This study utilizes secondary data obtained from the Central Bureau of Statistics, specifically employing pooled cross-sectional data from SUSENAS 2017 to 2021. The analysis focuses on households located in Java, the most populous island in Indonesia. According to the 2020 Census, Java is home to 56.01% of Indonesia's total population of 272 million people. Java was chosen due to its status as a center of development, with adequate non-cash access and facilities. Additionally, many of the pilot areas during the initial implementation of BPNT were in Java. However, it is important to note that Java still grapples with issues of inequality, contributing to rising levels of poverty, food insecurity, and obesity among the lower segments of the economy (Elmes, 2018)(Setiawan et al., 2020). (Rahayu et al., 2019) also found inequality in food security status across provinces in Java.

The data utilized in this research were collected from the SUSENAS modules, which included both household and individual data. The individual module provides sociodemographic

information, which will be used to predict household eligibility for receiving food assistance programs based on propensity scores. The household consumption module includes consumption expenditure details, which will be used to measure the Household Food Expenditure Share (FES) and Household Dietary Diversity Score (HDDS).

The researchers analyzed household data from Java, encompassing 96,174 households in 2017, 95,885 households in 2018, 96,751 households in 2019, 102,451 households in 2020, and 105,110 households in 2021. Using the consumption module, the researchers categorized these households into twelve food groups, as detailed in Table 1. Processed foods were included within the various food groups, whereas alcoholic beverages were excluded from the estimation of HDDS.

Although this study is based exclusively on household data from Java, it acknowledges the potential for regional variations in food prices over the years. Given that the SUSENAS dataset does not provide specific food prices, the researchers calculated the food price per unit by dividing weekly household food expenditure by weekly household food consumption. The researchers then aggregated these unit prices to determine the total food expenditure for each district on an annual basis. To address regional price disparities, the researchers applied the Laspeyres price index methodology, as used by (Kronebusch & Damon, 2019), which employs a reference state to control for price differences. In this research, the researchers used the Laspeyres price index formula to construct the food price index at the district level, dividing the total food expenditure in each district annually by the total food expenditure in each district for the base year of 2017.

Social safety net programs in developing countries often use proxy means-testing, as many potential beneficiaries are employed in the informal sector and lack verifiable income recordse (Alatas et al., 2012). Proxy means testing typically relies on extensive, periodic quasi-censuses of the population. Governments use household assets or per-capita consumption to estimate incomes. Consequently, in this research, eligibility for benefits is determined by predicted income, which is approximated through the analysis of 76 covariates. These covariates include the age, gender, marital status, education, and occupation of the household head; household size, dependency ratio, and educational attainment within the household; housing status and characteristics, drinking water source, toilet facility, sewage facility, and lighting source; asset ownership; and other social protection programs affecting household food insecurity (Ardianti & Hartono, 2022)(Hanna & Olken, 2018).

To evaluate the impact of the Sembako Program on household food expenditure, the researchers employed a quasi-experimental methodology combining propensity score matching (PSM) and ordinary least squares (OLS) regression. PSM was used as a pre-estimation step to create counterfactuals based on 76 observed characteristics, thus mitigating selection bias arising from non-randomization (Austin, 2011).

Given the availability of repeated cross-sectional data, the researchers conducted PSM separately for each year from 2018 to 2021 to ensure contemporaneous control. For the pre-treatment period in 2017, the researchers designated the lowest 10% of per-capita expenditure as a control group. TNP2K defines a proxy means-test (PMT) score cut-off of 30% as strongly associated with eligibility for food assistance programs. Thus, a lower cutoff indicates a higher likelihood of household eligibility for receiving food assistance.

The researchers used kernel-based PSM, which has been shown to outperform one-toone propensity score matching, regardless of the number of potential controls available (Berg, 2011). The assumption underlying kernel matching and caliper matching are slightly different. With kernel matching, the more "similar" the untreated observations are to the treated observations, the more weight they are given. With caliper matching, all untreated observations within the specified radius of the treated observation are used, and they all receive the same weight (Caliendo & Kopeinig, 2008). Kernel matching is more effective in reducing the mean standardized difference while preserving almost all observations from the original sample, compared to caliper matching. Kernel weighting also achieved the best covariate-specific balance between the treatment and comparison groups. Kernel matching assigns greater weight to better matches, thus maximizing precision while retaining sample size without exacerbating bias (Garrido et al., 2014). The pre-estimation model used in this research is specified in equation ( 1) and ( 2). The main estimation using OLS with multi-arm approach is specified in equation (3):

h = W(Pr(X_h ))= (Pr( h = 1| _h )) _h = 1 (1)

