Or, a “Forensic Analysis for the Heritage CDA results”
The Heritage Foundation Center for Data Analysis (CDA) simulation of the Ryan plan, on behalf of the House Committee, has come in for some criticism. Commentary has been provided by Paul Krugman, and perhaps most comprehensively by Macroeconomic Advisers. (My comments are here:   ). Yesterday, the Heritage Foundation CDA’s director, William Beach, posted a rebuttal to Krugman’s critique. While Big Picture posted an excellent rejoinder,
I want to deal with one particular aspect of Mr. Beach’s open letter. Consider this excerpt.
Claim # 2 – We crafted the Ryan plan results with the end in mind: While I can see how you may have forgotten the limited purposes of economic policy modeling (though it’s still shocking that someone of your stature could be so unmindful), it is simply bizarre that you argue that we designed the economic modeling of the Ryan plan to reach specific conclusions. Either you are intentionally lying about our work, or you are totally ignorant of the complex, widely used model we employed for this work and also failed to read the detailed description of what we did that is posted on the House Budget Committee web site along with our results (which you apparently did see).
We used the highly regarded U.S. Macroeconomic Model of IHS Global Insights [sic!!!], Inc. Perhaps this is a model you as a pundit “do not recognize, but most economists do. This model has been around in its various forms for nearly 50 years. It contains over a thousand equations and several thousand variables. The modeler’s ability to affect the mechanics of the model is very limited, and, given the fact that the Budget Committee gave us final inputs only a few days prior to publication of their budget, we only had time to make sure that this detailed model would solve with the enormous changes to public policy we had to introduce into it.
Then, as I mentioned, we published a detailed description of how we did this work. Even today, it is there for anyone to read, and I especially encourage you to do so. I don’t expect everyone to agree with our modeling of this plan, but I do expect the debate over our work to start from an understanding that our modeling is fully described in the methodological appendix to our publicly available report.
I’ll leave it Professor Krugman to rebut Mr. Beach’s claim that he did not recognize the IHS Global Insight model (I can’t find the reference using the NYT search command). I’ll address the non-responsiveness of Mr. Beach’s remarks to the general critique that the combined CDA plus IHS Global Insight results are implausible. In fact, I think Mr. Beach’s focus on the IHS GI model aims to distract attention away from the role played by Heritage’s in-house model, the CDA microsimulation model.
In other words, most people are not concerned with the IHS Global Insight model, which is as far as I can tell is utterly conventional in the world of large scale macroeconometric models. (Confirmation from Macroeconomic Advisers). Rather, the concern is focused on the nexus between the IHS Global Insight model and the CDA’s microsimulation model, and here Mr. Beach’s open letter is completely unresponsive.
There is some discussion in the documentation Heritage provided for the Ryan analysis, but in fact after some investigation I have found a more clear exposition in the 2006 documentation of how Heritage analyzed the Bush tax cuts in a dynamic fashion using their combined GI/CDA microsimulation model. After adjusting the IHS GI model to fit the CBO baseline:
Calibrating the Microsimulation Model. We next calibrate the microsimulation model of individual income tax returns to CBO’s baseline projections. The final CBO-like forecast provides income, price level, and some budgetary variables used in this calibration.
Primary Components of the Microsimulation Model. The microsimulation model consists of three primary components-the core base-year data, a federal income tax and payroll tax calculator, and an optimizing routine that ages (extrapolates) the core base-year data. The first component consists of individual tax return data and demographic data in the base year. The second component reads a data file and replicates the process of calculating individual income and payroll taxes in the base year and future years. The third component ages the base-year data to reflect projected changes in not only key demographic and economic aggregates but also the distribution of income.
Aging the core base-year data involves four major steps. In each, we target tax and non-tax variables in the microsimulation model.
Step 1.We use the CBO-like forecast to update all nominal income values on individual tax returns. We also update all targets for demographic variables.
Step 2.We sequentially target four broad measures of individual income by percentile class. Total income is divided into wages and salaries, business income, non-capital gains investment income, and income from other sources. It encompasses both gross income reported on individual tax returns (gross tax return income) and non-taxable income.
Step 3.We target more detailed measures of the components of gross tax return income. Most of the targets are for components of NIPA personal income, with some important exceptions. Those exceptions include small business corporation (S-Corporation) net income, taxable pension and annuity income, net capital gains, and gains from the sale of other assets.
