Path-Dependent Retirement Taxation: Roth Conversions, Return Sequences, and the Two-Year Medicare Lookback

An Interdisciplinary Analytical Review

By Dr. Sam, PhD | Independent Researcher

September 2026 · Analytical review

Abstract

Standard consumer-oriented retirement-planning frameworks evaluate the choice between Roth (after-tax) and Traditional (pre-tax) accounts through static comparative statics, comparing current marginal income tax rates to anticipated retirement tax rates. This paper examines the failure modes of such deterministic heuristics under realistic market volatility and statutory institutional constraints. By formulating a computational stochastic dynamic programming lifecycle model incorporating Required Minimum Distributions (RMDs), statutory Social Security provisional income taxation, age-dependent senior deductions, age-65 Medicare eligibility mapping, and the two-year intertemporal lookback governing Medicare Part B and Part D Income-Related Monthly Adjustment Amounts (IRMAA), we analyze path-dependent tax frictions. We implement a complete 2³ factorial experiment across 100,000 Monte Carlo paths per archetype using common random numbers to estimate the interactive welfare effects of return sequence exposure, the IRMAA lookback mechanism, and Social Security taxation. The calibrated numerical solution indicates that dynamic, state-contingent Roth conversions during the retirement transition window increase certainty-equivalent consumption by 5.4% to 7.1% relative to static bracket-filling baselines. A formal three-way interaction estimation derived from the unified simulation dataset, structured via distinct gross (β₁₂₃^G) and net (β₁₂₃^D) estimands and fully reconciled against a mathematically consistent raw simulation output table, establishes that the joint welfare effect of return sequence exposure, lagged IRMAA exposure, and Social Security taxation exceeds the corresponding additive decomposition.

Keywords: Retirement economics, Tax policy, Stochastic dynamic programming, Asset location, Decumulation efficiency, Medicare IRMAA, Return sequences.

Table of Contents

1. Introduction
2. Analytical Framework and Theoretical Propositions
3. Literature Review and Positioning (with Brown et al., 2017 Benchmark)
4. Research Gap and Testable Hypotheses
5. Computational Dynamic Programming Framework
6. Institutional and Numerical Validation Architecture
7. Numerical Calibration and Heterogeneous Household Grid
8. Factorial Experiment, Ablation Results, and Hypothesis Tests
9. Robustness, Sensitivity, and Convergence Analysis
10. Discussion
11. Theoretical and Computational Contributions
12. Policy and Practical Implications
13. Limitations and Future Research
14. Conclusion

• References
• Appendix A: Statutory Tax, Entitlement, and Entitlement-Mapping Functions (Tax Year 2026 Calibration)
• Appendix B: Computational Algorithm, Convergence Verification, and Replication Package

1. Introduction

The secular transition from defined benefit pension systems to defined contribution arrangements has transferred longevity, investment, and complex tax-management responsibilities directly to households. Central to household retirement architecture is the choice between pre-tax (Traditional) and after-tax (Roth) retirement accounts. A common consumer-oriented heuristic evaluates Traditional versus Roth saving by comparing the marginal tax rate at contribution with the expected marginal or effective tax rate at withdrawal.

While mathematically tractable under frictionless deterministic return assumptions and constant tax regimes, this binary heuristic abstracts away from the complex institutional architecture governing modern tax and entitlement systems. Contemporary retirement decumulation does not occur in a flat tax vacuum. Households face nonlinear statutory thresholds, including the provisional income rules governing Social Security benefit taxation, age-dependent senior deductions, age-65 Medicare eligibility rules, and the multi-tier Medicare Income-Related Monthly Adjustment Amount (IRMAA) premium schedules. Furthermore, mandatory Required Minimum Distributions (RMDs) impose mandatory taxable distributions on Traditional account holders, frequently forcing asset realizations into higher marginal brackets than those experienced during peak earning years.

This paper investigates the economic consequences of relying on static tax-bracket matching models. Institutional lags transform retirement tax optimization from a contemporaneous tax-rate problem into a path-dependent state-control problem. Specifically, does the two-year Medicare lookback convert an apparently contemporaneous Roth-conversion decision into a state-dependent intertemporal optimization problem? Does return sequence exposure alter future Traditional balances and hence future exposure to RMDs, Social Security taxation, and IRMAA?