βpsm = E(Y_h|FV_h=1, Sh=1)− E(W_h(Y_h|FV_h=0, S_h=1)) (2)

Yhrt = 0 + 1 hrt + 1 hrt + + + + + , (W(Pr(Xh )) x fwt) (3)

Where Wh is weight for household h based on propensity score Pr(Xh). Xh is vector covariates for household h in each year, to predict the probability of household h on receiving food voucher (FVh). Sh = 1 represents the on-support range for household h in the treatment and control groups that will be matched based on Wh. PSM will automatically estimate treatment effect βpsm based on differences in expected outcome E(Yh) of household h in the treatment group (FVh = 1) and control group (Wh (FVh = 0)) in terms of matched samples in the on-support range. Yhrt is dependent variables (FES & HDDS) of household h in region r at time t. β1 is the marginal change compared to the control group, of being in the treatment. Our treatment variables BPNThrt and Sembakohrt have binary values of 1 if the household h in region r at time t receives BPNT or the Sembako Program, and 0 otherwise. αZrt is a vector of regions and time control variables, including the food price index and urban-rural area. is year fixed effects, is region fixed effects, is interaction of year fixed effects and region fixed effects, is error term, and Fwt is sampling weight from the SUSENAS survey.

However, PSM cannot eliminate bias arising from unobserved variables. The researchers aggregated data from the lowest 10% of socioeconomic households in 2017 and post-PSM household data for each subsequent year (2018, 2019, 2020, and 2021). Using a pooled cross- sectional dataset, the researchers conducted OLS estimation in equation (3), incorporating the food price index at the district level, year-fixed effects, region-fixed effects, district-fixed effects, province-fixed effects, and the interaction of province and year-fixed effects to address unobserved confounders. Including fixed effects as control variables in the estimation helps mitigate the impact of both time-variant and time-invariant effects ((Kis-Katos & Sparrow, 2015); (Oster, 2019)).

Through the pre-estimation propensity score matching (PSM), the researchers established treatment and control groups with comparable characteristics across the 76 covariates used. The researchers interacted the kernel weight with the SUSENAS sampling weight to estimate the population level. By aggregating data from the lowest 10% of socioeconomic households in 2017 and household data post-PSM for each year from 2018 to 2021, the estimations are restricted to beneficiaries and their counterfactuals.

Table 2 and Table 3 shows that household data from 2017 was used as the control group from the pre-treatment period, therefore no PSM was applied to the 2017 data. After performing PSM for each year in the post-treatment period (2018-2021), both matched and unmatched samples data were obtained. For the main estimation, only the matched sample data, combined with the control group data from the pre-treatment period, will be used. The distribution of pooled data by province is shown in Table 4.

Figure 1illustrates that, following PSM, the bias across covariates was reduced, thereby improving the accuracy of the impact estimation. As a result, differences in outcomes between the control and treatment groups were primarily attributable to treatment variables and unobserved factors. After conducting annual PSM for the post-treatment period, the researchers acquired a contemporaneous control group. This contemporaneous control group was combined with the pre-treatment control group to form the common control group. Consequently, the pooled dataset now includes three treatment groups: the control group, BPNT beneficiaries, and Sembako beneficiaries.

Table 4 displays the distribution of the sample used in the primary estimations. The control group, predominantly composed of non-beneficiaries, represents the largest proportion, followed by the two treatment groups: BPNT beneficiaries and Sembako Program beneficiaries. The larger proportion of BPNT beneficiaries compared to Sembako Program beneficiaries is due to the longer duration of BPNT implementation during the observation period. The Sembako Program was only introduced in 2020. Among all provinces in Java, Yogyakarta exhibits the highest proportion of beneficiaries in both the BPNT and Sembako Program, followed by Central Java and East Java.

Using a multi-arm approach, the researchers employ a shared control group to estimate the effects of both the BPNT and Sembako Program. Since the Sembako Program is an enhancement and continuation of BPNT, the methodology aims to discern the individual impacts of each program by utilizing a unified control group. This approach allows us to compare the effects of both programs on beneficiaries, particularly in evaluating the efficacy of improvements in the food voucher program. By including control variables, the food price index, and all fixed effects, the researchers address potential upward biases in assessing reductions in Food Expenditure Share (FES) and Household Dietary Diversity Score (HDDS).

presents the estimation results obtained through this multi-arm approach. By comparing the two treatment groups with the common control group, it is evident that the Sembako Program has a more pronounced impact than BPNT on reducing beneficiaries’ FES. The coefficient estimates for both programs exhibit the same level of significance. Specifically, the Sembako Program leads to a 2.50% reduction in beneficiaries’ FES, whereas BPNT shows a 1.96% reduction. For the HDDS parameter, both programs show the same level of significance at α = 1%. The Sembako Program results in a 23.01% increase in HDDS, surpassing the 14.83% increase observed with BPNT.

