arising from: R. F. Savaris et al.; Scientific Reports https://doi.org/10.1038/s41598-021-84092-1 (2021).
Introduction
Identifying the impact of Non-Pharmaceutical Interventions (NPIs) against COVID-19 has been an important topic throughout the pandemic. It is, in many ways, one of the most difficult scientific challenges facing the research community, as these interventions are often implemented in parallel, and it is challenging to disentangle how various methods of avoiding infection may have impacted case and death rates across the world.
To this vital question, we read with interest the paper written by Savaris et al. entitled “Stay-at-home policy is a case of exceptional fallacy: an internet-based ecological study”1.
The authors found no evidence that COVID-19 deaths were reduced by more people staying at home in early 2020. However, we believe that several key deficiencies within the methodology make their conclusions unsupported.
Savaris et al. developed a custom approach to detect whether changes in mobility and COVID-19-related mortality in a country are associated: for each country in their dataset, and for each week between 2020-02-16 and 2020-08-22, they calculated the relative change in residential mobility and the relative change in COVID-19-related deaths. For every possible pair of countries A and B, they then subtracted the mobility/mortality time series of country A from the mobility/mortality time series of country B. Finally, the authors calculated the regression for “difference in change in mobility between the countries” versus ”difference in change of mortality between the countries”. They interpret a statistically significant correlation (i.e., the slope is not zero) as evidence of an association between mobility and mortality. As only about 5% of comparisons are statistically significant at p < 0.05 and only 1.6% passed further statistical tests—Savaris et al. argue that there is little evidence that staying home reduces the number of deaths/million.
Unclear suitability of the chosen approach
In our opinion, the chosen statistical approach may not be able to detect an effect of staying at home (or changes in mobility) on mortality at all.
The approach does not intuitively make sense to us as it intentionally removes potentially meaningful information from the raw data. For example, the temporal order of the weekly mobility/mortality changes is not taken into account. This means that—by design—the approach cannot determine causality, i.e., whether changes in mobility affect mortality or changes in mortality affect mobility. It also seems unusual that the inherent delay2,3,4,5 between changes in mobility and their potential effect on mortality is not considered in this approach at all. Furthermore, only the weekly change in mobility/mortality is analyzed, ignoring the absolute magnitude of both values. Góes6 argues that the approach is actually mathematically incapable of determining whether a correlation between mobility and mortality exists.
Given these observations, we see reason to assume that the novel approach proposed by Savaris et al. is not actually suitable for investigating the question at hand. The authors did not demonstrate that their approach would actually be able to detect an effect if it existed.
Failure to replicate
Thankfully, the authors published source code and data that allow for replicating their analysis. We could reproduce the results shown in the second figure of their paper using the provided dataset and the Jupyter notebook. It seems that the p-values in the second figure of their paper have not been corrected for multiple comparisons, so the legend might be misleading. In the following, we do not adjust for multiple comparisons, either.
To determine whether the approach could actually detect an effect of changes in mobility on mortality, we conducted four experiments using the code in Savaris’ Jupyter notebook.
First, we constructed a data set with five hypothetical countries where changes in mobility (or the lack thereof) had an obvious effect on mortality (Fig. 1). When running this dataset through the Jupyter notebook, none of the comparisons were statistically significant at p < 0.05 (Fig. 2). According to the authors’ criteria, this would mean that even for our handcrafted dataset, no effect of mobility on mortality could be found.
To address concerns that the approach might just be unsuitable for artificial datasets that don’t resemble real data, we created another dataset for which a working algorithm should detect an effect of mobility on mortality: for each of the 87 countries in the original dataset, we kept the mortality time series but generated a new mobility time series. This was done by inverting each value in the mortality dataset and shifting the data to the left by 3 weeks. Missing values at the end of the time series were set to the last valid value. This resulted in a mobility time series where each change in mobility would “result” in a change in mortality exactly three weeks later (Fig. 3).
Running this dataset through the notebook resulted in 414 of 7482 comparisons (5.533%) being statistically significant (Fig. 4). This is quite precisely the expected result if there was no association at all in the dataset8—indeed, the expected result if the code could not detect any effect would be 5% of values at p < 0.05 by definition. However, a majority of “significant comparisons” were caused by the obviously erroneous dataset from Peru. Following the reasoning by Savaris et al., all of the comparisons should have been statistically significant.
As a third experiment, we applied the algorithm by Savaris et al. to a dataset of 87 artificial countries with mobility/mortality data generated by a pseudo-random number generator (Fig. 5). This resulted in 742 of 7482 (9.917%) statistically significant comparisons at p < 0.05 (Fig. 6). That means that random data seems to provide better evidence for an effect of mobility on mortality than both datasets that we designed to demonstrate such an effect.
As we were unable to identify datasets for which the approach by Savaris et al. would actually result in statistically significant comparisons, we asked the authors for an example for which their method would demonstrate an effect of lockdowns. They sent a dataset with two hypothetical countries where lockdowns are indeed effective in reducing mortality and staying open results in a rising number of deaths. For this dataset, the notebook indeed outputs a p-value of < 0.0001 (Fig. 7, top). However, adding a small amount of random noise to the dataset resulted in a non-significant regression again (Fig. 7, bottom).
To us, the outcomes of these experiments indicate that the proposed method (a) cannot reliably detect an effect of mobility on mortality even if it should exist, (b) cannot distinguish meaningful data from random data, and (c) is highly susceptible to noise in the data.
Overall, this seems to make the approach unsuitable for the task at hand.
