0.0005	−12.1	0.000	0.0003	0.0003	1.0	0.310
Livestock ownership	−0.0794	0.0003	235.1	0.000	−0.0742	0.0003	271.7	0.000
Occupation	−0.1561	0.0030	−52.7	0.000	−0.1620	0.0017	−94.3	0.000
Education subsidy	−0.0192	0.0005	−36.1	0.000	0.0044	0.0003	13.2	0.000
Health Subsidy	−0.0117	0.0005	−21.4	0.000	0.0202	0.0004	54.6	0.000
Food assistance	−0.1724	0.0005	350.5	0.000	−0.1576	0.0004	445.4	0.000
Number of obs.		251			Number of obs.			251
F(13, 733)		263465			F(13, 733)			387186
Prob > F		0.000			Prob > F			0.000
Number of strata		42			Number of strata			42
Number of PSUs		726			Number of PSUs			726
Design df		684			Design df			684

Note. 9 strata omitted because they contain no subpopulation members.

Appendix C.

Robustness check with t test

Table

	Vulnerability to Poverty				Vulnerability to food insecurity
Variable of interest	Mean				Mean
	(1)	(0)	t	P>|t|	(1)	(0)	t	P>|t|
Sex (1 = male)	0.318 (0.0010)	0.266 (0.0009)	49.4	0.000	0.380 (0.0009)	0.296 (0.0010)	81.2	0.000
Marital (1 = not married)	0.252 (0.0010)	0.325 (0.0008)	−69.9	0.000	0.277 (0.0009)	0.389 (0.0008)	−110.9	0.000
Location (1 = rural)	0.328 (0.0008)	0.186 (0.0008)	126.6	0.000	0.369 (0.0008)	0.258 (0.0012)	76.5	0.000
Credit (1 = received credit)	0.144 (0.0022)	0.300 (0.0008)	−71.6	0.000	0.239 (0.0029)	0.347 (0.0008)	−37.6	0.000
Business Ownership (1 = owning)	0.265 (0.0013)	0.303 (0.0009)	−27.6	0.000	0.328 (0.0014)	0.348 (0.0009)	−13.9	0.000
Livestock (1 = own livestock	0.308 (0.0011)	0.283 (0.0009)	18.3	0.000	0.357 (0.0011)	0.331 (0.0010)	18.5	0.000
Occupation (1 = had occupation)	0.100 (0.0021)	0.298 (0.0008)	−87.5	0.000	0.193 (0.0037)	0.346 (0.0008	−40.5	0.000
Education Subsidy (1 = received)	0.241 (0.0011)	0.299 (0.0008)	−42.6	0.000	0.324 (0.0011)	0.346 (0.0009	15.0	0.000
Health Subsidy (1 = received)	0.227 (0.0015)	0.299 (0.0008)	−40.9	0.000	0.287 (0.0020)	0.347 (0.0008)	−27.6	0.000
Food assistance (1 = received)	0.265 (0.0072)	0.296 (0.0008)	−4.3	0.000	0.376 (0.0091)	0.344 (0.0008)	3.5	0.001
Number of strata:	51				Number of strata: Number of PSUs: Design df Subpop. no. obs:		51
Number of PSUs:	796						796
Design df	745						745
Subpop. no. obs:	844						844