Appendix: public dossier on DYNASIM

Purpose

The main comparison chapter treats DYNASIM strategically. This appendix asks a narrower question: what can an outside researcher actually know about DYNASIM from public sources alone?

The answer is: quite a lot about scope, inputs, modeled processes, and policy use cases; much less about the current codebase, full equation inventory, and day-to-day operating workflow. That distinction matters. For this repository, DYNASIM is not just a competitor model. It is the most important benchmark for how much detail a non-governmental dynamic microsimulation platform can accumulate over time.

1. What the public record establishes

Dimension What public sources show
Institutional home DYNASIM is maintained at the Urban Institute and has been developed there for decades (Favreault and Smith 2004; Favreault et al. 2015).
Current public version Urban’s current public fact sheet refers to DYNASIM4 (Urban Institute 2024).
Time horizon Public materials describe DYNASIM as producing 75-year projections for older adults’ economic and health outcomes (Favreault et al. 2015; Urban Institute 2024).
Starting sample Public documentation says the basic DYNASIM4 starting sample is 0.04 percent of the US population, with an expanded 0.4 percent version for some work (Urban Institute 2024).
Broad scope DYNASIM4 projects demographics, work, Social Security, pensions, income, wealth, taxes, health, disability, LTSS, and several other public programs (Favreault et al. 2015; Urban Institute 2024).
Rule-based components Public sources describe sophisticated rule-based calculators for OASDI and SSI, while also noting that some other government benefits are simulated statistically rather than through full eligibility rules (Favreault et al. 2015; Urban Institute 2024).
Alignment Public materials explicitly say some major outcomes are aligned to external targets produced by SSA’s Office of the Chief Actuary (Urban Institute 2024).
Public access model Urban publishes documentation, briefs, and some downloadable tabulations, but not an open-source codebase or a fully self-service public model (Favreault and Smith 2004; Favreault et al. 2015; Urban Institute 2024).

2. Lineage and versioning

Urban’s public primer states that DYNASIM was originally developed in the 1970s, that DYNASIM2 was built in the early 1980s, and that DYNASIM3 was a major redevelopment with a newer starting sample, expanded saving and pension modules, and updated Social Security and SSI calculators (Favreault and Smith 2004).

The later overview and fact sheet make clear that DYNASIM4 is not a small patch on top of DYNASIM3. It uses a newer SIPP base and a much broader health, health spending, and LTSS apparatus (Favreault et al. 2015; Urban Institute 2024). Publicly, that matters for two reasons:

  • DYNASIM has accumulated institutional memory over multiple generations rather than being assembled in one grant cycle.
  • The public paper trail is versioned and uneven. Older generations are often documented in more detail than the current operational model.

3. Starting sample, projection horizon, and scale

The 2015 overview notes that DYNASIM4 would be based on the 2004 and 2008 SIPP panels and would start projecting outcomes in 2006 (Favreault et al. 2015). Urban’s 2024 fact sheet adds two practical details that are especially useful for outside readers:

  • the basic starting sample is 0.04 percent of the population
  • an expanded version uses a 0.4 percent starting sample

The 2015 overview also notes a larger starting sample of 1.056 million people in 461,000 families, typically processed only through 2040 rather than through the full long horizon (Favreault et al. 2015). Taken together, these sources imply that Urban varies the effective scale of the model depending on the task, but the public record does not fully document which configuration is standard for which class of analysis.

Urban also publishes a small amount of direct model output, including a DYNASIM4 projections-by-birth-cohort page with downloadable tables. That is helpful, but it is still a long way from a public model release.

4. Weights and representation

The public DYNASIM materials describe starting-sample scale and alignment to external controls, but they do not describe a system of annual independent person weights that can be recalibrated after family events occur (Urban Institute 2024). That distinction matters for this project. Once marriage, divorce, fertility, leaving home, and household splitting are simulated, arbitrary person-level reweighting would make the relationship network incoherent.

The right design is to treat representation as part of the simulated population object. Base-year weights can be absorbed into the starting population, converted into stable replicate counts, or carried as network-consistent representation factors. Dynamic alignment should then operate through event selection, process parameters, and network-preserving resampling rather than by independently changing the importance of linked people. This is also the methodological warning in the dynamic-ageing microsimulation literature on weights (Dekkers and Cumpston 2012).

