The Development and Use of Models for Fiscal Policy Analysis Alan Auerbach September 23, 2016
Outline Types of models for fiscal policy analysis Different purposes for model use: implications Who should develop and maintain models? Particular issues to be resolved in model development and reporting of results Looking ahead
Types of Models Microsimulation models General equilibrium models Traditional large scale macroeconomic models Reduced form models
Microsimulation Models (1) Based on a large population of households Could be actual households (e.g., NBER Taxsim or TPC Tax Model) Could also use grouped data, for example by age, or income group Even when using actual data, may be for just one year, requiring adjustment for other years Sometimes utilized for firms, but typically with little connection to ultimate owners
Microsimulation Models (2) Typically includes detailed tax and/or benefit calculator, to accommodate changes in both broad and very specific provisions May also be linked to other microdata, to expand scope for analysis E.g., use consumption surveys to evaluate changes in taxes on particular commodities Direct linkage rarely possible, so need to use some matching procedure
Microsimulation Models (3) Most commonly analyzed without behavioral responses, although such responses may be implicit in incidence calculations (e.g., who bears a change in labor income tax?) Sometimes incorporates particular aspects of behavioral response, e.g., change in taxable income, change in demand for a taxed product, change in labor supply and demand But full range of responses typically not considered
Microsimulation Models (4) Most commonly evaluated on a cross-section basis, limiting ability to undertake analysis where life-cycle elements may be important
General Equilibrium Models (1) By definition, considers changes in equilibrium of economy taking behavioral responses into account But GE models generally also derive behavior from underlying optimization by households and firms Use of optimizing behavior allows direct measurement of welfare effects of policy changes Some GE models don t incorporate optimization (e.g., Solow growth model used by CBO)
General Equilibrium Models (2) Because of greater complexity, GE models typically are based on much smaller number of distinct, representative households (e.g., for different income groups or ages) Because behavioral responses are being accounted for, GE models typically also account for much less detail regarding tax and benefit systems
General Equilibrium Models (3) GE models can cover one year, many years, or the infinite horizon Infinite horizon models include representative agent and life-cycle models With life-cycle models, can look at effects of policy on different cohorts Dynamic models can incorporate government s intertemporal budget constraint (GIBC) and similar constraint on international accounts
General Equilibrium Models (4) GE models can incorporate short-run economic fluctuations and scope for monetary and fiscal policy, through explicit specification of rigidities (e.g., menu costs, delays in wage adjustments, etc.)
Large Scale Macro Models (1) Traditional models used for short-run macroeconomic forecasting, reaching intellectual peak in 1960s & 70s (e.g., MPS model, Wharton model) Large number of equations used to solve for equilibrium over time Considerable detail regarding government sector, including taxes and benefits
Large Scale Macro Models (2) Equations motivated by optimizing behavior, but not fully consistent Allows inclusion of traditional Keynesian features Subject to Lucas critique, which largely arrested further academic development But still used by private forecasters, governments and central banks
Reduced Form Models For pure macroeconomic forecasting purposes, small number of equations that incorporate few economic restrictions Some constraints may be imposed on equations, based on economic theory (e.g., SVAR) Benefit of transparency relative to, say, largescale macro models
Illustration: Evaluating 2009 U.S. Stimulus Package (1) In early 2009, US considered a bill with large tax cuts and spending increases. What type of model to use to evaluate proposal? Microsimulation model: could accommodate provisions, but not macro responses Dynamic GE model: not enough detail; no realistic short-run macro responses SVAR: not enough detail Traditional macro model: used
Illustration: Evaluating 2009 U.S. Stimulus Package (2) Implications: No welfare analysis, or even distributional analysis Use of complex model (rather than simple forecasting model like VAR) leads to concerns about need for judgment to arrive at forecasts Similar focus on short-run macro responses and detailed provisions means that such models remain broadly used E.g., OBR-Treasury model*
Different Objectives, Models (1) As these examples illustrate, what model to use depends on objectives For short-run macro analysis of detailed provisions, only traditional models offer necessary features But for long-run welfare analysis, a dynamic CGE model would be more appropriate Can models be integrated?
Different Objectives, Models (2) Illustration: CBO Dynamic Analysis* Current approach: Short run: traditional macro model Long run: Solow-type growth model or life-cycle growth model Question: how to integrate results between short- and long-runs?
Different Objectives, Models (3) Even within model type, e.g., microsimulation, different objectives require different models Example 1: evaluating changes in capital gains or estate/inheritance taxation Virtually all taxes paid by those at the top of the income distribution need concentrated sample, not full sample of taxpayers or households Intergenerational linkages may be needed
Different Objectives, Models (4) Even within model type, e.g., microsimulation, different objectives require different models Example 2: evaluating changes in public pension system s taxes and benefits Can t evaluate incidence or behavioral responses without looking simultaneously at tax and benefit impacts for same individual Need life-cycle model with long horizon to do this, requiring projections for existing households
Who Should Develop & Maintain Models? To tie to latest research findings, want some academic input But need for ongoing development and maintenance for policy analysis likely requires government or research/policy organizations, given incentives for academics Challenge: how to maintain credibility of analysis?
