Can a supranational medicines agency restore trust after vaccine suspensions? The case of Vaxzevria

Abstract

Over the first half of March 2021, the majority of European governments suspended Astrazeneca's Vaxzevria vaccine as a precaution following media reports of rare blood clots. We analyse the impact of the European Medicines Agency's (EMA) March 18th statement assuring the public of the safety of Vaxzevria and the immediate reinstatement of the vaccine by most countries on respondents' intention to get vaccinated. By relying on survey data collected in Luxembourg and neighbouring areas between early March and mid-April, we observe that the willingness to be vaccinated was severely declining in the days preceding the EMA statement. We implement a regression discontinuity design exploiting the time at which respondents completed the survey and find that the vaccine reinstatement substantially restored vaccination intentions.

Fig 1. The impact of the EMA statement on the willingness to be vaccinated.

Notes: These graphs show RDD plots for the dependent binary variable equal to 1 if the individual is willing to be vaccinated. Graph (a) is obtained using a local linear polynomial regression with triangular weights and bandwidth following the optimal mean squared error criterion in [16]. Graph (b) is obtained using a global -quartic polynomial and uniform weights.

Fig 2. Google Trends searches for “Astrazeneca” and “blood clot” in the Greater Region of Luxembourg.

The vertical dashed line indicates the day of the EMA statement. Notes: Google Trends index for searches for “Astrazeneca” and “blood clot” in Luxembourg, Wallonia (Belgium), Saarland and Rhineland-Palatinate (Germany) and Lorraine (France) weighted by regional population size. Moving average over two days.

Fig 3. Density of survey responses over time.

Notes: Manipulation test using the local polynomial density estimators proposed in [17, 18]. Stata command rddensity. A local quadratic approximation with kernel triangular weights is used to construct the density estimators, while a cubic approximation is used for the bias-corrected density estimator. The density estimation method is unrestricted (two-sample). Robust bias-corrected statistic with jackknife standard errors and uniform confidence interval at 95% level (2000 of simulations).

Fig 4. Placebo tests.

Notes: Placebo tests point estimates and confidence intervals. The dependent binary variable is equal to 1 if the individual answers that they are willing to get vaccinated. In these placebo tests, z is the running variable on the time of survey completion, with a cut-off at 17:00 for each day to the left (graph (a)) or the right (graph (b)) of the true discontinuity (17:00 on the 18th of March). The missing point estimates are due to the low number of observations around some false cut-off points, which prevents the implementation of the RDD estimator. We follow Cattaneo14 with the following options: triangular kernel; variance-covariance matrix estimated using the heteroskedasticity-robust nearest-neighbour variance estimator and local linear polynomial regression based on the MSE-optimal bandwidth selector for each side of the cut-off.

Fig 5. Evolution of the vaccination rates (full vaccinations) in the EU, the UK, Israel and the US.

Notes: Source: https://ourworldindata.org/covid-vaccinations.

Fig 6. Cumulative number of doses administered in Luxembourg, by manufacturer.

Notes: Source: https://vaccinetracker.ecdc.europa.eu.

Fig 7. Proportion of administered doses in Luxembourg, by manufacturer.

Notes: Source: https://vaccinetracker.ecdc.europa.eu.