Wissenschaftliche Arbeiten › MSc Dissertation
MSc Dissertation | LSE — Behavioural Science | 2024 | Supervised by Dr. Christian Krekel | Grade: 65%
Standard policy analysis aggregates total welfare — it treats a unit of wellbeing gained by the richest and the poorest as morally equivalent. This dissertation challenges that assumption. Drawing on the Prioritarianism in Practice project initiated by Matthew Adler and colleagues, it constructs the first Prioritarian Social Welfare Function (SWF) using life satisfaction (WELLBYs) as the welfare measure.
The empirical application evaluates Victoria's second COVID-19 lockdown (July–October 2020) — one of the most stringent lockdowns outside China. By applying Difference-in-Differences to Australia's major household panel survey (HILDA), the study isolates the lockdown's causal effect on life satisfaction from the general pandemic context. This is, to my knowledge, the first study to causally identify the impact of a COVID-19 lockdown on life satisfaction.
The results are then fed into an Atkinson Social Welfare Function (with inequality-aversion parameter γ = 1) to formally compare the wellbeing consequences of the lockdown policy with a Business-as-Usual counterfactual. The comparison provides an evidence-based recommendation framed in terms that capture fairness to those worst-off — something a simple cost-benefit analysis cannot do.
Policy evaluation has long been dominated by Utilitarianism, which maximises the sum of individual welfare. While tractable, this approach violates the Pigou-Dalton Axiom: it is morally indifferent to who receives a welfare gain, ignoring inequality. Behavioral economics provides substantial evidence that citizens are inequality-averse (Clark & D'Ambrosio, 2015; Hurley et al., 2020), justifying a policy framework that weights improvements for the worse-off more heavily.
Two alternatives satisfy the Pigou-Dalton Axiom: Egalitarianism (comparative fairness, relative wellbeing) and Prioritarianism (absolute wellbeing, independent of relative position). Prioritarianism is preferred for two reasons:
The Atkinson SWF implements Prioritarianism formally through a strictly concave and increasing transformation of individual wellbeing, giving greater weight to improvements for those with lower wellbeing:
Setting γ = 1 reduces the SWF to the sum of log-transformed life satisfaction scores. The logarithm is strictly concave: the same unit increase in life satisfaction contributes more to social welfare for a person at 3/10 than for a person at 8/10. This formally encodes the Prioritarian principle that improvements for the worse-off matter more. γ = 1 represents a moderate inequality-aversion — a balance between efficiency and equality.
A WELLBY (Well-Being Year) equals one point of life satisfaction for one person for one year, measured on a 0–10 Likert scale. WELLBYs are used as the welfare measure because: (1) asking citizens directly about their life satisfaction aligns with democratic principles, respecting individual judgements; (2) life satisfaction predicts key behaviours like voting; and (3) the Atkinson SWF requires ratio-rescaling invariance, which WELLBYs on a 0–10 scale satisfy, unlike positive-affine invariance required for Utilitarian SWFs.
The analysis uses 20 annual waves (2001–2021) of the Household, Income and Labour Dynamics in Australia (HILDA) survey — Australia's major household panel survey, initiated in 2001. Initial sample: 7,682 households; 2,153 additional households added in 2011 (top-up sample); re-interview rates rose from 87% (wave 2) to over 95% (wave 8+).
The treatment window is the 20th wave (2020), when over 90% of interviews were completed before end of October 2020. The sample was restricted to respondents who provided data in 2020 and who were exposed to the lockdown (Stage 3 mobility restrictions in their postcode):
The DiD design compares the change in life satisfaction over time in the treatment group (Victoria, August–November 2020) with changes in the control group (all other Australian states, same period). This isolates the lockdown's specific effect from the general pandemic effect.
The DiD design requires that, absent the lockdown, life satisfaction in Victoria would have followed the same trajectory as in other Australian states. A graphical pre-trends analysis (Figure 4) plots mean life satisfaction for treatment and control groups from 2001 to 2019.
Treatment group (Victoria, red) and control group (other states, blue) track closely from 2001–2019, supporting the parallel trends assumption. In 2020 the treatment group drops sharply to its lowest recorded level — consistent with the lockdown effect.
