JMIR Publications

ABSTRACT

Background:

Leaders make critical decisions during public health emergencies. Although traditional surveillance systems regularly update data and aid planning, they are retrospective in nature and lead to reactionary measures. Forecasting and analysis of behavior-related data may complement traditional surveillance systems in developing timely infection prevention and control strategies.

Objective:

In this study, we assess the use of behavioral indicators, such as public interest on the risk of contracting SARS-CoV-2 and change in mobility, in building a vector autoregression model for forecasting the daily number of COVID-19 cases in the National Capital Region.

Methods:

The stationarity of each variable was tested using Augmented Dickey-Fuller test. We determined the lag length by combining the knowledge on the epidemiology of SARS-CoV-2 and the information criteria measures. We fitted a VAR(8) containing change in mobility and daily number of cases with a seasonal dummy variable for the day of the week to the data from 11 January to 10 August 2021. Another VAR(8) model that includes public interest as an additional variable was also fitted to the training data set, and its forecast accuracy from 11 to 18 August was compared with that of the smaller model. The models were recalibrated to determine their performance during the peak of cases due to B.1617.2 and the resurgence following the spread of B.1.1.529.

Results:

The forecasted number of cases obtained from the model that includes public interest exhibited a similar trend with the actual data, and has higher forecast accuracy for two periods of forecasting than the model that contains only change in mobility and number of COVID-19 cases (Model 1: 17.9% and 74.2% vs Model 2: 17.6% vs. 21.4%). The importance of providing attention to public interest as well as the shift in the interrelationship of the three variables were also suggested by the Granger causality test. During the period of 11 to 18 August, only change in mobility (p-value=.002) was found to improve the forecast of the daily number of cases. On the other hand, public interest was also found to Granger-cause the daily number of cases from 15 to 22 September (p-value=.001) and the statistically significant result was also found when forecasting the number of cases from 28 January to 4 February (p-value=.003).

Conclusions:

To the best of our knowledge, this is the first study that attempted to forecast the number of COVID-19 cases and explored the relationship of behavioral indicators with the number of COVID-19 cases in the Philippines using time series analysis. Further investigation of how the number of cases will respond to changes in mobility and public interest using real-life data may be warranted to improve the planning and implementation of prevention and response measures.