The final CBO-like baseline forecast provides a number of NIPA measures of personal and business income. These include wage and salary income, investment income, proprietors’ income, other business income, transfer payments to persons, and corporate profits.
The difficulty in disentangling the two models is shown by the following:
Simulating the Economic and Budget Effects of a Change in Tax Policy. Calibrating a macroeconomic model of the U.S. economy and a microsimulation model of the federal individual income tax to a common baseline yields a consistent starting point for dynamic policy analysis. We apply an additional calibration process to ensure that final dynamic revenue estimates from the macroeconomic model are broadly consistent with revenue estimates from the microsimulation model.
We regularly calibrate both the Global Insight model and the microsimulation model to CBO’s baseline projections. We also regularly use the calibrated macroeconomic and microsimulation models to analyze a variety of tax proposals. Tax data in the microsimulation model can be used to provide a “stand-alone” revenue estimate. A revenue estimate from the microsimulation model can also be introduced into the GI model to generate a “first-round” dynamic estimate of a proposal’s economic and budget effects.
A fully dynamic tax policy simulation proceeds in three steps.
First, we use the microsimulation model to estimate the revenue effects of the proposed change in tax policy under baseline economic assumptions. The proposed tax policy can involve a change in current-law federal income tax rates, a change in the federal individual income tax base, or both. The microsimulation model is used to estimate the change in federal income tax revenues. It also produces estimates of marginal tax rates on three types of income-ordinary income, long-term capital gains realizations, and dividend income-under the proposed policy and current law.
Second, we use the Global Insight model to estimate the dynamic revenue effects of the same policy change. Estimated changes in federal tax revenues and marginal tax rates from the microsimulation model are used as inputs into a simulation with the GI model. The macroeconomic simulation produces an alternative to the CBO-like baseline forecast. That alternative (non-baseline) forecast includes the dynamic effects of the proposed policy on GDP, prices, interest rates, employment, and personal and corporate incomes, among other variables.
Third, we update the microsimulation model to reflect the dynamic effects of the proposed tax policy on individual and business incomes. This is done using procedures similar to those developed for baseline calibration. Thus, NIPA components of individual and business income along with price level variables and some NIPA budget variables from the alternative forecast are used to estimate target values for non-taxable income and gross tax return income on individual income tax returns. We use those targets to update individual and business incomes in the microsimulation model so that they are consistent with the Global Insight model’s alternative forecast for the components of NIPA personal income.
For major tax proposals, we typically continue to iterate between the microsimulation model and the Global Insight model. Thus, we use revenue estimates and marginal rates from the updated microsimulation model to adjust the alternative forecast from the GI model so that it better reflects the effects of the tax proposal. [Emphasis added — mdc]
I think the underlined portions of the text are key to understanding in part the revenue aspects of the Heritage CDA simulations. However, they cannot be the reason for the fantastical investment responses implied in the simulations. I repeat three figures on investment from this post below.
Figure 2: Real equipment investment (blue), baseline (dark blue) and simulated under Ryan plan (red), in bn Ch.05$. Source: BEA, Heritage CDA, Appendix 3.
Figure 3: Real nonresidential structures investment (blue), baseline (dark blue) and simulated under Ryan plan (red), in bn Ch.05$. Source: BEA, Heritage CDA, Appendix 3.
Figure 4: Real residential investment (blue), baseline (dark blue) and simulated under Ryan plan (red), in bn Ch.05$. Source: BEA, Heritage CDA, Appendix 3.
In order to rebut the critics, Mr. Beach will have to justify the implied supply elasticities, in light of the extant empirical literature.
By the way, I still think, as before, it is the add factors that do the trick on the investment side. In any case, until Mr. Beach addresses these concerns, as well as those (mundane things like Okun’s Law) raised by Macroeconomic Advisers, with some actual numbers, we will just have to guess at what is exactly going on in these simulations.
I know that some will have tired of my discussion of this topic. However, I think it is incumbent upon us in the policy analysis community to fully understand what Heritage CDA did in this simulation, which provides the intellectual basis for an important policy proposal, before moving on. I for one intend to keep at this until a fuller explanation (involving numbers) is forthcoming.
Postscript: If you want to see what passes for a defense of the Heritage CDA analysis, see here (written by a former “director of The Heritage Foundation’s Center for Media and Public Policy”). No numbers, so the innumerate need not fear. Just ad hominem attacks.