Maintaining Credibility (1) 1. Use independent advisory boards Example: CBO panel of economic advisers* For governments, auditing function may be served by fiscal councils, which may also be responsible for modeling decisions Who appoints the experts?
Maintaining Credibility (2) 2. Present results from different models Example: US Joint Committee on Taxation (JCT) uses different models simultaneously for dynamic analysis, to satisfy different views of the appropriate model* How should one interpret potentially very different results from different models?
Source: https://www.jct.gov/publications.html?func=startdown&id=4564
Source: https://www.jct.gov/publications.html?func=startdown&id=4564
Maintaining Credibility (3) 3. Use open source models Example: AEI Open Source Policy Center* Makes analysis transparent, so that individual researchers can see what is driving final results Allows other analysts to vary parameters, to produce results consistent with alternative views Limits ability to make off model adjustments if results don t make sense How to deal with risks of reverse engineering?
Particular Issues to be Resolved
The Government Budget (1) In dynamic models with proper specification of government policy, the government s intertemporal budget constraint (GIBC) must hold: PV(Taxes) = PV(Expenditures) + Initial Debt But even if underlying policy satisfies the GIBC, policy changes generally won t. What to do about this, i.e., how to close the model?
The Government Budget (2) 1. Consider and present different alternatives for closing model (Auerbach, N. Tax J. 2002*)
Source: www.ntanet.org/ntj/55/3/ntj-v55n03p387-407-bush-tax-cut-national.pdf
Source: www.ntanet.org/ntj/55/3/ntj-v55n03p387-407-bush-tax-cut-national.pdf
The Government Budget (3) 1. Consider and present different alternatives for closing model (Auerbach, N. Tax J. 2002*) But which alternative is likely? 2. Use estimated fiscal feedback rules (e.g., taxes and spending respond to debt) But which taxes and spending? Also, if done by government agency, an issue of assuming actions not yet taken by government
Filling the Knowledge Gaps (1) In model that depends on behavioral parameters, what to do for parameters where estimates are scant or missing? Example 1: Suppose have estimates of the taxable income elasticity (TIE) sometimes a sufficient statistic but want GDP impact of policy change (for which responses on different margins have different effects)
Filling the Knowledge Gaps (2) In model that depends on behavioral parameters, what to do for parameters where estimates are scant or missing? Example 2: How does broadening income tax base affect labor supply? Often ignored in models, and in estimation of labor supply responses, but theory says it should have an impact
Filling the Knowledge Gaps (3) In model that depends on behavioral parameters, what to do for parameters where estimates are scant or missing? Typical solution for missing parameters target aggregate levels or moments, assuming some common response But hard to do with many missing parameters, and choices may have important effects on estimated policy impacts
Out-of-Range Model Use What to do when policy experiments go far outside historical experience? Example 1: Increase marginal income tax rate to 95% Example 2: Replace a territorial corporate income tax with a destination-based one Model may be set up to provide estimates, but how plausible are they?
Accounting for Uncertainty (At least) three elements of uncertainty: Parameter uncertainty Model uncertainty Economic uncertainty Parameter uncertainty in principle the most straightforward to deal with report implied confidence bounds for estimates, based on standard errors. But
Parameter Uncertainty Parameter estimates typically drawn from different sources what is their covariance? Some parameters are assumed or calibrated how to incorporate uncertainty for them? For some purposes (e.g., budget scoring) a single number is required Should point estimate be used when deviations have asymmetric social costs, or should it be adjusted to account for this asymmetry?
Model Uncertainty As discussed, different models sometimes used to represent a range of results In principle, could choose among models using some out-of-sample accuracy criterion (e.g., RMSE) But some models may be better for forecasting some elements of behavior and worse for others
Economic Uncertainty Models are often deterministic, but this leaves inconsistencies given the presence of uncertainty Example 1: What should the model use for the rate of return, when the actual economy has several rates, depending on risk? Example 2: How to incorporate uncertainty about policy, where first and second moments should matter?
Ensuring Internal Consistency (1) One cause of susceptibility to reverse engineering is internal inconsistency Example 1: US tax-preferred saving accounts traditional (EET) vs. Roth (TEE) Largely equivalent, but models of annual distributional incidence and budget cost show vastly different results (which has distorted policy)
Ensuring Internal Consistency (2) One cause of susceptibility to reverse engineering is internal inconsistency Example 2: Two-part consumption tax (e.g., business cash-flow tax plus labor income tax) vs. traditional consumption tax (e.g., VAT) Largely equivalent, but models typically treat business income taxes differently and would assign different incidence and behavioral effects
Looking Ahead (1) Two major advances that should aid in moving process forward Big data access to more data and faster computation Data access increasing ability of economists outside official government agencies to access administrative data, including tax returns Will aid in our ability to deal with many of the problems mentioned here
Looking Ahead (2) But challenges will remain, many relating to the consistent application of economic theory and technique in a very public environment where politics is central and many groups have much at stake in the policy process