Google Trends is used in research and surveillance as a proxy for community infection incidence. Signals are difficult to validate, as most surveillance biases towards severe outcomes and certain demographics.
Using Winter COVID-19 Infection Study (WCIS) data in England, symptom prevalence is estimated via generalized additive model with multilevel-regression and poststratification. Symptom duration was estimated using interval censored time delay modelling, converting prevalence to incidence. Google Trends and WCIS incidence and growth rates were compared using cross-correlation.
Google Trends and WCIS agreement varied by symptom and age group. The national maximum growth rate cross-correlation for sore throat was 0.81, with 90% prediction intervals of [0.69, 0.90]. Google Trends growth rates generally lagged the WCIS growth rates across symptoms (cough: −5.0 days [−8.0, 0.0], fever: −3.0 days [−6.0, 1.0], loss of smell: −9.0 days [−13, −3.0], shortness of breath: −12 days [−16, −5.0], and sore throat: −4.0 days [−5.0, −2.0]).
This work shows Google Trends and community symptom incidence can align, although substantial variation between symptoms and age groups exists, underscoring utility in predicting other surveillance indicators.