COVID-19: The Trouble With the Case Curve During the Holidays

The familiar curve chart that shows new Covid-19 cases will be disrupted by testing and data reporting interruptions just as people prepare for another holiday season in the United States. Last year, the national case curve showed pronounced declines after Thanksgiving and Christmas that did not reflect real decreases in new infections.

The impact of holidays may be even more noticeable this time around, as illustrated by the recent Labor Day holiday, because states are reporting data less consistently than they did a year ago. In response, we’re updating the averaging methodology that turns the daily case and death data we report into a curve chart. ...

...

The average number of cases reported each day declined after both Thanksgiving and Christmas, but these drops corresponded to decreases in the number of tests performed, meaning that more cases were likely going uncounted. Once tests returned to previous levels, cases quickly climbed again to their peak in early January. Those deceiving decreases were not reflected in the chart showing the number of people hospitalized with Covid-19 during the holiday season.

The number of people hospitalized with Covid is a lagging indicator of community infection rates, but is less sensitive to holiday data interruptions because data from hospitals is reported directly to the Department of Health and Human Services. And if a hospital fails to report for a single day, the same number of Covid-19 patients from the previous day can be easily carried forward. ...

I've pointed out the distortionary effects of holiday dates on COVID-19 (and other healthcare and epidemiological) reporting previously. That's occasionally been met by some strong pushback by those insisting that the changes in reported data reflect actual incidence fluctuations.

The simple truth is that they do not, that the same fluctuations appear across numerous data series ... and are curiously absent from others. As the New York Times article notes, the pronounced "notches" which appear at Thanksgiving and Christmas in US data (and smaller notches appearing at Memorial Day, Independence Day, and Labour day, all large US national holidays often coinciding with travel) are not reflected by fluctuations in hospital admissions, which are both a lagging and more persistent trend.

I'd first noticed this about two decades ago in an earlier life doing health care data anlysis. I'd been tasked with making sense of a set of outpatient data for a clinic --- mostly dealing with minor or routine issues. The daily-level data were quite noisy, and what I quickly learnt applying different smoothing intervals was that at a weekly level virtually all of the daily "noise" disappeared, though what remained was a pronounced 7-day trend with large drops on weekends. This pattern also persisted when longer multiples-of-seven smoothing periods were applied: 14, 21, and 28 days. I (and most official data analyses) tend to stick with the shortest useful smoothing period, though for smaller countries with thinner data, I often wish that a longer (2--4 week) smoothing period would be applied as data over shorter periods remain quite noisy and give the spurious appearance of real changes which are not in fact warranted..

The other strongly evident pattern was that claims dropped precipitously on and around holidays, with the largest declines appearing at Thanksgiving and Christmas for US data. This was evident across several years' worth of data. And so when I started looking at COVID19 data, the holiday dips stood out to me as they occurred, first with a late-May dip (Memorial Day, US), then the 4th of July (Independence day), Labour Day, and of course the larger Thanksgiving and Christmas / New Years dips (there's a trend that flows through the period from the 25th of December through the 1st of January).

And now the New York Times seem to have recognised this.

You'll find similar patterns for data in other countries, of course. In Canada, Canada Day (July 1), Labour Day (Sept. 6), and Thanksgiving (October 12) are visible, along with the much more pronounced Christmas holiday:

https://www.worldometers.info/coronavirus/country/canada/

In Germany, Easter Week shows a major dip:

https://www.worldometers.info/coronavirus/country/germany

Again, what characterises a holiday data reporting dip is:

• A fall in reported data that are themselves based on direct and often voluntary assessment or filing times. Covid tests and (in many countries) coronors' reports if based on filing date rather than date of death will follow these trends. Hospital admissions rather less so as these facilities both remain operational and are accessed when truly needed.
• A resumption of an earlier trendline. You can usually imagine a consistent smoothed line running through the time period.
• An absence of a sharp exponential growth trend. Again, where there are triggering events leading to outbreaks, what is not seen is a sudden spike in cases, but a slow growth following a doubling trend. As has been observed, "Exponential growth doesn't mean 'fast', it means 'imperceptibly at first, then suddenly overwhelming'." A holiday-induced outbreak would look much more like South Dakota following the 2020 Sturgis Rally (~8--19 August, 2020), with a slow rise appearing by the 23d, and a sharp climb from the 25 of August through September 3. Cases subsequently fell, but began rising again after the 14th of September, and continued to climb until their November 15 peak. For South Dakota in 2020, both the Thanksgiving and Christmas notches appear as the overall case rate was falling rapidly.

https://www.worldometers.info/coronavirus/usa/south-dakota/

https://www.nytimes.com/interactive/2021/11/22/us/covid-data-holiday-averages.html

#COVID19 #epidemiology #pandemics #thanksgiving #christmas #statistics