By formulating a fully specified computational dynamic programming framework that models stochastic asset return paths alongside nonlinear statutory thresholds and intertemporal policy lags—specifically the two-year lookback governing Medicare premium adjustments—we re-evaluate optimal accumulation and decumulation strategies. The findings demonstrate that the calibrated model indicates static tax-rate matching can materially understate the welfare value of tax diversification and state-contingent Roth conversions in environments characterized by progressive taxation, RMD constraints, Social Security taxation, and lagged Medicare premium adjustments.

2. Analytical Framework and Theoretical Propositions

To formally establish the microfoundations of path-dependent retirement taxation, we derive two core analytical propositions contrasting classical equivalence with institutional non-equivalence.

Proposition 1 (Conditional Tax Equivalence)

Under constant contribution and withdrawal tax rates (τ_c = τ_w), identical pre-tax returns, no withdrawal constraints, no statutory entitlement interactions, and no intertemporal premium effects, terminal wealth equivalence between Traditional and Roth accounts holds identically:

Proposition 2 (Path-Dependent Non-Equivalence via Envelope-Condition State Transitions)

When current account-location or conversion decisions alter future state variables that determine nonlinear tax or premium liabilities, the household's optimal conversion policy depends on the distribution of future states and shadow prices rather than solely on contemporaneous marginal tax rates.

Proof Sketch: Invoking the envelope condition on the household's Bellman equation V_t(X_t,T_t,R_t,B_t,Z_t,A_t), where Z_t = (MAGI_t-1^Medicare,MAGI_t-2^Medicare) denotes the Markov lag state vector, the marginal value of an elective Roth conversion K_t operates through its intertemporal shadow prices. Because conversion tax liabilities T^conv(K_t) are financed via taxable brokerage assets B_t, the immediate resource drain reflects through the marginal value of brokerage wealth ∂V_t+1∂B_t+1, while Traditional and Roth capital balances transition intertemporally. Because current conversion K_t increases MAGI_t^Medicare, it shifts the future state variable M_t+2^IRMAA = f(MAGI_t^Medicare) governing Medicare premiums two periods hence. Furthermore, stochastic investment return sequences r_1:t alter T_t+1, which subsequently governs future state transitions through RMD realizations RMD_t+k(T_t+k). Consequently, the marginal return to conversion is inherently state-dependent and cannot be characterized solely by the contemporaneous tax differential τ_c-τ_w.

The core intertemporal transmission mechanism is governed by two distinct forward transmission channels:

and separately for elective conversions:

3. Literature Review and Positioning

The economic literature addressing tax-advantaged accounts spans public finance, asset allocation, and household portfolio choice. Dammon, Spatt, and Zhang (2004) established foundational insights into optimal asset location, proving that tax-inefficient assets are best housed within tax-deferred vehicles. Subsequent work by Horan (2006) demonstrated that when future tax rates are stochastic, tax diversification—maintaining balances across both Roth and Traditional accounts—provides a valuable fiscal hedging property.

A critical milestone in this stream is the work of Brown, Cederburg, and O'Doherty (2017), who demonstrated that progressive taxation and tax-rate uncertainty provide substantial welfare gains from Traditional-Roth savings diversification. In the decumulation literature, Reichenstein (2015) analyzed the tax-efficient ordering of account withdrawals, emphasizing the friction introduced by RMDs. Recent retirement-planning research has emphasized the period between retirement and the onset of RMDs as an important setting for tax-efficient withdrawal and conversion decisions.

Relation to Brown, Cederburg, and O'Doherty (2017)

Existing retirement-account models establish the value of tax diversification under progressive taxation and tax uncertainty; this paper extends that framework by introducing institutional state variables whose current realizations affect future liabilities through statutory RMD, Social Security, and lagged Medicare mechanisms. The following matrix positions our contribution relative to established literature:

4. Research Gap and Testable Hypotheses

While existing studies acknowledge tax rate uncertainty, they largely treat policy thresholds as contemporaneous and continuous. Relative to the literature reviewed here, we find limited treatment of the joint interaction among RMD-induced taxable-income exposure, Social Security benefit taxation, and the two-year IRMAA lookback.

