How a Small Think Tank’s Mathematical Models Are Shaping Modern Politics

The Penn Wharton Budget Model was formed only five years ago, and you’ve already made a pretty impressive footprint on policy. What did the organization hope to achieve when it set out?

We’ve been able to make a big impact quickly with rigorous, integrated modeling. We’re really trying to fill a gap in serious analytics. What makes us attractive is that we don’t advocate, and we don’t have an opinion. We let the chips fall where they may. 

At times the White House might love our numbers and cite them, and at other times they lambaste them, like they did with our analysis of their infrastructure and tax reform policies. Sometimes the Speaker of the House cites our stuff, and other times it gets ignored. We don’t try to be strategic in that sense. We think that intellectual honesty is a long-run game, not a short-run game. We also try to play well with the scoring agencies — the Joint Committee on Taxation and the Congressional Budget Office — and we don’t view them as competition.

You share the analysis you’re doing with policymakers so they can make informed decisions. For example, you ran an analysis of Vice President Joe Biden’s presidential campaign platform. How are public officials using your modeling to inform their plans?

Policymakers often come to us before they write bills. It’s very clear when our footprint is on those bills, because we give feedback — usually off the record — about what the impacts would be if they try to achieve something one way versus another. During this presidential campaign, for example, six of the candidates came to us when designing their plans. 

We don’t try to influence outcomes or make a judgment whether it’s good or bad. What we do is provide policymakers with a sandbox that allows them to explore their ideas before they actually put pen to paper. It gives them the economic impact, the distributional impact, the budgetary impact — and it’s not biased through conventional scoring mechanisms such as a ten-year budget window. 

We did that in our analysis of Biden’s campaign platform, and you saw that there were definitely short-run costs, but by 2050, the economy was roughly neutral, and part of that is the returns to productivity that’s associated with some of his spending.

How does the Penn Wharton Budget Model use data to tell the story of how policy affects the economy and the federal budget?

It was fashionable for a while to say that big data can solve everything. Now, the pendulum has swung pretty hard the opposite way. We see that actually leads to a lot of cherry-picking. You need more than just data. You need data plus a conceptual framework that is subject to rigorous, honest testing that can’t be manipulated. So we try to bring the best of academic modeling using cutting-edge tools and actually applying them to real policies. 

The scoring agencies in D.C. really understand a lot of policies, but they often analyze them with spreadsheets. We’re combining the best of both. That helps us use data appropriately, and not in an abusive way.

Tell me more about the framework itself. What was the process for developing your model, and what kind of testing goes into it?

It’s what’s known as a micro-simulation framework, and it’s very detailed with incredible heterogeneity that allows us to see relationships over time. It starts with several hundred thousand representative households with over 60 different attributes. These people grow up, get education, get married, sometimes divorce, have kids, decide how to educate their kids. 

The framework is nice for creating rich detail and seeing the impact of budget. But by itself it’s not going to give what we call macroeconomic dynamics, incorporating the impact of these decisions on the capital stock and labor of the entire economy, GDP, things like that. So that micro-simulation framework feeds into what’s known as a heterogeneous overlapping generations framework, which allows us to capture, in incredible rigor, the macro impact on the economy, things like when households save more or less, when debt impacts the amount of capital in the economy, when a certain tax encourages more or less work, and how that impacts things like wages, interest rates, and retirement savings. What’s neat about the framework is that it’s very rigorous, and it’s not gameable in a way that more traditional frameworks are.

In combining those methods for analyzing data, how are you teasing out the ways that policy will actually impact real people’s lives? 

It’s really easy to have a model about averages. The problem is that a model that works well on average can leave a lot of people on each side of that average worse or better off, sometimes by a lot. For example, you could say that a policy either raises or lowers the annual or lifetime income of someone at the poverty line, and that sounds somewhat informative. You’re not just talking about averages — you’re talking about someone who’s at the poverty line. 

But even that’s not adequate, because the policy may look good, bad, or neutral on average for that person, but it also may incur certain risk-sharing properties. It might lower that person’s wage a little bit, but also give them a great insurance program that will buffet recessions and prop them up during times of need, which has great value over time. 