5. Software structure and processing sequence

The 2015 overview is unusually useful because it exposes part of the model’s processing architecture. Figure 1 describes a core FORTRAN model that loops annually, along with a SAS postprocessor that handles much of the benefit-calculation logic (Favreault et al. 2015).

Public documentation also makes clear that DYNASIM is not just one monolithic equation block. It is a staged simulation with modules that feed one another. The public processing sequence includes:

  • demographic updates such as birth, death, migration, marriage, divorce, schooling, and leaving home
  • labor-market outcomes such as employment, hours, wages, job change, pension coverage, and retirement
  • health and disability states including general health, ADLs, IADLs, chronic conditions, cognitive impairment, and work limitations
  • program calculators for Social Security, SSI, disability insurance, taxes, Medicaid, and other payers
  • wealth, pension, and insurance updates that feed back into later eligibility and spending outcomes (Favreault et al. 2015)

For an outside researcher, this is enough to understand DYNASIM’s conceptual architecture, but not enough to recreate the current model faithfully.

6. Data sources visible in public documentation

Urban’s public materials are also fairly informative about the data sources behind DYNASIM. The 2024 fact sheet says DYNASIM4 uses models estimated on datasets such as the NLSY, PSID, HRS, and SCF, as well as SIPP linked to administrative data (Urban Institute 2024).

The longer overview describes those sources in more operational detail. Publicly documented inputs include:

This matters for transparency. DYNASIM is not based only on public microdata. Public documentation itself says some model components are estimated using survey files linked to administrative records (Urban Institute 2024). That is one reason an open-source project should not assume it can reproduce DYNASIM simply by copying the public descriptions.

7. What DYNASIM models

Urban’s public materials make DYNASIM’s scope unusually visible. Between the overview and the 2024 fact sheet, the following categories are publicly documented:

  • demographics and family structure
  • education
  • employment, hours, wages, sector, and job changes
  • pension coverage and benefit accrual
  • Social Security, SSI, and disability insurance
  • health insurance coverage
  • medical spending
  • LTSS use, costs, and payer allocation
  • taxes
  • wealth, including home equity, financial wealth, pension wealth, and Social Security wealth (Favreault et al. 2015; Urban Institute 2024)

Two design details are especially important for interpreting DYNASIM’s policy uses:

  1. Public documentation describes Social Security and SSI as strongly rule-based components.
  2. The 2024 fact sheet says the simulation of other government benefits is based solely on statistical modeling rather than explicit eligibility rules (Urban Institute 2024).

That means DYNASIM is not best understood as a universal rules engine. It is a dynamic microsimulation model with some highly detailed calculators and some more reduced-form program modules.

8. Alignment and calibration

Public documentation repeatedly emphasizes that DYNASIM is not a purely free-running simulation. The 2024 fact sheet explicitly says that some outcomes such as fertility, mortality, disability, immigration, and labor force participation are aligned to targets produced by SSA’s Office of the Chief Actuary (Urban Institute 2024).

That point should shape how this repository talks about DYNASIM. Urban’s benchmark model is not valuable because it avoids alignment. It is valuable because it combines rich micro-level state transitions with external anchoring to trusted aggregate forecasts. Any competing public model will need to be equally explicit about where it aligns and where it lets the microsimulation run on its own.

9. LTSS-specific machinery

For long-term care modeling, the public record is strong enough to be substantive. The 2015 overview describes a real LTSS module, not a placeholder (Favreault et al. 2015).

Care settings and transition logic

Public documentation says DYNASIM models:

  • nursing home use
  • residential care or assisted living
  • paid home care

Those equations are estimated on pooled HRS data for respondents ages 65 and older. Predictors include ADL and IADL limitations, self-reported health, marital status, spouse disability, race and Hispanic origin, number of children, age, sex, nativity, income, wealth, and prior care use. Any nursing home care, residential care, and paid home care are jointly estimated as a trivariate probit with persistence built in through lags and correlated errors (Favreault et al. 2015).

