Among politically motivated attempts to find bright spots in a grim outlook for control of covid 19 in the US, I find it disturbing to see a chronic bias in reports and visualizations of trends in mortality. Today's Washington Post, for example, has published today the latest in a series of graphs (p. A9) that shows coronavirus deaths in the US at 524 on August 3rd (7 day moving average), down from levels a few days before of around 1,000. National and state leaders have pointed to "decreases" of this sort as evidence of a downward trend in hospitalization and mortality rates and reasons for returning to business as usual. In their efforts to generate timely predictions, predictive modelers take up-to-the- minute data and show downward trends in infections, hospitalizations, and deaths. What is biasing these statistics and projections? I contend that incomplete data due to lags in reporting, whether intentional or due to difficulties in collecting time data during a pandemic, drag down trends in time series
As part of a little pro bono project, I collected end of week counts of all cause deaths (ACD) at different as of times during the past few months from
https://data.cdc.gov/NCHS/Provisional-COVID-19-Death-Counts-by-Week-Ending-D/r8kw-7aab . CDC/NCHS has accelerated collection of coronavirus infection, mortality, and other time series data to an astonishing degree during the pandemic. They caution against lags in reporting, but other sources of similar data and predictive modelers using these data have not. Though sparse, the end week ACD counts for multiple as of dates show in general that the longer the lag in reporting, the smaller the changes in counts.
This graph of the end week counts of ACD in Florida by lag in reporting. It shows substantially incomplete counts for more than two weeks.
Predictions and tends that extrapolate incomplete time series values will be biased downward. Adjustments based on sparse data will lead to more accurate predictions and visualizations.
I have placed data and a program adapted from one by Rick Wicklin of SAS(R) here:
https://github.com/SWHermansen/PutDataBackIntoDataScience------------------------------
Sigurd Hermansen
------------------------------