Abstract

Introduction
Public finance through social protection programmes is essential for reducing poverty, especially in developing countries. Social protection policies and programmes are one potential approach that simultaneously tackles poverty and inequality while empowering poor individuals to participate equitably in market activities (Slater, 2011). In Indonesia, several social protection programmes are conducted by both the central government and local governments. Social assistance programmes conducted by the central government include Conditional Cash Transfer, Social Assistance for Elderly People, Wage Subsidised Assistance, Noncash Food Assistance Program, Fuel Cash Transfer, Village Cash Transfer, and Village Labour Intensive Cash Assistance. In addition, there is a wide variety of social protection programmes applied by local governments.
The effectiveness of social assistance should be evaluated not only by its impact on reducing short-term poverty, but also by its capacity to facilitate upward mobility from vulnerability to at least middle-class status. Therefore, it is necessary to quantitatively evaluate the effectiveness of social assistance programmes by assessing their impact the class mobility. It is commonly accepted that Randomised Controlled Trials (RCTs) are the gold standard for explaining the causality of an intervention's outcomes. However, implementing RCTs is time-consuming and costly compared to survey data collection. The main objective of this paper is to predict the effect of social assistance programmes on poverty reduction by applying predictive modelling, including a machine learning approach to cross-sectional data. In addition, we compare the economic status of digitalised agricultural households and non-digitalised agricultural households.

Methodology
This research uses advanced quantitative methods using the 2023 Indonesia National Socio-Economic Survey as a data source. This data contains approximately 345,000 representative sample households selected using a multistage sampling method. The analytical procedures could be determined as the following steps:
1. Filtering agricultural households from the datasets to be used as the study population
2. Determining the economic status of agricultural households in the sample data based on the poverty line. The poverty line is defined using the basic needs approach. There are 5 classes of economic status constructed:
a. Poor households: households with expenditure per capita lower than the poverty line
b. Households at risk of falling into poverty: households with expenditure per capita lying between the poverty line and 1.5 times the poverty line
c. Emerging middle-class household: households with expenditure per capita lying between 1.5 times the poverty line and 3.5 times the poverty line
d. Middle-class households: households with expenditure per capita lying between 3.5 times the poverty line and 17 times the poverty line
e. High-class households: households with expenditure per capita that lies higher than 17 times the poverty line.
3. Finding the best method to predict the class of economic status based on predictors outside expenditure, such as asset ownership, housing condition, education, financial access, digitalisation, food insecurity, and unemployment status for households that do not receive social assistance. Some advanced statistical methods and machine learning approaches are performed and evaluated for their predictive accuracy.
4. Using the determined best method to predict the class of economic status of agricultural households that receive social assistance from the government. This predicted class of economic status could be assumed as the predicted economic status of agricultural households if they do not receive social assistance from the government.
5. Comparing the class of economic status and the predicted class of economic status if they do not receive social assistance from the government to estimate the likelihood of upward economic mobility.
6. Comparing the class of economic status between digitalised agricultural households and non-digitalised agricultural households

Preliminary Results
The distribution of agricultural households across economic status depicts that 11.8% are classified as poor, 26.56% are at risk of falling into poverty, and 47.06% belong to the emerging middle class. Moreover, the remaining 14.29% of households are categorised as middle class, and the high-class households are about 0.29%. Considering that the three lowest groups belong to the vulnerable population, the main analysis only focuses on these groups. A random forest classification model is applied due to its high predictive accuracy in classifying households based on economic status. In general, social assistance could prevent the vulnerable population downward its economic status.
Empirical results indicate that, in the absence of social assistance, 3.36% of households at the group at risk of falling into poverty would descend into the poor class. Similarly, if social assistance were withdrawn from emerging middle-class households, 7.40% would fall into the at-risk group, and an additional 1.90% would fall into poverty. While these findings suggest that social assistance has a mitigating effect on downward economic mobility among vulnerable populations, the magnitude of this effect is relatively low. This might be caused by a major issue in inaccurately delivering social assistance to the right targeted households. Furthermore, there are significant differences in economic status between digitalised agricultural households and non-digitalised agricultural households.