New data excite me because I might be able to use them in a research paper. Because they are valuable to me, I expect them to also be valuable to society. So I keep attempting papers on whether better information reduces illegal fishing. But I keep getting null results! A recent paper on illegal gold mining in Colombia, however, found that better information did reduce extraction. What made that paper succeed where mine have failed?
Failing
To see what has to go right, take my fishing example. Suppose new data become available on where vessels fish. This new information might help the regulator better detect fishing in places where it is prohibited, like “no-take” Marine Protected Areas (MPAs). The information could reduce illegal fishing inside no-take MPAs if the regulator is now more able to punish rule-breakers after the fact, or even send a patrol boat to stop them during the act. If the benefits of lower illegal fishing exceed the costs of producing or using the new data, then the information is valuable to society.
All three steps in this simple story must be satisfied for the information to have value. First, the regulator has to receive the information. There are so many reasons this might not happen! The relevant officials might not be aware of the new data. Or the new data might be difficult to access. “Fisheries transparency”—making data on fishing activity publicly available—aims in part to address this accessibility barrier. Even if accessible, however, officials might choose to ignore it if they don’t believe the data add information to what they already know. This uncertainty—not knowing what information officials received, much less whether data that seem novel to me actually tell them anything new—is a persistent challenge when working with natural experiments and secondary data.
Second, the regulator needs to act differently because of the information. Better data could make enforcement more effective by deterring vessels from fishing illegally in the first place (if they now believe they are more likely to be punished), or by enabling regulators to stop violations as they occur. Since each unit of enforcement effort now reduces illegal fishing more, regulators might also increase total enforcement effort. Either way, regulatory behavior must change. If, for example, there is no fuel for patrol vessels, then better information is unlikely to reduce illegal fishing. Here again, I typically don’t observe regulators’ actions, further increasing the difficulty of diagnosing my null results on illegal fishing.
Third, benefits must exceed costs. Reaching this step is encouraging—it means the information has already changed outcomes. Officials received the data, changed their enforcement actions, and this reduced illegal fishing. The remaining question is whether the benefits of lower illegal fishing exceed the costs of better information, such as producing the data, making the data accessible, accessing the data, and any additional enforcement resources deployed.
Succeeding
“Technology, Information and State Capacity: Experimental Evidence from Illegal Mining in Colombia” by Santiago Saavedra satisfies and measures each step of the causal chain, finding a significant reduction in illegal mining and showing why it occurred. The experiment randomly assigns municipalities to receive machine learning predictions of gold mining locations shared only with local authorities, only with national authorities, or with both, compared to a control group that receives nothing.
Saavedra knows officials received the information because he sent it directly. Cleverly, he doesn’t just send it—he asks officials to confirm whether the predicted locations contain actual mining operations. Requesting a response makes it more likely that officials open, read, and engage with the information.
The paper also documents changes in regulatory behavior with a survey of mayors and national police administrative data. Enforcement increases because of the intervention. But a revealing pattern emerges from the surveys: when Saavedra’s independent verification confirms a predicted site contains an illegal mine, local authorities are significantly less likely than national authorities to acknowledge it.
This differential response suggests local authorities, who serve as the first line of defense against illegal mining, may be more prone to capture than national authorities. Among this group, the intervention design permits separating whether the illegal mining reduction came from the specific mining locations Saavedra shared, or simply from local authorities learning that a monitoring technology exists—which could reveal their failure to shut down illegal mines. Two patterns point to the monitoring technology threat as the key mechanism. First, more accurate predictions don’t reduce illegal mining significantly more—suggesting the specific locations matter less than awareness of monitoring itself. Second, local officials’ misclassification rates don’t increase for sites farther from the municipal office, which we’d expect if distance simply made verification harder. The threat of transparency—that monitoring could expose inaction or collusion—appears to drive the reduction in illegal mining.
Third, Saavedra demonstrates that benefits exceed costs. The machine learning predictions are cheap to produce and share, while illegal mining imposes large pollution costs. In his conclusion, Saavedra notes this technology could be deployed in other countries and for other natural resources—but cautions that improvements in outcomes depend on whether authorities are interested and able to act on the information.
This caveat resonates with my null results in non-experimental fisheries settings. Officials must receive the information, act differently because of it, and those actions must successfully reduce illegal behavior. The Colombia mining experiment shows all three steps working, explaining both its success and the difficulty of replicating it. Information has value only when it travels the full distance from production to improved outcomes.
Join the Conversation