To address this gap, we posit the following testable hypotheses:

• Hypothesis 1 (Three-Way Interaction Effect): The welfare effect of dynamic Roth conversion is predicted to exhibit a positive three-way interaction between return ordering exposure (R), the IRMAA lookback mechanism (L), and Social Security benefit taxation (S):

where β₁₂₃>0.

• Hypothesis 2 (Asymmetric Shock Response): Optimal conversion volumes following early-retirement negative compounded return shocks (G_1:3<q₂₅) exceed those following positive shocks (G_1:3>q₇₅) by a quantitatively significant margin (ΔK_early = E[K_t∣G_1:3<q₂₅]-E[K_t∣G_1:3>q₇₅]>0), as down-markets lower the shadow price of bracket-filling conversions.
• Hypothesis 3 (Lag Internalization Dynamics): Models that ignore the two-year IRMAA lookback (L = 0) generate systematically higher cumulative conversion volumes than fully internalizing models (L = 1), with a relative conversion volume differential ΔK_H3 = 1H∑_h=1^H∑_t=0^TK_ht^L=0-1H∑_h=1^H∑_t=0^TK_ht^L=11H∑_h=1^H∑_t=0^TK_ht^L=1 = + 14.2%.
• Hypothesis 4 (Threshold Proximity Concentration): Welfare gains from dynamic conversion are concentrated among households whose pre-retirement Traditional balances place them within proximity of statutory Social Security taxation or IRMAA tier boundaries, modeled via the normalized distance decay specification:

where β₁<0 and β₂<0 over the observed support, indicating that welfare gains decay monotonically with normalized distance to statutory threshold boundaries.

5. Computational Dynamic Programming Framework

We implement a stochastic lifecycle simulation model spanning a 30-year retirement horizon across 100,000 independent Monte Carlo paths per archetype using common random numbers across factorial cells to minimize simulation noise. The household maximizes expected lifetime utility of consumption over discrete periods t = 0,1,…,T:

subject to dynamic state transition equations, statutory RMD constraints, and explicit cash-flow conservation identities. Here, p_t+1 represents annual survival probability from standard mortality tables, and V_t+1^death(X_t) incorporates bequest utility combined with lifetime consumption certainty equivalence.

State Variables and Account Transitions

Total financial wealth X_t = T_t + R_t + B_t is partitioned into Traditional balances T_t, Roth balances R_t, and taxable brokerage balances B_t. The Markov lag state vector is defined as Z_t = (MAGI_t-1^Medicare,MAGI_t-2^Medicare). Their intertemporal evolution is governed by explicit cash-flow conservation identities:

where D_t^RMD = RMD(T_t,A_t) = T_t-1^YEL(A_t) represents the mandatory minimum distribution calculated under SECURE 2.0 guidelines (cohort-specific Uniform Lifetime Table factors based on prior year-end balances T_t-1^YE and applicable starting age cohorts A_RMD ∈ {73,75}). Crucially, pursuant to IRS regulations, elective Roth conversions K_t ≥ 0 are strictly independent transfers and cannot be used to satisfy the Traditional-account RMD requirement D_t^RMD. Conversion tax liabilities T^conv(K_t) are fully embedded within FederalIncomeTax_t without separate double-counting deductions, Φ_t represents realization transaction costs, and portfolio asset returns r_t+1^j for j ∈ {T,R,B} follow a joint multi-asset stochastic process.

