Figures and data in Refining the simulation of social assistance with monthly income data and calibration

Abstract

As a last-resort benefit, social assistance can cushion the impact of cuts in other benefits on low-income households. Therefore, it may play a crucial role for the effects of benefit cuts on income distribution. Measurement errors in its simulation can significantly affect the estimated distributional indicators. At the same time, simulating social assistance is challenging because it is influenced by many factors, not all of which are typically observed in the data or captured with an existing simulation model. In this study, we refine the simulation of social assistance in the Finnish static tax-benefit model SISU by considering monthly fluctuations in household income. Moreover, we use regression-based calibration technique to account for oversimulation caused by, among other things, non-take-up and unobserved household wealth. Using the refined simulation model, we estimate that social assistance compensates the substantial cuts to housing and unemployment benefits in Finland 2024-2025 much less than without refinements. Based on the analysis, the reforms would increase the number of recipient households of social assistance by approximately 26,000 households – a substantial downward revision from the pre-refinement estimate of 50,000 households. At the same time, the benefit cuts are estimated to increase at-risk-of-poverty rates and income inequality more than without the refinements. The increase of the child at-risk-of-poverty rate are 60% larger than before refinements. Although demonstrated in a Finnish context, the results underline the importance of accounting both monthly income variation and oversimulation when simulating means-tested benefits also in other contexts.

This article contains 3 figures and 5 tables.

Figures

**Figure 1.** Over- and undercoverage of social assistance recipients using different simulation methods in 2022, and root mean square error compared to actual data.

**Figure 2.** (**a-b**) Distribution of social assistance recipient households by income decile and number of benefit months according to data, simulation with official model, and simulation with refined model.

**Figure 3.** The impact of the 2024-2025 reforms on total disposable income by income decile, simulated with the official and the refined model.

Tables

Table 1. Sources of error in simulating social assistance by the direction of the bias.

Undersimulation	Oversimulation	Indefinite
Use of annual income	Non-take-up	Inaccuracies in simulating first-tier benefits
Incomplete data on reimbursed costs and payments (e.g. OOP payments)	Overcoverage of population registers (e.g. under-reported emigration)	Discretionary decisions are not modelled
Inaccurate consideration of the family unit definition in SA (e.g. adult children living in the same household)	Incomplete asset data (esp. savings)	Changes in household composition or housing costs during the year are not observed
Priority treatment of garnished income not considered	Tax refunds not included	Survey-related biases such as sampling error, errors in survey design and recall bias
Taxes simulated according to final taxation, not withholding taxation	Sanctions not modelled
No data on gifts and informal support

Table 2. Methods for monthly allocation by income item.

Income item	Allocation to monthly level
Unemployment benefits	Observed at monthly level
Part-time labour income	Observed at monthly level
Study grants and loans	Estimated based on additional information
Sickness and parental benefits	Estimated based on additional information
Full-time labour income	Estimated based on additional information
Pensions	Estimated based on additional information
Child benefits	Estimated based on additional information
Capital income	Annual average (no allocation)
Taxes	Annual tax rate (applied to monthly income)
General housing allowance	Allocated through simulation
Social assistance	Allocated through simulation

Table 3. Differences between results from the refined model and the official model compared to data reference values in 2022.

Data			Difference with data		Difference, calibrated vs. official model
Data			Official model	Calibrated model	Difference, calibrated vs. official model
Social assistance	Expenditure, M€	670	+38 %	0 %	-28 %
	Households	231 000	-3 %	0 %	+2 %
	Indiviudals	392 000	-14 %	-6 %	+9 %
General housing allowance	Expenditure, M€	1 576	-4 %	+3 %	+7 %
General housing allowance	Households	504 000	-15 %	+2 %	+21 %
Gini		29.2	-0.2	-0.1	+0.1
Median income, eqv. €/yr		27 614	-202	-179	+23
AROP rate (60), %		13.6	+0.4	+0.3	-0.1
AROP rate (50), %		6.8	-0.2	0.0	+0.2
AROP gap (60), %		16.7	-1.3	-0.6	+0.7
AROP gap (50), %		21.2	-7.2	-5.5	+1.7
AROP rate (60), age less than 18, %		11.9	+0.8	+0.2	-0.6
AROP rate (60), age 65+, %		14.0	-0.2	-0.1	+0.1
Income share of the 1st decile, %		3.5	+0.4	+0.2	-0.2

Table 4. The estimated effects of the reform on benefit expenditure and distributional indicators with the official and refined model.

Official model			Refined model	Sensitivity test: Change of overcoverage rate
Official model			Refined model	+20%	-20%
Social assistance	Expenditure, M€	+313	+182	+122	+242
	Households (1000)	+50	+26	+8	+43
	Individuals (1000)	+81	+37	+12	+62
Housing allowance	Expenditure, M€	-885	-972	-972	-972
Housing allowance	Households (1000)	-218	-251	-251	-251
Gini index		+0,4	+0,6	+0,6	+0,6
Median income, €/yr		+292	+270	+269	+271
AROP rate, %	Total pop (60)	+1,7	+1,9	+2,0	+1,9
	Total pop (50)	+1,3	+1,8	+1,9	+1,8
	Age less than 18 (60)	+1,7	+2,7	+2,7	+2,6
	Age 65+ (60)	+0,4	+0,4	+0,4	+0,4
AROP gap, %	Total pop (60)	+1,2	+2,6	+2,7	+2,5
Income share of the 1st decile, %		-0,2	-0,3	-0,3	-0,3

Note: In the sensitivity analysis the overcoverage rate after the reform is assumed to be 20% higher or lower than in the reference-year calibration. The increased overcoverage reflects a scenario where oversimulation rises, for example due to unobserved wealth among new recipients, while the reduced overcoverage can reflect higher take-up if increased benefit amounts spur higher claiming probabilities.

Table A1. Coefficients of the logistic regression model analysing overcoverage (1/0) among the simulated social assistance (SA) recipients.

	β	p
Intercept	4.60	0.00
Simulated months of SA (1–12)	-0.12	0.00
SA as % of hh income, avg per month	-0.02	0.00
Socio-economic position (ref=unknown)
Upper-level office workers	0.81	0.00
Lower-level office workers	0.23	0.00
Employees	-0.44	0.00
Students	0.44	0.25
Pensioners	-1.01	0.00
Other	-1.17	0.00
Household life stage (ref=single parents)
Single, less than 35 years	-0.14	0.00
Single, 35–64 years	0.45	0.00
Single, 65+ years	-0.20	0.00
Couple, less than 35 years	0.25	0.00
Couple 35–64 years	0.19	0.00
Couple, 65+ years	0.10	0.56
Family with 2 parents and children	-0.03	0.00
Other	-0.24	0.00
Disposable income (ref=more than 7500 €/year)
Less than 1500 €	3.49	0.00
1500–4500 €	0.24	0.00
4500–7500 €	-1.00	0.00
Income decile (ref=I)
II	-0.46	0.00
III	-0.51	0.00
IV	-0.33	0.00
V-X	0.67	0.00
Migrated during the data year (0/1)	0.48	0.00
Institutionalized (0/1)	2.11	0.00
Debt for execution (0/1)	-0.69	0.00
Immigrant (0/1)	-0.50	0.00
Rental income (0/1)	2.38	0.00
Extra adults in the household (0/1)	-0.49	0.00