To ensure robustness, the researchers assess program impact using various cutoffs of the pre-treatment control group, representing the lowest 10%, 20%, and 30% socioeconomic levels, as depicted in Table 6. Increasing the cutoff level results in a slight decrease in the household FES reduction for BPNT from -1.96% to -1.95%, while the Sembako Program’s reduction remains consistent at -2.50%. For the HDDS indicator, raising the cutoff level leads to a rise in HDDS from 14.83% to 14.98% for BPNT, and a slight decrease in the increase in HDDS from 23.01% to 22.91% for the Sembako Program. Nevertheless, minor discrepancies in the coefficient estimates for both BPNT and the Sembako Program are observed, with no variation in significance level, indicating the consistency of the estimation model.

Household Food Expenditure Share (FES) and Household Dietary Diversity (HDDS) are a vital indicator of household welfare and food security, reflecting the extent to which households can satisfy their dietary needs (Yuliana et al., 2020)). Engel's theory further elucidates this phenomenon by suggesting that household welfare correlates inversely with the proportion of income allocated to food expenditures. As family income increases, the percentage of income spent on food decreases. Consequently, Engel's theory posits that variations in expenditure patterns can serve as a proxy for assessing changes in population welfare levels, with modifications in expenditure composition acting as an indicator of shifts in welfare status (Wirba, 2023). HDDS as a for household diet quality and indicator of access to the food component of food security (Ali et al., 2024)(Mehraban & Ickowitz, 2021).

During the pre-estimation stage, we performed Propensity Score Matching (PSM) for each post-treatment year, using 76 household characteristics as covariates to estimate the likelihood of households receiving treatment. For the pre-treatment household data, we selected households from the lowest 10% of per capita expenditure in 2017 to serve as the control group for the pre-treatment period. This approach enabled us to construct pooled data from a matched sample for each post-treatment year and the pre-treatment control group, as shown in Table 3. The distribution of samples by treatment group and province from pooled dataset, is shown in Table 4. We have two types of food voucher treatments, with a common control group. These will be estimated using OLS with a multi-arm approach to compare the FES and HDDS outcomes between the two food voucher treatments.

The main estimation results for FES and HDDS in Java, shown in Table 5., indicate that the impact of a decrease in the percentage of FES and an increase in HDDS is greater for the Sembako Program beneficiaries group compared to the BPNT beneficiaries group, relative to the common control group. The results of this study fill a gap in the existing literature, as there are few studies that compare the impact of two types of food voucher flexibility. Previous studies have discussed and compared between cash, in-kind, and non-cash transfers (Banerjee et al., 2023) (Doocy et al., 2020) (Hidrobo et al., 2014) (Michelson et al., 2012). In principle, non-cash assistance offers more flexibility than cash and in-kind assistance, thereby providing a greater welfare impact. In line with this principle, the findings of this study support previous research, showing that food voucher with higher flexibility lead to better welfare outcomes.

A more detailed analysis of the impact of food assistance at the provincial level, as shown in Table 7 and Table 8. , reveals that the BPNT and Sembako Program have no effect on FES but do lead to an increase in HDDS in Jakarta Province. In contrast, in West Java and Banten provinces, the BPNT and Sembako Program reduce FES but do not significantly affect HDDS. Meanwhile, in Central Java, Yogyakarta, and East Java provinces, these programs both reduce FES and increase HDDS. The varying impact of food assistance across provinces can be influenced by several factors, including the initial conditions of beneficiaries, as well as the characteristics and availability of foodstuffs in each region (Sinaga et al., 2021).