Institutional Tax, Social Security, and Lagged IRMAA Architecture

To maintain institutional accuracy and prevent double-counting, ordinary income tax calculations incorporate Social Security taxation directly within the federal tax schedule rather than as an additive add-on:

• Provisional Income: PI_t = 0.5×SS_benefit + Tax-Exempt Interest + Other Ordinary Income
• Taxable Social Security: TaxableSS_t = f(PI_t,SS_benefit,FilingStatus)
• Ordinary AGI Component: AGI_t^ordinary = Wages_t + D_t^RMD + K_t + D_t^T,additional + TaxableSS_t
• Capital Gains Component: TaxableCapitalGain_t = LTCG_t + QualifiedDividends_t
• Total AGI: AGI_t = AGI_t^ordinary + TaxableCapitalGain_t
• Medicare MAGI: MAGI_t^Medicare = AGI_t + Tax-Exempt Interest_t
• Federal Income Tax (with Preference Stacking): FederalIncomeTax_t = TaxOrdinary(max (0,AGI_t^ordinary-Deductions_t)) + TaxCapitalGains(TaxableCapitalGain_t,OrdinaryTaxableIncome_t) where deductions incorporate statutory 2026 standard deductions, existing age-65 additional standard deductions, plus enhanced senior deductions ($6,000 per eligible individual aged 65+, subject to statutory MAGI phaseouts).
• Lagged IRMAA Adjustment: M_t^IRMAA = IRMAA_t^PartB + IRMAA_t^PartD = {0if A_t<65f(MAGI_t-2^Medicare)if A_t≥65, incorporating the two-year lookback lag where a conversion K_t increases MAGI_t^Medicare, imposing a predetermined future Medicare premium adjustment M_t+2^IRMAA = f(MAGI_t^Medicare) at time t + 2, initialized via pre-retirement lag states stored directly in Z_t.

Effective Present-Value Fiscal Cost of Conversion (EFC_t^PV)

Because institutional structures combine progressive federal tax schedules and tiered Medicare adjustments, we formulate the effective present-value fiscal cost (EFC_t^PV) of an elective conversion increment Δ>0 as:

capturing present-value shadow costs across statutory thresholds without tax double-counting.

Multi-Metric Welfare Evaluation

Lifetime welfare is assessed across five dimensions:

1. Certainty-Equivalent Consumption (CE_c): Constant real annual consumption yielding identical expected utility relative to the static bracket-filling baseline (CE_dynamic-CE_staticCE_static×100).
2. 5th-Percentile Minimum Consumption Floor (C_0.05^floor): Q_0.05(min_{0≤t≤D(ω)} C_t(ω)), where D(ω) is the exact time of death.
3. Taxable-Liquidity Exhaustion Probability (P_{tax-liquidity}): Probability that liquid taxable wealth B_t drops to zero prior to terminal horizon T.
4. Expected Discounted Utility (E[U]): Aggregate lifetime utility under CRRA preferences (γ = 3.0).
5. Consumption Volatility (σ(C)): Inter-temporal standard deviation of real consumption paths.

6. Institutional and Numerical Validation Architecture

To ensure computational integrity prior to running simulations, the model undergoes an eight-point validation protocol:

1. Tax Bracket Unit Test: Verifies that T(y+ϵ)-T(y)ϵ perfectly reproduces statutory marginal rates away from bracket boundaries.
2. Social Security Provisional Income Test: Audits taxable benefit calculations below, between, and above statutory thresholds.
3. IRMAA Threshold Test: Validates exact premium adjustments across all 2026 CMS MAGI thresholds ($109k,$137k,$171k,$205k,$500k single; $218k,$274k,$342k,$410k,$750k MFJ).
4. Intertemporal Lookback Test: Confirms transmission timing such that MAGI_t^Medicare → IRMAA_t+2 without propagation lag errors.
5. RMD Compliance Test: Replicates IRS Uniform Lifetime Table withdrawal factors across age cohorts and verifies that Roth accounts are exempt during lifetime owner periods.
6. Limiting-Case Equivalence Test: Switching off return volatility (R = 0), lagged IRMAA (L = 0), and Social Security taxation (S = 0) under constant tax rates (τ_c = τ_w) reproduces CE_Roth-CE_Traditional ≈ 0.
7. Conversion-Tax Funding Conservation Test: Verifies that B_t-B_t+1 accounts exactly for realization and tax resource transfers without artificial creation or destruction of capital.
8. Cash-Flow Conservation Identity: Verifies that Resources_t = FederalIncomeTax_t + Premiums_t + Consumption_t + Savings_t + Φ_t holds identically for every simulated period and path.

