Hey Google! will New Zealand vote to legalise cannabis? Using Google Trends data to predict the outcome of the 2020 New Zealand cannabis referendum

• Published in: The International journal of drug policy Vol. 90; p. 103083
• Main Authors: Raubenheimer, Jacques Eugene, Riordan, Benjamin C., Merrill, Jennifer E., Winter, Taylor, Ward, Rose Marie, Scarf, Damian, Buckley, Nicholas A.
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.04.2021; Elsevier Science Ltd
• Subjects: Aggregate data; Cannabis; Cannabis legalisation; Decriminalization; Drug legalization; Drug policy; Google Trends; Google Trends extended for health; Internet; Legalization; Marijuana; New Zealand; Polls & surveys; Predictions; Public opinion surveys; Referendum; Referendums; Sampling; Time; Trends; Voting; New Zealand; Google Trends; New Zealand; Cannabis; Cannabis legalisation; Google Trends extended for health; Referendum
• ISSN: 0955-3959; 1873-4758; 1873-4758
• DOI: 10.1016/j.drugpo.2020.103083

New Zealand held a referendum on the legalisation of recreational cannabis in October 2020. Polls preceding the referendum provided contrasting outcomes. We investigated whether internet search data from Google Trends could provide an alternative estimate of the referendum outcome. We assessed various methods for accessing Google Trends data, downloading search probability data for google.com searches from New Zealand via trends.google.com, PyTrends and Google Trends Extended for Health. We used daily data for the three months prior to the final referendum date, and hourly data for the final week. We defined two smaller time frames each from daily and hourly data, allowing comparisons over the entire time frames, and progressively closer to the end. Using the selected keyword combination of ‘cannabis referendum yes/no’ we calculated the proportions of ‘yes’ and ‘no’ searches for each time frame/data source combination, aiming for a prediction within 2% of the final result. Data from different sources varied slightly. The method used to aggregate search probabilities over the selected time frame (mean/median) resulted in changes in the predicted outcome for hourly-, but not daily data. On 20 October we predicted the ‘no’ vote at 51.9%–55.4% for daily-, and 60% for hourly data when aggregated using the median, but only 49% for mean hourly data. Hourly data performed poorly at predicting the final 51.2% ‘no’ result, while predictions based on mean daily data for the full voting period provided the best prediction, differing by 0.1–0.2%. Predictions based on Google Trends data broadly agreed with polling predictions, but the exact method used affected the eventual prediction. While polls are subject to influence from methodological considerations (e.g., sampling), it is clear that Google Trends data can be used to make a prediction, but do not present a magic bullet solution to polling problems.