The (a)symmetric effects of income and unemployment on popular demand for redistribution

Abstract

Numerous studies show that those with lower income and the unemployed support more redistribution, which is attributed to material self-interest. However, recent studies assessing within-individual changes result in smaller and less consistent effect estimates. To explain why preferences do not narrowly follow material self-interest, this study argues that the effects of income and unemployment may be asymmetric, implying that improving and deteriorating material circumstances exert differently sized effects. The claims are tested using panel data from Great Britain and a weighted difference-in-difference estimator. The results show that only income increases (negatively) affect redistribution support while income decreases have null effects. In contrast, unemployment is estimated to have a strong and symmetrical effect in line with self-interest theory. These results add further evidence for the validity of self-interest theories but suggest that individuals are only boundedly rational.

Supplemental data for this article can be accessed online at: https://doi.org/10.1080/01402382.2021.1963139 .

Keywords:

• Inequality
• public opinion
• political economy
• self-interest
• rational

Acknowledgements

This article benefitted from the help of many colleagues, to whom I express my gratitude. Firstly, I thank Frank Bandau, Fabio Bothner, Anselm Hager, Lukas Hakelberg, Nikolaus Jopke, Simon Linder, Thomas Rixen, and Nadja Wehl. Secondly, helpful comments were provided by participants of the ECPR General Conference 2020, where I presented a previous draft of this article. Finally, I thank two anonymous reviewers and the editors of West European Politics for thoughtful feedback and support.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1 Of course, other pitfalls of effect identification and estimation must also be avoided (e.g. post-treatment controls).

2 This blow may be softened by the market income generated by other household members.

3 More than two periods can be used in DID, but canonical DID requires that included periods can be strictly categorized into pre- and post-treatment periods, which closely resembles the two-period data setup introduced here.

4 This is achieved by assigning all respondents in the treatment group a weight of 1 and respondents in the control group a weight of varying size.

5 Note that other models can be used for analyses of asymmetric effects, most notably models with first-differenced variables in which increases and decreases of explanatory variables are included separately (Allison 2019; Haffert and Ergen 2019). This approach can also incorporate continuous explanatory variables, but it is less straightforward to address unobserved heterogeneity.

6 TWFE requires the assumption of linear-additive effects (Imai and Kim 2021) with treatment effects that are constant across individuals and periods (de Chaisemartin and D’Haultfoeuille 2020).

7 These include the dependent variable: the demand for redistribution; the explanatory variables: income and employment; as well as matching and control variables: age, gender, and education. Other included variables (vote intention, perceived unemployment risk) are not included here because they are not required in the matching procedure for the analysis of unemployment.

8 The main reason for this inconsistent data coverage is that some concepts are simply not measured in a particular wave. Relatedly, some concepts are measured only periodically in the sense that, e.g., over five panel waves respondents only indicate information about their working status the first time they participate in the panel over these five waves.

9 The proportions of respondents left after listwise deletion are: 11% (wave 10), 5% (wave 11) and 2.5% (wave 12). These low proportions occur because some concepts are only measured the first time a respondent participates in the panel between, e.g., waves 6 to 12. The unfortunate implication of this design is that even when respondents participate in the same wave, it is not guaranteed that they all respond to the same items.

10 Respondents in the highest category are assigned an income of £175,000, which keeps the spacing constant compared to the preceding category.

11 The spouse of a rich individual, for example, benefits from their partner’s income and should adjust their preference accordingly. Furthermore, income is often taxed at the household rather than the personal level, but this is not the case in the United Kingdom.

12 It must be noted that respondents with ‘constant income’ may experience a degree of income change. For example, a respondent within income category ‘£20,000–£24,999’ in both waves may experience income variation within this range.

13 The cross-sectional models include gender, age, and age squared in addition to the controls outlined in the data and model setup section (education, employment, and unemployment risk).

14 The median income changes of the different treatment groups are: +44% (20% increase or more treatment), −40% (20% decrease or more treatment), +67% (40% increase or more treatment) and −54% (40% decrease or more treatment).

15 Equivalence income scales aim to facilitate the comparison of consumption potential between households with different composition because households with more members benefit less from a given income amount than households with less members. Equivalization is applied by dividing household income by the square root of household members.

16 Household size does not inherently affect the relationship between market income and expected tax burden/transfer income. When the household size increases, e.g. due to a newborn child or older relative becoming a household member, using an equivalence scale would downscale household income. It is unreasonable to expect that, as a result, rational household members will increase their demand for redistribution. Their income decreases simply due to the equivalization; the amount of taxes the household pays does not necessarily decrease due to the ‘lower’ income, and the amount of transfer income it receives does not necessarily increase.

17 This is not a major surprise because unequivalised and equivalised household income are highly correlated (r = .93).

18 A difference in treatment effects may arise, for example because of compositional differences between those with stable and changing circumstances. However, the evidence presented in Figures A1–A6 in the online appendix suggests that, while treated and untreated respondents do differ in their socio-economic background, the differences are not substantial. There is thus no strong evidence for compositional differences.

Additional information

Notes on contributors

Leo Ahrens

Leo Ahrens is a doctoral candidate at the Free University of Berlin. His research focuses on the political economy of redistribution and taxation. His work has appeared in Socio-Economic Review, New Political Economy, and Regulation & Governance, among other journals. [[email protected]]