Jakarta has the lowest poverty rate in Indonesia (B.P.S., 2019). Despite having the highest monthly food expenditure, the household FES remains below the national average. According to Engel's theory, this indicates that food consumption in Jakarta has reached a saturation point, where any increases in income are allocated to non-food expenses or savings. (Hoynes & Schanzenbach, 2009) argue that households spending less on food relative to non-food items may face constraints when utilizing food voucher. Although this may not significantly impact the household FES, as food voucher reduce out-of-pocket food expenses and shift spending towards non-food items, it results in an overall increase in both food and non-food consumption. In the other five provinces, both the BPNT and Sembako Program led to significant decreases in FES. Despite being a major food production hub with substantial food resources, studies conducted by BKP in 2017 indicate that East Java, West Java, and Banten are still classified as food-insecure regions (Prasada & Masyhuri, 2019). The Sembako Program, however, resulted in a greater reduction in FES compared to the BPNT program across all provinces. Since the Sembako Program provides higher food voucher amounts than BPNT, beneficiaries receive more additional income, leading to a more substantial decrease in FES (Athoillah et al., 2022)(Sugiharti, 2020). Based on the FES reduction and its statistical significance, it can be concluded that the largest beneficiaries of both food voucher programs are in Central Java and East Java.

The BPNT and Sembako Program have no significant impact on HDDS in West Java and Banten provinces. However, both programs have a significant effect in the other four provinces. Generally, the increase in HDDS resulting from the Sembako Program was greater than that of BPNT, except in Jakarta province. In Jakarta, where agricultural land is scarce and population density is high, the demand for rice exceeds local production capacity (Alfa & Subagyo, 2018). Consequently, beneficiaries may prioritize rice purchases using food voucher. The situation in Jakarta, West Java, and Banten provinces highlights that food consumption in Indonesia remains low and less diverse, with a preference for local food and a reliance on rice as a staple (H. P. S. (Rachman & Ariani, 2008). As a result, food assistance does not alter beneficiaries' preferences, nor does it encourage them to allocate their additional income toward purchasing a more diverse range of foodstuffs. Households tend to substitute commodities to buy staple foods when their income is insufficient (Yuniarti et al., 2022). Both programs show significant increases in HDDS in the provinces of Central Java, Yogyakarta and East Java. Based on the increase in HDDS and its statistical significance, it can be concluded that the largest beneficiaries of both food voucher programs are in Central Java and East Java.

Looking at the overall impact of the two programs in urban and rural areas shows that both programs have a significant impact on reducing FES and increasing HDDS in urban and rural areas. Rural areas experienced a greater decline in FES than urban areas, but a lower increase in HDDS than urban areas. Figure 2, where at baseline, rural areas have higher FES than urban areas. FES tends to be higher for households in the lower income quintiles, which are the middle- and low-income groups, with rates often exceeding 50%, indicating food insecurity (B.P.S., 2017) (Sinaga et al., 2022). Therefore, additional income has a greater impact on reducing FES in rural areas.Figure 2 shows that the mean HDDS in urban areas is slightly lower than in rural areas at the initial level. Despite the common perception that urban areas have greater food supply, better infrastructure, and improved access to food diversity, HDDS is also influenced by sociodemographic and cultural factors (Hirvonen, 2016) (Kolliesuah et al., 2023). (Weerasekara et al., 2020) discovered that dietary diversity was richer in rural areas compared to urban areas, possibly due to the prevalence of agricultural households in rural settings that sell their produce in urban markets. However, the relationship between dietary diversity and market access versus 'own production' remains a subject of debate. Notably, rural areas in Indonesia are experiencing a decline in dietary diversity over time, with households consuming fewer fruits, vegetables, and legumes while increasing their consumption of dairy, eggs, and meat as their incomes rise. This trend indicates a complex interplay of nutritional gains and losses (Mehraban & Ickowitz, 2021).

Table 9 shows that BPNT results in a more significant decrease in FES than the Sembako Program in rural areas. This could be attributed to the dominance of agricultural households in rural areas, particularly in Java, which is the largest rice producer in Indonesia. In such regions, access to rice is relatively easier, even without additional income from food assistance programs. Additionally, rice consumption tends to be higher among agricultural households (B.P.S., 2017). By limiting BPNT benefits solely to rice and eggs, beneficiaries can allocate more vouchers towards purchasing eggs, thereby allowing them to reallocate their cash towards acquiring other food items. Research indicates that egg consumption is particularly high among animal protein sources in rural areas of Indonesia. Conversely, other animal protein sources such as beef, chicken, fish, and milk are often perceived as luxury items in rural settings (Khoiriyah et al., 2019) (Suryana et al., 2021). The flexibility offered by the Sembako Program allows beneficiaries to choose other animal protein sources, which may be more costly than eggs, or to purchase additional fruits and vegetables using food voucher. However, this flexibility may still fall short of meeting all dietary requirements, compelling beneficiaries to continue using their cash to acquire additional food items. Consequently, the reduction in FES among Sembako Program beneficiaries is relatively smaller compared to BPNT beneficiaries.