High-throughput computing and the Open Science Grid (OSG) can be as critical to research in the social sciences as in other fields. In the last of our series of three interviews with University of Wisconsin-Madison (UW-Madison) social science researchers, we talk with Jesse Gregory, who has used the power of the OSG to revolutionize his work.
Jesse Gregory (photo courtesy of Jesse Gregory)
Gregory is an assistant professor of economics at UW-Madison, and he specializes in labor and public economics. He has been accessing the OSG via UW-Madison’s Center for High Throughput Computing for about one and a half years. In the past year, Gregory has taken advantage of more than 2 million CPU hours, with more than 500,000 hours on the OSG.
With a particular interest in disaster relief policy, Gregory has sought to address the question: When individuals struck by disaster choose where to live or whether to rebuild, in what way (and to what degree) do neighbors, government policy, and available grant funding influence their decisions? Gregory has been using the OSG to develop a model of New Orleans homeowners’ post-Hurricane Katrina rebuilding choices.
How does policy influence behavior? Gregory says there are two ways empirical economists address this kind of question. One way is to look for settings in which otherwise similar groups exist—some who were exposed to a given policy, and some who were not—and compare the outcomes of the two groups. Gregory points out that these kinds of treatment-versus-control comparisons are not possible when the groups being studied are exposed to the same policy. This often occurs with disaster relief or when a new policy has not been enacted yet. “In these cases, a common approach is to estimate an economic model of the underlying trade-offs that are being made,” said Gregory. If he can estimate the parameters of the model, then he can assess a policy’s impact on rebuilding choices. “We characterize the distribution of people’s preferences for returning to their prior home location, for having more of their neighbors rebuild, and for the consumption of other goods. In the specific case of New Orleans, we do this by simulating their rebuilding choices with any incentives and simulating choices without incentives.”
Estimating the preference parameters is computationally intensive because it requires simulating households’ choices under a large number of candidate parameter values. Gregory is looking for the parameter values that cause predicted rebuilding choices to most closely match households’ actual rebuilding choices. “The ability to perform many of these calculations at the same time with high-throughput computing such as the Open Science Grid makes the estimation feasible,” notes Gregory.
After severe disasters, Congress sometimes provides funding for an affected area that is in addition to standing disaster relief. Block grants may be given to state or municipal governments, who then have some control over how they want to use the money. After Katrina, Congress gave a block grant to Louisiana, which it used to fund “Road Home” grants to homeowners whose storm damages weren’t covered entirely by insurance.
“One feature of the program was that they gave a more generous grant package to households that chose to rebuild at the same location,” said Gregory. It may seem like a strange policy to encourage people to rebuild in an area that was just hit by a flood or devastated by a disaster. “One question I’ve studied, in collaboration with my colleague Chao Fu, is whether there is some sort of spillover effect on the surrounding neighborhood from one household rebuilding. In economics, one motivation for subsidizing a particular activity is if it provides benefits to someone other than the person who’s choosing whether to engage in the activity. Here, we’re asking if one household choosing to rebuild makes the surrounding area more attractive to neighbors, and if so, how the value of that positive spillover compares to the inefficiency of inducing some households to rebuild and live in a place that wouldn’t otherwise be their first choice.”
Gregory begins by addressing a narrow question: Does one household choosing to rebuild increase the probability that other neighbors rebuild? “Since there are many non-causal reasons why neighbors would make similar choices, we focused on a set of households who had damage estimates very close to a threshold in the Road Home grant formula above which grant offers jumped dramatically,” said Gregory.
Households just above the threshold were otherwise in very similar circumstances to households just below. “We found that households with damage estimates just above the threshold rebuilt at a rate about five percentage points higher than those just below the threshold,” added Gregory. “In turn, these households’ immediate neighbors, who were not directly affected by the more generous grant offer, were about two percentage points more likely to rebuild, suggesting a positive influence from the subsidized households’ rebuilding. The idea behind the estimation of the choice model that we use for our policy experiments is to determine what sorts of underlying preferences are consistent with these patterns in the data.”
Figures: The Louisiana Road Home grant program offered rebuilding grants to all Louisiana homeowners with home damages not entirely covered by insurance. The program used a grant formula that resulted in significantly larger grant offers when the estimated cost of repairing a home was more than 51% of the cost of replacing the home than when the estimated cost of repairing a home was less than 51% of the cost of replacing the home. The top figure shows that the likelihood of a household rebuilding in New Orleans within five years of Katrina was about five percentage points higher if the estimated cost of repairing their home was just above this threshold than if the estimated cost of repairing their home was just below that threshold. The bottom figure shows that the rebuilding rate among a household’s immediate neighbors was about two percentage points higher if the household’s repair cost estimate was just above this threshold than if the household’s repair cost estimate was just below this threshold, suggesting that one household rebuilding generates positive externalities for the surrounding area. Treating these sorts of raw patterns in the data as inputs, Gregory has used the OSG to estimate models of New Orleans homeowners’ rebuilding choices that can be used to evaluate the effects of disaster relief programs. (Images courtesy of Jesse Gregory)
Gregory sets up a computational model that includes parameters describing the distribution of households’ preferences for living in their previous location, for having more of their neighbors rebuild, and for the consumption of other goods. The estimation involves solving the model a very large number of times and comparing the model’s predictions under each parameterization to actual household choices. Having access to a large number of processors that can run at roughly the same time allows him to search many parameters at once. This has great advantages over modeling once on a PC, waiting to get the result, and then trying the next parameter value. “The results of each model solution don’t depend on one another, so using Condor [high-throughput computing software], I can simply try a very large number of parameter values all at the same time, compare the results at the end, and see the prediction that most closely matches the choices that we saw in the data. This is the perfect environment to work with.”
The policy implication of Gregory’s research is that the optimal design of post-disaster grant programs depends on the context. “Requiring that grant money be used to rebuild in the same location only makes sense in settings where rebuilding generates large positive externalities,” noted Gregory. “Otherwise requiring someone to rebuild in a disaster-prone location in order to get relief would seem like a crazy policy to enact.”
The OSG has helped Gregory show that on average there were large positive spillover effects of rebuilding as opposed to leaving a lot undeveloped in post-Katrina New Orleans. “One neighbor being induced by a rebuilding grant to rebuild significantly increased the probability that additional neighbors would also rebuild,” Gregory said. He was also able to identify particular settings in which that effect is the strongest. “We find that these positive spillover effects occur almost entirely in neighborhoods where a majority of households would have rebuilt even without subsidies. The policy implications seem clear,” he added. “In highly devastated areas, it is probably best to provide disaster relief with few strings attached. However, it can be sensible to explicitly encourage rebuilding in places where somewhat dense rebuilding is already likely to occur.”