So one of the things we brought in from academics is what’s called ​“dynamic distributional analysis.” We go a step further and ask, what is the impact of a potential policy on providing risk reduction? What is the impact not just on this person this year, but on a lifetime basis? It’s much more holistic.

Now that you’ve been at this for five years, are there certain themes that you’ve seen emerge over time in your work?

I think that, by far, the biggest one is that integration of modeling matters tremendously. There are certain analysts who just take on one policy area — immigration, health, tax — and that’s all they focus on. The problem — and this is true in our Biden work — is that these things interact a lot. You don’t want to think about tax without immigration, because immigration actually plays a big role when we think about taxes. So integration is more and more important in our analyses. 

The second big theme is the importance of proper distributional analysis. Our view — and we saw this with the Tax Cuts and Jobs Act — is that advocates often push the envelope on the economic benefits of a policy, while those who are opposed push the envelope on the distributional impact. We believe it’s best to think about these things in an integrated way, not about a person this year but over a lifetime. Ignoring the macroeconomic effects of a policy is very limiting. 

And then the third theme is the value of bottom-up analysis versus top-down. So much of D.C. analysis starts with aggregate numbers. That’s where this focus on the averages comes from. With a top-down approach, you’re always trying to learn from the past, but you’re never learning the right things from the past. With the bottom-up approach, you capture the right things from the past, and you can use that to analyze brand new policies.

Why is it so important that you are nonpartisan — that you don’t have a viewpoint?

You have to trust the source you’re getting analysis from. We’re not being paid by people advocating for a smaller government or a larger government. We’re not being funded by people whose agenda influences us. We constantly update our own model, and we feel like we’re giving you the best of what economics has to say about policy. 

We do a lot of events on Capitol Hill. At least once a month, we have meetings with congressional staff, and we count how many Republicans come and how many Democrats come. It’s shocking that it’s almost an even split all the time. It’s because we’ve built that trust.

You’ve talked about some interesting impacts your model has had. Are there particular achievements you’re most proud of — ways you’ve seen policy direction shift as a result of your work?

It’s always hard to get in policymakers’ heads, but we do know that various bills have been written differently because of our feedback. We came out with a primer before the Tax Cuts and Jobs Act with different elements that have trade-offs, and we think some of that got in there. 

For the most recent CARES Act, we showed three different ways of sending out stimulus checks to households, and it appears they picked one option from that list — they changed the check size to hit a round number, but the phase-out amounts they used were exactly what we modeled. 

We were also the first to explain how much money would really be achieved from a wealth tax, and it turned out not to be as large as some people were hoping. More recently, we developed a very detailed model where we showed the impacts of COVID-19 on the Social Security trust fund finances, and that really forced a conversation in D.C., because the actuaries that originally came out seemed to ignore the COVID stuff. So we’ve been playing a role, not only in helping policymakers but also in helping the agencies really think through these trade-offs. We consider that a win.

You’re mentioning huge areas of policy you’ve had your hands in already. What’s in the future for the Penn Wharton Budget Model?

There are a couple of things, and, in fact, Arnold Ventures is supporting one of these. Medicaid is almost as big as Medicare now, and it’s going to get bigger. Long-term care programs in the U.S., for when people go into a nursing home, are paid for by Medicaid. That has always been tracked on a ten-year basis, whereas with Medicare, everybody knows we should go out at least 75 years. So we’re tracking Medicaid farther out in time. 

That’s an interesting one, because Medicaid is split between the federal government and the states, and the states will face tremendous fiscal pressure. We’re excited about that project. We’re also developing a new generation of modeling that is several thousand times more complex than our current model. It doesn’t have the same richness in certain areas, but it can handle really complex questions. 

Economists all agree that carbon taxes play a necessary role as a corrective, because we agree that the science is pretty darn clear on carbon. But modeling a carbon tax well is extremely challenging. We’re going to have a model that can do it with great rigor within a year or two, we hope. We don’t use math for the sake of using math, but with certain questions you just need it. We’ve been able to use mathematical theory to solve that problem, which is actually at the quantum level. That’s going to give us much deeper insights into issues that we think are of first-order importance.