Three models are estimated: Model 1 (overall effect), Model 2 (excluding COVID-diagnosed individuals to further isolate the lockdown policy effect from the direct health impact of infection), and Model 3 (stratified analysis comparing worst-off to better-off).
| Variable | Model 1 Overall Effect |
Model 2 Excl. COVID-19 |
Model 3 Stratified |
|---|---|---|---|
| Treatment (β1) Causal lockdown effect |
−0.0303*** (0.0077) [−0.0486, −0.0120] |
−0.0297*** (0.0080) [−0.0487, −0.0106] |
−0.0260** (0.0103) [−0.0505, −0.0013] |
| Group (life sat. < 4) Worst-off indicator |
N/A | N/A |
−4.1795*** (0.0442) [−4.2841, −4.0749] |
| Treatment × Group Additional effect on worst-off |
N/A | N/A |
−0.3343*** (0.0175) [−0.3756, −0.2929] |
| Constant | 7.9022*** [7.9019, 7.9025] |
7.9020*** [7.9016, 7.9023] |
7.9528*** [7.9513, 7.9544] |
| Year Fixed Effects | Yes | Yes | Yes |
| State Fixed Effects | Yes | Yes | Yes |
| Individual Fixed Effects | Yes | Yes | Yes |
| Observations | 215,733 | 215,555 | 215,733 |
| State clusters | 8 | 8 | 8 |
| All models use SCQ HILDA population weights. Standard errors clustered at state level (adjusted for heteroskedasticity). Group = 1 for life satisfaction < 4 (threshold from ONS definition of "misery", Oman, 2016). *** p < 0.01, ** p < 0.05, * p < 0.10. | |||
Model 1 establishes a highly significant average treatment effect: the lockdown reduced life satisfaction by 0.0303 points (p = 0.006). Model 2 virtually replicates this result after excluding COVID-positive individuals (−0.0297, p = 0.008), ruling out direct health impacts as a confounder. Model 3 reveals the distributional story: the interaction term (Treatment × Group) is −0.334*** — those with initial life satisfaction below 4 experienced an additional 0.334-point decline, on top of the average effect of −0.026. The lockdown was strongly regressive.
Stratified DiD analyses (in the appendix) reveal which groups were most harmed by the lockdown:
These findings have direct implications for future intervention design: targeted mental health support for those with low baseline wellbeing and poor mental health would be warranted if lockdowns are ever repeated. The gender differential also warrants specific policy consideration.
While DiD controls for time-invariant individual characteristics and common time trends, it does not account for time-varying observable differences between treatment and control groups. Propensity Score Matching (PSM) addresses this by creating a matched sample that is balanced on observed covariates likely to influence both lockdown exposure and life satisfaction.
Covariates used for matching: Big Five personality traits (openness, conscientiousness, extraversion, agreeableness, emotional stability), risk preferences, age, sex, number of children, occupation, and employment status. All variables were standardised before logistic regression. Matching was performed using nearest-neighbour with a caliper of 0.05.
Standardised % bias across all 11 covariates after nearest-neighbour PSM (caliper 0.05). All covariates below the 4% threshold — the matched sample is well-balanced on all observable characteristics.
PSM result: −0.299 points (95% CI −1.255 to 0.658, p = 0.48). The substantial sample reduction during matching (from 215,733 to 1,077 observations) severely reduces statistical power, preventing significance. However, the direction of the effect is consistent with the main DiD finding — PSM confirms the negative sign and cannot reject the main result.
Additional robustness checks in the appendix: Hausman test for fixed vs. random effects, placebo regressions for alternative treatment years, varied time windows, reverse causality test, and further stratified DiD analyses by age, gender, family type, and mental health baseline.
The DiD estimates are used to construct an Atkinson SWF (γ = 1) for two scenarios, comparing their Prioritarian WELLBYs to determine which policy yields greater social welfare when the worst-off receive priority weight.
Life satisfaction scores predicted from the main DiD model (Equation 1) for Victorian citizens. Predicted data, rather than observed data, is used to isolate the causal lockdown effect. Scores are scaled to the full Victorian population using SCQ survey weights, then transformed using the Atkinson SWF with γ = 1 (logarithmic transformation).
Life satisfaction from the 2020 control group represents what Victorian citizens' wellbeing would have looked like without the lockdown policy (but still including broader pandemic effects). In addition, a Business-as-Usual scenario would have led to 44,756 additional deaths (applying a 0.67% mortality rate, Ferranna et al. (2022), to Victoria's 6.68 million population). These lost WELLBYs — calculated using a remaining lifetime of 6 years and an indifference point of 2 (Peasgood et al., 2018) — are subtracted from the Business-as-Usual welfare total.
Where: D = 44,756 (estimated additional deaths); I = 2 (indifference point between life and death, Peasgood et al. 2018); Y = 6 (remaining lifetime in years, Layard et al. 2020); γ = 1 (moderate inequality-aversion parameter).