Conclusion

This paper has shown that the inability of workers to accurately perceive their marginal tax rates imposes a welfare cost on the order of 0.1% of GDP. Under quasilinear-isoelastic preferences, the expected deadweight loss per worker takes the closed form $\frac{1}{2}\varepsilon\sigma^{2}/(1-\tau)$ times earnings. Calibrated to a Frisch elasticity of 0.33, a misperception standard deviation of 12 percentage points ($\sigma = 0.12$), and a mean marginal rate of 0.30, the stylized central estimate is $187 per worker and $30 billion in aggregate---0.11% of GDP annually. Applying the formula to household-level marginal tax rates from PolicyEngine-US microsimulation yields a higher aggregate of $37 billion, as the convexity of the DWL formula amplifies losses from the heavy tails of the income and tax rate distributions. The top income quintile bears 59% of total DWL, while the bottom quintile bears less than 2%. The full sensitivity range spans 0.04% to 0.25% of GDP ($10 to $71 billion), reflecting plausible variation in labor supply elasticities and misperception magnitudes. These losses arise not from future policy uncertainty but from the current opacity of the tax code.

Policy implications¶

The quadratic dependence on $\sigma$ means that proportional reductions in misperception variance have a larger effect on DWL than proportional reductions in the tax rate, holding all else equal. A 5 percentage point reduction in the misperception standard deviation (from $\sigma = 0.12$ to $\sigma = 0.07$) lowers $\sigma^{2}$ by 66%, cutting per-worker deadweight loss by roughly two-thirds. By contrast, a 3 percentage point reduction in the statutory rate (from $\tau = 0.30$ to $\tau = 0.27$) reduces the $1/(1-\tau)$ term by only about 4%. The relative cost-effectiveness of these approaches depends on implementation costs that are outside the scope of this model. The model is agnostic about how misperception is reduced. Several channels could lower $\sigma$, each with different costs and tradeoffs:

• Better information provision: More accurate withholding tables, clearer pay-stub reporting of marginal rates, or interactive tools that show workers their true rate could reduce $\sigma$ without any change to the tax code itself.

• Tax simplification: Consolidating brackets, phase-outs, and overlapping credits would make the rate schedule more legible, but provisions typically exist for distributional or incentive reasons. The net welfare gain from simplification depends on what is lost in the process.

• Professional tax preparation: Wider access to competent preparation could reduce misperception, though the compliance cost literature suggests current preparation spending is already substantial National Taxpayers Union Foundation, 2023.

The model does not rank these channels; it quantifies the welfare cost of a given $\sigma$ regardless of its source. Because DWL is quadratic in $\sigma$, a reduction from 0.12 to 0.10 lowers DWL by 31%, and a reduction to 0.07 lowers it by 66%.

The optimal tax analysis reinforces this point from a different angle. A utilitarian planner who accounts for misperception chooses a lower optimal linear tax rate of 42.9%, compared with 44.5% under perfect information. The gap of 1.6 percentage points reflects the additional efficiency cost that misperception imposes at the margin: each dollar of revenue raised creates more deadweight loss when workers cannot accurately perceive the rate they face. Within the model, lower $\sigma$ shifts the optimal linear tax rate upward, allowing more redistribution at a given level of total deadweight loss.

Limitations¶

Several caveats apply. The model is static: workers choose labor supply once, facing a single marginal rate. In reality, labor supply decisions unfold over time, and workers may learn about their true rate as the year progresses. This suggests the static model may overstate the cost for workers with stable employment but understate it for those making discrete labor market transitions (entering or leaving the workforce, choosing between jobs with different hours).

The analysis assumes a linear tax schedule with a single marginal rate. The actual U.S. tax code is piecewise linear with multiple brackets, phase-outs, and cliffs. Misperception of a non-linear schedule may differ qualitatively from misperception of a single rate, as workers may be uncertain not only about the level of their marginal rate but also about where bracket thresholds fall.

I have assumed that the misperception error $\delta$ is normally distributed with mean zero and constant variance across the population. In practice, misperception may be correlated with income (lower-income workers may face more complex effective schedules due to benefit phase-outs), with financial sophistication, and with access to tax preparation services. The model treats $\sigma$ as a population-level parameter and does not capture this heterogeneity. Moreover, the calibrated $\sigma = 0.12$ is drawn from Rees-Jones & Taubinsky (2020), who measured misperception of federal income tax rates only. Since the comprehensive marginal tax rate includes payroll taxes, state taxes, and benefit phase-outs, the true comprehensive $\sigma$ may differ from 0.12. If errors across tax components are partially independent, the comprehensive $\sigma$ would exceed 0.12; if workers who misperceive federal rates also misperceive other components in the same direction, the difference would be smaller.

Finally, the analysis focuses exclusively on labor supply and ignores other margins of response---savings, portfolio allocation, tax avoidance, and organizational form---that may also be distorted by rate misperception. These omitted margins could increase or decrease the total welfare cost depending on the sign and magnitude of the distortions they introduce.

Future work¶

Two extensions would strengthen and generalize these results. First, while this paper applies the DWL formula to heterogeneous $\tau_i$ via PolicyEngine microsimulation, it maintains homogeneous $\sigma$. The assumption of uniform misperception can be relaxed by linking misperception variance to observable household characteristics (income, filing status, use of tax preparers) using the survey data from Gideon (2017) and Rees-Jones & Taubinsky (2020). Heterogeneous $\sigma_i$ would sharpen the aggregate estimates and identify which populations and provisions contribute most to misperception-induced welfare loss.

Second, measuring comprehensive misperception---not just federal income tax misperception---is essential for accurate welfare accounting. One approach would be to replicate the Rees-Jones & Taubinsky survey instrument but expand it to ask about total marginal rates (including payroll, state, and benefit phase-outs). An alternative approach would use large language models as a benchmark: presenting realistic household scenarios to AI systems and comparing their estimated MTRs to true rates from PolicyEngine could validate the structural misperception hypothesis while isolating the contribution of tax system complexity from individual cognitive limitations.