Seconder of the Vote of Thanks and Contribution to the ‘First Discussion Meeting on Statistical Aspects of the Covid-19 Pandemic’
Sheila Macdonald Bird · Journal of the Royal Statistical Society Series A (Statistics in Society) · 2022
Two of this evening's papers are retrospective: modelling COVID-19 infection trajectories using a piecewise linear quantile trend model and the first weeks of COVID-19 in China – small data, big time. The third is economic. Intriguingly, in forecasting Purchasing Managers' Indices (scored 0–100), Nason and Wei introduce COVID-19 mitigation stringency indices and monthly COVID-19 death-rates per million of population as time-series of exogenous variables in their Generalised Network Autoregressive model. All three papers concern on how pandemic-related events evolve over time but, in infectious disease epidemics, the event-time which matters is occurrence-date. Prior to both Delta-variant and immunization, one-third of SARS-CoV-2 infections were never symptomatic. Since symptom-onset-date is not generally applicable, focus falls on swab-date. Next steps after the swab is taken are: receipt by diagnostic laboratory, laboratory-analysis, swab-result reported-back to citizen, swab-positive-results reported-in centrally, daily count of newly-diagnosed SARS-CoV-2 infections reported to the media; additionally, report-date to the World Health Organization (WHO). However, WHO-date, media-date and swab-date for new SARS-CoV-2 infections are seldom the same. When analysing COVID-19 infection trajectories internationally, the epidemiological-date that matters is swab-date: neither media-date nor WHO-report-date. The piecewise linear quantile trend model by Jiang et al. can, of course, be implemented on a diversity of event-scales. But the scale closest to when the pandemic takes citizens out-of-circulation is swab-date. Others are confounded by a series of delays: country-specific, time-varying or both. No10 press conferences during the first wave of COVID-mention deaths in England and Wales displayed a date-scale on which COVID-mention hospitalised deaths evolved. Statisticians, but neither the media nor the general public, immediately knew that report-date, not death-date, was the undisclosed date-scale. Why? Because the tell-tale dropping-off of reports for the most recent few days was not in evidence! In late March 2020, the Department of Health and Social Care (DHSC) began to release publicly the death-dates for each day's daily-reported COVID-mention deaths in hospitals in England. Hence, Bent Nielsen and I were able to estimate the reporting-delay distribution and put into the public domain (via Science Media Centre) our nowcasting of hospitalised COVID-deaths in English regions and by age-group, see http://users.ox.ac.uk/∼nuff0078/Covid/index.htm. Fellow scientists were grateful. Modellers' more sophisticated work, which informed SAGE, was not then routinely in the public domain but now is, for example: https://www.mrc-bsu.cam.ac.uk/now-casting/nowcasting-and-forecasting-7th-september-2021/. Progress in RSS's campaign to end the late registration of deaths in England, Wales and Northern Ireland has included: correct labelling of table-legends and time-axes by the Office for National Statistics (ONS) and provision by NHS Digital of registration-date as well as death-date when cohorts are flagged for mortality. As National Statistician, Sir Ian Diamond resolved to sort the issue but the pandemic supervened. Importantly, ONS reports COVID-mention deaths by occurrence-date, as well as by registration-date. For example, England's first COVID-mention death occurred on 30 January 2020 but was not registered as such until September 2020 after this 84-year-old gentleman's family, who suspected COVID-19, asked for his autopsy to be reviewed. Jiang et al. do not model the ‘COVID-19 infection trajectory’ – rather the trajectory of report-dates for newly diagnosed SARS-CoV-2 cases. Report-dates are confounded by reporting-delays which vary internationally, suffer weekend and vacation perturbations, which contribute to heteroscedascity. The authors' historical piecewise linear quantile trend model is useful in segmenting time-series into periods with relatively stable behaviour. But, inevitably, forecasts are generated from observations in the last segment where, in practice, variance may itself change: for example, due to variant-pressure on laboratories. Neatly, the authors propose 80% confidence intervals based on their forecasting ahead for the 10th, median and 90th quantiles. There is much to like in Nason and Wei's quantification of national economic response to the stringency of COVID-19 mitigations (highly influential) and death-rates per 1 million of population (less so) via Purchasing Managers' Indices (PMIs) using Generalised Network Auto-Regressive models with exogenous variables (themselves time-series). Specifically, parsimony (order of N not N2), dynamic networks (weighted by exports between countries), missingness managed (nearest neighbours with data ‘represent’ those without data), global versus local alpha (global better), experimentation enabled (e.g. fully connected trade network vs. nearest neighbour(s), say: country to which UK exports most). But I am supposed to be critical. Hence, the public's and Purchasing Managers' response to stringency measures (e.g. lockdown) may be different second or third time around versus during the analysed first wave; available covariates are not necessarily the best – I note with interest that stringency measures, an early intervention, are more explanatory than registration-delayed mortality series. Vaccination or variant can disrupt the age-related patterns of progression from SARS-CoV-2 infection to death that applied during the first wave. Age-structure in network-countries matters: thus, COVID-mention deaths per million of population has different meaning in United Kingdom (ageing population) versus India or Africa. A week is a long time in politics: a month of fatalities is a very long time in this pandemic. Finally, PMIs were computed by aggregating the views of senior purchasing executives from around 400 companies: 400 in total or per network country; response-rate; how were executive selected and refreshed; and concordance between PMI-scores from trading-neighbours? Finally, reporting delays in Hubei feature also in Zhao's account of ‘Small Data, Big Time’. This paper was completed before infectious diseases physician, Sir Jeremy Farrar, published his inside story, SPIKE, The virus vs the People – in which he recounts how international scientists collaborated, and supported each other in the face of potential sanctions, in ‘outing’ quickly not only the sequence for SARS-CoV-2 which enabled polymerase chain reaction diagnostic tests to be immediately developed but also critical early information about asymptomatic transmission of the virus. Media – whether printed or social – can either be sampled representatively (with difficulty) or included selectively for illustrative purposes. I think that Zhao has chosen the latter but greater clarity on this aspect would be welcome. My second remark relates back to SPIKE: early appreciation of a new infectious disease requires a diversity of experts: infectious disease clinicians, virologists, immunologists, veterinarians (if zoonosis is suspected), parasitologists in addition to infectious disease modellers and biostatistical expertise. Statisticians do not work alone. If we do, we miss out on anticipatory insights that other basic sciences bring to the table well before data emerge; and they miss out on what the theory of infectious diseases portends. Experts matter … for their anticipation. President, ladies and gentlemen, we have enjoyed a fascinating trio of read papers. I congratulate all tonight's co-authors and have great pleasure in seconding the vote of thanks. The vote of thanks was passed by acclamation.