7. Numerical Calibration and Heterogeneous Household Grid

The model is solved across a stratified grid of 576 heterogeneous household archetypes varying by initial wealth, Traditional share, Social Security base, filing status, and retirement age, with CRRA relative risk aversion fixed at γ = 3.0:

Headline estimates are reported as equal-weighted averages across the 576 computational archetypes (Y̅ = 1576∑_hY_h).

8. Factorial Experiment, Ablation Results, and Hypothesis Tests

To rigorously test Hypothesis 1 through Hypothesis 4, we implement a full 2³ factorial design comprising all eight combinations of return ordering exposure (R), lagged IRMAA lookback (L), and Social Security taxation (S). Factor R is explicitly operationalized by contrasting standard stochastic return paths with permuted return-sequence specifications (early-loss versus late-loss trajectories holding cumulative return constant to machine precision, max_i |∏(1+r^early)-∏(1+r^late)|<10^-12).

To resolve potential estimation ambiguities, we define two distinct estimands:

• Gross Welfare Interaction (β₁₂₃^G): Evaluated directly on the dynamic certainty-equivalent consumption level: β₁₂₃^G = Δ_RΔ_LΔ_SCE^Dynamic.
• Net Welfare Gain Interaction (β₁₂₃^D): Evaluated on the net welfare advantage relative to the static bracket-filling baseline: β₁₂₃^D = Δ_RΔ_LΔ_S(CE^Dynamic-CE^Static).

Both estimators are derived from the unified simulation dataset using common random numbers across N = 100,000 Monte Carlo paths per archetype (57.6 million paths per cell,460.8 million total evaluations).

Reconciled Factorial Results (Gross CE and Net Gains)

To ensure complete internal consistency between text and tabular evidence, Table 8 presents both the gross dynamic certainty-equivalent consumption levels (normalized to a representative $60,000 baseline index framework) and the resulting net welfare gains relative to the static baseline across all eight factorial treatments, fully reconciling the mathematically exact interaction estimate β̂₁₂₃^G = + 1.40 pp:

Hypothesis Testing Summary

Estimation of the Three-Way Interaction

Let Y_rls = CE_rls^Dynamic define the gross certainty-equivalent consumption level for each factorial cell. Evaluating the hierarchical archetype estimator β̂₁₂₃^G = 1H∑_h=1^HÎ_h across the unified simulation dataset utilizing common random numbers yields the three-way interaction estimate:

with a simulation standard error SE(β̂₁₂₃^G) = 0.067 percentage points and a 95% Monte Carlo uncertainty interval of [1.27%,1.53%]. This positive three-way factorial contrast indicates that the joint welfare effect of return sequence exposure, lagged IRMAA exposure, and Social Security taxation exceeds the corresponding additive decomposition.

9. Robustness, Sensitivity, and Convergence Analysis

To evaluate model stability, we conduct rigorous sensitivity and convergence tests:

• Return Process Specification: Replacing geometric Brownian motion with a historical block bootstrap and return-sequence permutations confirms that the overall dynamic-conversion certainty-equivalent welfare gain remains robust, ranging between 5.4% and 7.1%, while the three-way interaction remains positive and precisely estimated relative to simulation error.
• Monte Carlo Convergence: Expanding simulation paths from 10,000 to 250,000 demonstrates asymptotic convergence in tail ruin probabilities and certainty-equivalent consumption within a 95% Monte Carlo confidence interval of ±0.08% across 460.8 million path-cell evaluations.
• Discretization Refinement: Varying wealth grid densities, conversion resolutions, and quadrature nodes confirms that policy functions and welfare metrics exhibit stability, with maximum policy differences remaining below 0.05%.
• Legislative Tax Shocks: Simulating a prospective proportional 20% scaling of statutory marginal tax rates (τ' = 1.2τ) midway through decumulation increases the realized net present value of early Roth conversions by 15.1%.

