Abstract: While they are rare, superspreading events (SSEs), wherein a few primary cases infect an extraordinarily large number of secondary cases, are recognized as a prominent determinant of aggregate infection rates. Existing stochastic SIR models incorporate SSEs by fitting distributions with thin tails, or finite variance, and therefore predicting almost deterministic epidemiological outcomes in large populations. This paper documents evidence from recent coronavirus outbreaks, including SARS, MERS, and COVID-19, that SSEs follow a power law distribution with fat tails, or infinite variance. We then extend an otherwise standard SIR model with estimated power law distributions, and show that idiosyncratic uncertainties in SSEs will lead to large aggregate uncertainties in infection dynamics, even with large populations. That is, the timing and magnitude of outbreaks will be unpredictable. While such uncertainties have social costs, we also find that they on average decrease the herd immunity thresholds and the cumulative infections because per-period infection rates have decreasing marginal effects. Our findings have implications for social distancing interventions: targeting SSEs reduce not only the average rate of infection but also its uncertainty. To understand this effect, and to improve inference of the average reproduction numbers under fat tails, estimating the tail distribution of SSEs is vital.
Publication Bias under Aggregation Frictions: From Communication Model to New Correction Method (Job Market Paper)
(R code of stem-based method, Data appendix)
Abstract: To make informed decisions, readers often need to aggregate the results of multiple publications. However, due to cognitive limitations, they may focus on signs of effects rather than magnitudes or precision. This paper presents a model about how researchers will communicate their results under such aggregation friction, and uses its implications to develop an improved bias correction method. First, in the model, publication bias will emerge even when research is communicated optimally for readers. In particular, there will be omission of some noisy null estimates and inflation of some marginally insignificant ones. Second, the model suggests publication selection process will not be parsimonious, and thus cannot be adequately corrected by commonly used bias correction methods. This paper presents evidence consistent with these implications, and proposes a new, non-parametric stem-based bias correction method that is robust to the selection process implied by both the communication model presented here, and other models proposed in meta-analysis literature.
Health Benefits of Replacing Kerosene Candles by Solar Lamps: Evidence from Uganda, 2017
Abstract: A randomized controlled trial in rural Uganda shows that there can be improvement in air-quality-related health such as headaches, chest pain, fever, and eye irritation if non-electrified households switch from kerosene to solar lamps. This five-month study worked with a sample of 155 schoolchildren. Those who received solar lamps reported having better overall air-quality-related health (0.25 standard deviation of baseline distribution, 6% lower probability of any symptoms), although there was no statistically significant change in self-reported cough symptoms or lung capacity, as measured by spirometer tests. The health improvements were concentrated in school exam periods, most likely because students who switched to studying under solar lamps were exposed to less indoor air pollution. While health benefits exist, a follow-up survey shows that poor maintenance and low adoption remain major challenges for scaling up this new technology.
Cognitive Inertia in Active Learning under Imperfect Recall
Abstract: How do human memory imperfections affect experimentation decisions? This paper studies an active learning model augmented with two features of human memory: imperfect recall of past actions and information, and persistence of prior belief about one's ability. The model shows that under this memory structure, the experimentation decision will exhibit cognitive inertia, or a bias towards prior beliefs relative to other signals received. Unlike in settings with perfect recall, stabilizing this information over time improves experimentation decisions by alleviating and asymptotically eliminating this inertia. The paper further argues that this inertia illustrates some cognitive aspects of psychological distress and that it can be mitigated through meditation.