Examining Colorado’s Covid-19 Outbreak Using Control Charts

What helps us understand what’s going on, in a useful, actionable way?

Published: Wednesday, May 6, 2020 - 12:02

Story update 5/6/2020: The charts and some data have been updated to reflect the data available on the date this article was published.

During the Covid-19 stay-at-home order in Colorado, I've become increasingly frustrated by Covid-19 charts. Most of what I see are cumulative column charts, which don't give any real insight into what's going on. Are we really flattening the curve?

So I decided to use the state's Covid-19 statistics for Colorado and Denver county, and see what I could learn using control charts. Control charts have been around for almost 100 years. They use formulas to calculate control limits that encompass 99.7 percent of the data points. This makes it easy to monitor any process and detect process shifts and "out of control" conditions.

Source: https://covid19.colorado.gov/case-data Click image for larger view.

Quality Digest's editor in chief, Dirk Dusharme, ran this article by Donald J. Wheeler to see what he thought of this approach. Wheeler commented that I was not the first to use a control chart on Covid-19 data, but questioned its value. He offered this observation: "The large amount of variation from day-to-day creates wide limits, and even then points fall outside the limits. The question we need to ask of these data is different from the question asked by a process behavior chart. Instead of looking at data that we think came from an unchanging system, we are looking at data from a system that is known to be changing. Rather than asking, 'Has a change occurred?' we need to ask, 'In what way is the system changing?'" In his article "Tracking Covid-19," Wheeler looked at U.S. Covid-19 cumulative data using a semilog plot.

So I'm knowingly violating a prerequisite of the XmR chart, but I still find the results interesting. My question is always: "Does it help me understand what's going on, in a useful, actionable way?"

I found the Colorado statistics here.

If you want to try this with your state, The New York Times is collecting state and county data here. You will need to use Excel PivotTables to summarize the data.

I had to organize and clean up the data before charting it because each file is a snapshot of an individual day. Colorado's stay-at-home order went into effect March 26, 2020, so I added a process change at that point. And, as of May 5, 2020, it looks like this:

Click image for larger view.

Hospitalizations and deaths are both lagging indicators of the disease. Cases by day are a more reliable indicator of the current state of the disease. There are also cases by reported date of illness, but I'm guessing that is more unreliable.

As Wheeler points out, the variation from various outbreaks around the state makes it difficult to use XmR charts. But I think of this like a startup process in a factory: It takes a while for the process to stabilize and get up to speed. This often creates a lot of scrap. But it does become stable.

Although the case count has stabilized since I first created this dashboard on April 14, 2020, if I were a politician, I'd want to see a downward trend or run of eight points below the center line to know that existing countermeasures are working to not only slow the disease, but also reduce it.

I also wanted to know about my home county, Denver. There was an upward trend from March 23–28, 2020, signaling a process shift before the March 26, 2020, stay-at-home order. Without the stay-at-home order, the trend may have continued upward, overwhelming hospitals.

It took a couple of days for Covid-19 case counts to stabilize. While it looks like the cases may have shifted upward again on April 10, 2020, there's no clear trend or run to signal the shift. Counties are more likely to stabilize quickly after stay-at-home orders.

Click image for larger view.

Some people will say that the official case counts are below actual. True, but the official case counts reflect the actual, so it's a useful indicator. And the best one we have.

I chose XmR charts (the "Swiss Army knife" of control charts) because cases per day is a ratio, easily charted with an XmR chart. And the R chart gives us a view of daily variation so easily lost in cumulative data. Some people might have used a c-chart, but I found XmR charts to be more informative.

These charts were created using QI Macros XmR Control Chart dashboard.

To reopen the economy, we will need to be able to identify hot zones, cold zones, and zones that are heating up. As of April 12, 2020, one Denver hospital has had no Covid-19 patients; it's a cold zone. Control chart rules tell us that six points in a row ascending indicate that there is a process shift; a growing hot zone. Both the Colorado cases and Denver cases show a hot zone trend during the initial weeks of Covid-19 tracking. Six points in a row descending or eight points below the center line is a cooling zone.

Here's a map of Denver Public Health's Covid-19 zone . There are 11 hot zones and a variety of cool to warm zones. I live and work in a cool zone near the bottom of the map.

Click image for larger view.

What would control charts of your state and local Covid-19 statistics tell you? Are you heating up or cooling down?

About The Author

Jay Arthur—The KnowWare Man

Jay Arthur, speaker, trainer, founder of KnowWare International Inc., and developer of QI Macros for Excel, understands how to pinpoint areas for improvement in processes, people, and technology. He uses data to pinpoint broken processes and helps teams understand their communication styles and restore broken connections. Arthur is the author of Lean Six Sigma for Hospitals (McGraw-Hill, 2011), and Lean Six Sigma Demystified (McGraw-Hill, 2010), and QI Macros SPC Software for Excel. He has 30 years experience developing software. Located in Denver, KnowWare International helps service and manufacturing businesses use lean Six Sigma tools to drive dramatic performance improvements.