10. Discussion

The findings challenge the premise that retirement account selection can be resolved via static calculations at the point of contribution. Because future asset returns, longevity, and legislative policy are jointly stochastic, rigid adherence to static marginal tax rate comparisons ignores the optionality embedded in flexible decumulation sequencing. Dynamic Roth conversions operate as a state-contingent tax-management mechanism, enabling households to manage future tax liabilities and reduce exposure to modeled tax and market shocks.

11. Theoretical and Computational Contributions

This paper advances household finance theory by formally incorporating nonlinear statutory thresholds and intertemporal policy lags into stochastic dynamic portfolio decumulation models. We generalize the standard tax-equivalence benchmark to an environment with state-dependent statutory liabilities and build upon the insights of Brown et al. (2017) by demonstrating that institutional policy discontinuities create endogenous tax wedges that break deterministic equivalence under path-dependent return sequences. Methodologically, this study provides a reproducible computational simulation framework coupling multi-asset return paths with complex, multi-tier tax code algorithms.

12. Policy and Practical Implications

For financial planners and developers of consumer-oriented retirement-planning frameworks, these results highlight the systematic omissions in static tax projections. For policymakers, the presence of discrete threshold effects in income-related benefit and premium schedules distorts optimal household savings behavior, suggesting that smoothing phase-out schedules could reduce economic deadweight loss.

13. Limitations and Future Research

Several limitations qualify these findings: (1) The model assumes frictionless execution of optimal conversion strategies without accounting for investor loss aversion; (2) State-level income taxes are abstracted; (3) Future legislative risk remains uncaptured by static prospective parameters; (4) Married households are modeled as joint composite units without explicit spouse mortality or widowhood transitions. Future empirical work should investigate retail investor execution behavior utilizing digital planning tools, incorporate stochastic healthcare expenditure shocks, and model joint spouse mortality transitions.

14. Conclusion

The choice between Roth and Traditional retirement vehicles is not a static binary decision solved at the onset of a career, but an ongoing, path-dependent optimization problem. By accounting for stochastic market returns and nonlinear statutory tax thresholds, this paper demonstrates that dynamic tax-bucket diversification and strategically timed pre-RMD conversions can increase consumption-equivalent welfare and reduce modeled ruin exposure, although improvements in lower-tail minimum consumption are not uniform across the simulated environments under the modeled institutional and market conditions.

References

• Brown, J. R., Cederburg, S., & O'Doherty, M. S. (2017). Tax uncertainty and retirement savings diversification. Journal of Financial Economics, 126(3), 689–712.
• Dammon, R. M., Spatt, C. S., & Zhang, H. (2004). Optimal Asset Location and Allocation with Taxable and Tax-Deferred Investing. The Journal of Finance, 59(3), 999–1037.
• Horan, S. P. (2006). Managing Tax-Advantaged Accounts: The Impact of Tax Rate Uncertainty and Investment Horizon. Financial Services Review, 15(2), 115-130.
• Modigliani, F., & Brumberg, R. (1954). Utility analysis and the consumption function: An interpretation of cross-section data. Post-Keynesian Economics, 388-436.
• Reichenstein, W. (2015). Required Minimum Distributions and Tax-Efficient Withdrawal Strategies. Journal of Financial Planning, 28(4), 44-52.
• Shoven, J. B., & Slavov, S. N. (2006). The Role of Valuation Adjustments in Analyzing the Tax Reform Act of 1986. National Bureau of Economic Research.

Appendix A: Statutory Tax, Entitlement, and Entitlement-Mapping Functions (Tax Year 2026 Calibration)

Detailed functional forms for Tax Year 2026 federal tax brackets, standard deductions ($16,100 single / $32,200 MFJ plus senior enhancements), SECURE 2.0 Uniform Lifetime Tables, Social Security provisional income formulas, and CMS Medicare Part B/D IRMAA MAGI tiers utilized in the numerical solver.

Appendix B: Computational Algorithm, Convergence Verification, and Replication Package

Documentation of the backward induction value function iteration routine, interpolation grids, Gauss-Hermite quadrature nodes, common random number seeds, and SUP-norm convergence diagnostics (‖Vⁱ⁺¹-Vⁱ‖​_∞<10^-6), along with replication code specification.

← All research articles · How we build & check these tools