Policy Design in Long-run Welfare Dynamics
Authors: Jiduan Wu, Rediet Abebe, Moritz Hardt, Ana-Andreea Stoica
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We illustrate our theoretical findings using simulations, which highlight the risks of evaluating policies based solely on their short-term effects. Our results underscore the necessity of considering long-term horizons in designing and evaluating welfare policies; the true efficacy of even well-established policies may only emerge over time. (...) We illustrate our theoretical results through simulations using a real-world dataset. We compare the average social welfare under a finite time horizon for all proposed policies. (...) We use data collected from the Survey of Income and Program Participation (SIPP) (Bureau, 2023), which is a longitudinal survey of households in the U.S. containing variables related to economic well-being such as income, employment, etc. (...) We measure social welfare at timestep t as the individual growth rate up to time t averaged over all individuals (equation 2 up to time t). The average social welfare (solid lines) converges to the theoretical expected welfare (dashed lines) for all policies (Figure 1). |
| Researcher Affiliation | Academia | 1Max Planck Institute for Intelligent Systems, T ubingen and T ubingen AI Center 2Department of Computer Science, ETH Z urich 3ELLIS Institute, T ubingen |
| Pseudocode | No | The paper describes algorithms and policies in prose (e.g., max-fg, max-U, min-U policies in Section 2.2) and mathematical formulations, but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | All simulations are ran on commodity hardware, using Python 3.8. All code and data used in our simulations is available in this repository. |
| Open Datasets | Yes | We use data collected from the Survey of Income and Program Participation (SIPP) (Bureau, 2023), which is a longitudinal survey of households in the U.S. containing variables related to economic well-being such as income, employment, etc. Among numerous indices, we use the income variable as a proxy for the initial individual welfare level, (Ui(0))i. (...) U.S. Census Bureau. Survey of income and program participation (SIPP) 2014 panel. https://www.census.gov/programs-surveys/sipp/data/datasets/2014-panel/, 2023. Accessed: 2024-09-25. |
| Dataset Splits | No | The paper states: "We group the whole population 39, 720 into 13 bins and treat every 200 samples as one individual, and every $1, 000 as one welfare unit in our model. In the end, we obtain a population of 206 individuals." This describes data pre-processing and aggregation into a simulation population, but it does not specify explicit training, validation, or test splits for experimental reproduction. |
| Hardware Specification | No | The paper states: "All simulations are ran on commodity hardware, using Python 3.8." This statement is too general and does not provide specific details such as CPU/GPU models, memory, or other hardware specifications. |
| Software Dependencies | Yes | All simulations are ran on commodity hardware, using Python 3.8. |
| Experiment Setup | Yes | We simulate an instance of the general model from equation 1 with Gaussian noise, specified as: Ui(t + 1) Ui(t) = ai(t) fi(Ui(t)) (1 ai(t)) gi(Ui(t)) + ξi(t), t 0, i [N]. where {ξi(t)}i,t N(0, σ2) and capped within uniform bounds, for some noise parameter σ. We generate homogeneous bounds f , f +, g , g+, and then generate the functions fi( ), gi( ) as segment linear functions. Our results are averaged over 100 draws, reporting standard deviation in the error bands. See Appendix C for further experimental details. (...) Appendix C: For the simulations in which the return and decay function bounds are uniform, we choose threshold parameters F i , F + i , G i , G+ i s.t. fi(x) = f , gi(x) = g+ for x F i , x G i , respectively, and fi(x) = f +, gi(x) = g for x F + i , x G+ i , respectively, and we linearly interpolate between these thresholds. Choosing F i < F + i and G i > G+ i ensures that fi is increasing and gi is decreasing on the non-constant segments. We generate F i , F + i , G i , G+ i randomly in the interval (0, ] for some > 0. For Figures 1, 2, and 4, we filter to ensure that fi( ) + gi( ) is increasing. For Figure 3, we filter to ensure that fi( ) + gi( ) is increasing under some threshold τ, and decreasing above threshold τ. For all figures, we average over 50 iterations and report the social welfare obtained at every timestep. Our results are qualitatively the same for other functional forms of fi, gi, such as sigmoid functions. |