Intensity and duration

Public documentation also says DYNASIM projects:

  • number of nursing home nights
  • duration of paid home care
  • hours of paid home care

The overview describes zero-truncated negative binomial models for nursing home nights and home-care hours, plus NHATS-based adjustments to convert HRS monthly home-care measures into annual quantities (Favreault et al. 2015).

Prices, payers, and Medicaid

The public LTSS documentation is also unusually concrete about costing:

  • private-pay LTSS prices use state-specific Genworth data
  • Medicaid LTSS prices rely on published reimbursement summaries
  • future LTSS prices are wage indexed
  • payer allocation distinguishes out-of-pocket spending, Medicare, Medicaid, insurers, and uncompensated or other public care (Favreault et al. 2015)

The overview further states that DYNASIM reflects a composite set of state-specific Medicaid eligibility rules, compares those rules against income and assets, and uses a relatively simple spenddown equation to link LTSS expenses to Medicaid entry (Favreault et al. 2015).

This is enough to say that DYNASIM already covers several pieces that an LTC-focused public model would need:

  • care need and care setting
  • service intensity
  • payer assignment
  • Medicaid interaction
  • private long-term care insurance

But the public record also reveals some limitations:

  • the 2015 LTSS description focuses on ages 65 and older
  • home-care measurement requires an annualization workaround because HRS observes paid care only in the prior month
  • wealth spenddown is described publicly as relatively simple, not as a highly detailed state-specific depletion model (Favreault et al. 2015)

10. Demonstrated policy uses

DYNASIM’s importance is not only methodological. Urban has used it in published policy analysis across multiple domains. Public sources show applications in:

This matters because it distinguishes DYNASIM from a merely notional research model. It is a live production research platform with a long publication trail.

11. Operational clues from public materials

Urban’s 2024 fact sheet gives a rare public glimpse into how DYNASIM is used in practice. It distinguishes between:

  • a baseline under current law and forecast assumptions
  • a counterfactual under alternative policy or behavioral assumptions

It also says that some baseline-only studies can be completed in days, whereas new policy scenarios can require weeks of involvement from the DYNASIM team depending on the complexity of the provisions (Urban Institute 2024).

That is an important clue. Publicly, DYNASIM looks less like a downloadable package and more like an internal expert-operated research platform.

12. What remains opaque

Even after reading the public record closely, several important things remain unknown or only partially known:

  • the full current source code
  • the exact version history for each live module
  • the current parameter values and full equation inventory
  • the testing and quality-assurance workflow
  • the exact operational representation of state Medicaid rules
  • how run-specific choices differ across Urban publications
  • which components rely on restricted linked data in ways an outside team cannot reproduce exactly

The DYNASIM3 primer explicitly said that more detailed documentation was available on request from the authors rather than fully public (Favreault and Smith 2004). That is useful context. DYNASIM has been publicly described for years, but it has not been publicly exposed as an inspectable end-to-end software artifact.

13. Implications for this repository

The public record suggests five practical conclusions for this project:

  1. DYNASIM is a serious benchmark for breadth. Any proposal that talks about retirement, disability, wealth, or LTC as though this terrain is empty will look underinformed.
  2. DYNASIM is not an open reference implementation. The gap is not just “another model”; it is transparency, inspectability, and the ability for outside researchers to reproduce the full pipeline.
  3. For Social Security, the real comparison point is not only benefit calculation. It is lifecycle data construction, alignment, and validation discipline.
  4. For LTC, DYNASIM already appears to cover much of the essential national-model machinery. A public entrant should differentiate on reproducibility, modular policy rules, and better visibility into intermediate validation targets.
  5. The right ambition is not to dismiss DYNASIM. It is to document, as precisely as possible, what DYNASIM already does and then build the parts that remain inaccessible to the field.

Bottom Line

Publicly, DYNASIM is knowable enough to take seriously. Urban has disclosed its broad architecture, major data sources, alignment logic, LTSS machinery, and several classes of policy application. What remains inaccessible is exactly what matters for true reproducibility: the live code, the full equation system, and the operational workflow.

That is why this repository should treat DYNASIM as both a benchmark and a boundary marker. It shows how far a non-governmental model can go. It also shows how much remains unavailable without an explicitly open alternative.