Seth Stephens-Davidowitz on Google Trends and Coronavirus

Episode Summary

Executive Summary: The conversation explores how Seth Stevens-Davidowitz uses internet search and platform data to infer hidden human behavior, with a focus on COVID-19. He explains the promise and pitfalls of Google Flu-style models, argues that refined search signals can help detect symptoms, and discusses broader economic and social shifts visible in online data. He emphasizes that these tools are best for nowcasting and complement, rather than replace, expert public-health judgment.

Main Topics: Why internet data became Seth Stevens-Davidowitz’s focus (Priority: 5/5): Seth describes moving from traditional economics toward studying what people really think and do, using Google Trends and other digital traces to analyze behavior across sociology, politics, health, and consumer activity. What different internet datasets reveal (Priority: 5/5): He contrasts Google search data, Wikipedia, Stormfront, and Facebook data, showing each can illuminate different phenomena such as notability, racism, fandom, or disease, but none is universally sufficient. Google Flu, its failure, and the case for a second-generation model (Priority: 5/5): The discussion revisits Google Flu Trends, why it initially worked, how H1N1 and media curiosity caused failure, and why more nuanced modeling of specific searches and changing patterns may make the approach useful again. COVID-19 symptom detection through search behavior (Priority: 5/5): Seth explains how searches like 'I can’t smell' and related symptom terms may serve as real-time indicators of COVID-19 spread, especially where testing is limited, while warning about news-driven noise. Risks of spurious correlations and search-data hype (Priority: 4/5): The hosts discuss the danger that search trends can reflect curiosity rather than illness, or produce false correlations, making careful statistical validation essential before drawing conclusions. Economic and behavioral changes during the pandemic (Priority: 4/5): Seth notes that search and purchase data mostly show expected pandemic effects—teleconferencing up, sports tickets down, toilet paper and home exercise equipment up—while some outcomes like pregnancy and divorce searches are less intuitive. Public policy, surveillance, and future uses of digital data (Priority: 4/5): He argues that digital data can help policymakers track conditions in near real time, especially where official statistics lag or testing is weak, but these tools are complementary and still evolving.

Key Arguments: Google search data is valuable because it captures what people are thinking or experiencing even when they do not say it publicly. Different digital data sources answer different questions: Wikipedia for notability, Stormfront for extremist motivations, Facebook for lifetime fandom patterns, and Google for nearly any topic. Google Flu failed partly because it could not distinguish actual flu from news-driven curiosity during H1N1. A stronger second-generation flu model should use more precise symptom searches and adapt as search behavior changes over time. Searches like 'I can’t smell' may be especially informative for COVID-19 because they are less common in other illnesses and less likely to be driven by general curiosity. Eye pain and burning eyes may also correlate with COVID-19 outbreaks, though the evidence is weaker and needs more validation. The biggest value of search data is often nowcasting—estimating what is happening right now—rather than forecasting far into the future. Internet data can help identify hidden secondary effects of shutdowns, not just the obvious changes in consumption and behavior. Official public-health models may not yet incorporate enough digital or high-frequency data, leaving room for broader data use in surveillance. Search data can be distorted by media attention and awareness campaigns, so any model must account for these confounders.

Data Points: Date of recording: Friday, April 10th, 2020 - The episode was recorded early in the COVID-19 pandemic. Google Flu lag: 1-2 weeks - The CDC took about a week or two to collate flu data before reporting outbreaks. H1N1 effect on Google Flu: Extraordinary rise in predicted flu without corresponding actual flu rise - Search interest increased because of curiosity/fear rather than actual illness. State-level symptom signal: New York, New Jersey, Louisiana, Michigan - These states repeatedly appeared in search signals for COVID-related symptoms. Search term example: "I can't smell" - A highly correlated search term for COVID-19 spread in Google Trends over the prior seven days. Search term example: "burning eyes" rose sixfold - Observed in Italy in March as a possible COVID-related symptom signal. Search term example: "eye pain" topic rose 4-5x - Observed in Spain and also in Iran as a possible symptom-related search signal. Seasonal behavior example: 8 years old - People whose teams won championships around age eight often became lifelong fans. Behavioral search impact: Teleconferencing up 200% - A large, expected economic shift during pandemic shutdowns. Behavioral search impact: Sports tickets down 80% - A large, expected economic shift during pandemic shutdowns. Purchase/search category example: Toilet paper up; home exercise equipment up - Examples of obvious consumer responses during lockdown. Pregnancy searches: Big drop - Seasonally adjusted searches for pregnancy tests suggested a possible baby bust rather than baby boom. Divorce searches: Down - Searches for divorce declined unexpectedly during lockdown. Anti-China sentiment searches: Moderate rise, not explosive - Searches insulting Chinese people increased, but less dramatically than in some previous periods of racial animus.

Pivotal Quotes: ""Google Trends is... a tool that showed what people are searching in different parts of the world at different times."" — Seth Stevens-Davidowitz: He explains why Google search data attracted him as a researcher. ""The next Google Flu 2.0 is going to be effective, but it's going to be a little more subtle than some of the earlier models."" — Seth Stevens-Davidowitz: He argues that refined symptom-based modeling can work if it adapts to changing behavior. ""We're not really hurting anybody by noting the correlations and telling people to explore."" — Seth Stevens-Davidowitz: He defends exploratory symptom analysis despite uncertainty and possible false positives.

Implications: Digital traces can improve real-time public-health and economic monitoring, especially where official data lag or testing is weak. But they require careful controls for media noise, changing behavior, and false correlations.