Google mobility data
While Google’s mobility data is opaque and dimensionless it is not unreasonable to use such data as rightfully pointed out by the authors who cite previous work9,10. However, unlike prior works, the authors used only the “residential” category of the mobility data. Residential changes, as noted by Google11, are the least likely to show any impact when behaviour changes, since people already spend a significant amount of time at home in normal circumstances. This does not mean that residential mobility scores are always low, but that they will always be the lowest of the mobility metrics and thus most likely to result in a null comparison. Moreover, they are the least likely to change on a weekly basis, and since the author’s model only takes into account the change, this would inherently bias the analyses towards a null result.
Ecological fallacy
Ecological studies of this kind are also geared towards null findings. As one example, the country of Australia is included as a single unit of analysis. However, during the time period of the study, Australia had a unique situation, with a single state (Victoria) locking down for months while the rest of the country opened up. At the country level this difference becomes almost negligible, with an increase of < 10% in “residential” mobility across Australia. However, in Victoria alone, the increase exceeded 20% for this entire period. Any comparisons to Australia as a whole during the Victorian lockdown are hence unreliable as a measure of the key metrics of interest, and while this is perhaps most significant for Australia it is clearly not limited to a single nation. Russia, the United States, and many other places suffer from the exact same issue. This limitation is noted in the discussion, but it is not just a notation but a fundamental limitation that makes the final numbers largely arbitrary. In addition, it should be noted that the death numbers from Belarus are overtly implausible, and the numbers from Peru contain obvious reporting errors. Both countries should have been excluded from the analysis.
Confounding by indication
Finally, including only countries with more than 100 overall COVID-19 deaths in the analysis excludes countries such as New Zealand, Singapore, and Vietnam that had success with rapid, strict measures. This introduces a bias towards countries without effective measures and therefore increases the chance that no difference between countries can be found5,12.
Discussion
While the question of whether NPIs can influence COVID-19 deaths is of great importance, the issues we have highlighted seriously weaken the conclusions made by the authors. That they found no statistically significant differences between regions on these metrics may simply be a function of their chosen methodology and the inherent limitations of the mobility dataset, and might have little to do with the matter at hand. It appears likely that the methodology could not detect an effect of staying at home on mortality even if one were to exist.
Data availability
The source code and datasets used for our attempt at replication can be found on our GitHub repository7.
References
Savaris, R. F., Pumi, G., Dalzochio, J. & Kunst, R. Stay-at-home policy is a case of exception fallacy: An internet-based ecological study. Sci. Rep. 11, 5313. https://doi.org/10.1038/s41598-021-84092-1 (2021).
Gupta, S., Simon, K. & Wing, C. Mandated and voluntary social distancing during the COVID-19 epidemic. Brook Pap. Econ. Act. 10, 269–315. https://doi.org/10.3386/w28139 (2020).
Brauner, J. M. et al. Inferring the effectiveness of government interventions against COVID-19. Science 371, 6531. https://doi.org/10.1126/science.abd9338 (2020).
Li, Y. et al. The temporal association of introducing and lifting non-pharmaceutical interventions with the time-varying reproduction number (R) of SARS-CoV-2: A modelling study across 131 countries. Lancet. Infect. Dis 21, 193–202. https://doi.org/10.1016/S1473-3099(20)30785-4 (2020).
Besançon, L., Meyerowitz-Katz, G., ZanettiChini, E., Fuchs, H. & Flahault, A. Challenges in determining causality: An ongoing critique of Bendavid et al’s ‘Assessing mandatory stay-at-home and business closure effects on the spread of COVID-19’. Eur. J. Clin. Investig. https://doi.org/10.1111/eci.13599 (2021).
Góes, C. Pairwise difference regressions are just weighted averages. Sci. Rep. 11, 23044. https://doi.org/10.1038/s41598-021-02096-3 (2021).
COVID-19-Mobility-vs-Mortality, GitHub Repository (2021). https://github.com/RaphaelWimmer/COVID-19-Mobility-vs-Mortality (Accessed 15 March 2021).
Munroe, R. “Significant.” xkcd. https://xkcd.com/882/ (Accessed 20 July 2021).
Delen, D., Eryarsoy, E. & Davazdahemami, B. No place like home: Cross-national data analysis of the efficacy of social distancing during the COVID-19 pandemic. JMIR Public Health Surveill. 6, e19862. https://doi.org/10.2196/19862 (2020).
Vokó, Z. & Pitter, J. G. The effect of social distance measures on COVID-19 epidemics in Europe: An interrupted time series analysis. Geroscience 42, 1075–1082. https://doi.org/10.1007/s11357-020-00205-0 (2020).
Understand the Data—Community Mobility Reports Help. https://support.google.com/covid19-mobility/answer/9825414?hl=en&ref_topic=9822927 (Accessed 15 March 2021).
Confounding by Indication (2019). https://catalogofbias.org/biases/confounding-by-indication/ (Accessed 04 April 2021).
Author information
Authors and Affiliations
Contributions
G.M.K., L.B., R.W. and A.F. drafted the document. R.W. performed the modelling and analysis and generated all figures.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Meyerowitz‐Katz, G., Besançon, L., Flahault, A. et al. Impact of mobility reduction on COVID-19 mortality: absence of evidence might be due to methodological issues. Sci Rep 11, 23533 (2021). https://doi.org/10.1038/s41598-021-02461-2
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-021-02461-2
This article is cited by
-
What scientists have learnt from COVID lockdowns
Nature (2022)
-
Retraction Note: Stay-at-home policy is a case of exception fallacy: an internet-based ecological study
Scientific Reports (2021)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.