5
$\begingroup$

I'm trying to understand/reproduce/revise this chart in the Southeast section of the United States National Climate Assessment. The truncated scale for the bars is what caught by eye as a classic chart lie, but I'm also wondering about other features: different thresholds (besides 3 inches) and different aggregations (besides decades).

enter image description here

Original caption for this chart plus a map next to it:

Figure 19.3: The figure shows variability and change in (left) the annual number of days with precipitation greater than 3 inches (1900–2016) averaged over the Southeast by decade and (right) individual station trends (1950–2016). The number of days with heavy precipitation has increased at most stations, particularly since the 1980s. Sources: NOAA NCEI and CICS-NC.

The vague source citations are not particularly helpful, but I did find NOAA's Global Historical Climatology Network Daily downloadable data. That has daily precipitation data for 100,000 stations around the world, with about 12,000 in the Southeast. However, in a given year, I'm seeing 100 to 200 days with precipitation above 3 inches, but the chart indicates a different order of magnitude.

What does "days with precipitation greater than 3 inches" mean in this context?

I'm guessing there must be a "per station" part to it, but I still can't guess the exact math. The simplest treatment, a weighted average of each stations rate of heavy rain, produces the chart below.

enter image description here

It's close (same order of magnitude and peaks in 40s, 70s and recent), but not quite the same. Is there a different data source? Or a different treatment for irregular samples? Or a way to account for increasing number of stations?

Update after comments: If I filter on stations with at least 110 years of data and use only years with 200+ days of data, I get the following, which is closer. (Using the same scale as the original, but avoiding bar truncation by just showing the tops.)

enter image description here

Most decades have lower values than the original.

$\endgroup$
  • $\begingroup$ I certainly haven't spent that much time closely monitoring or analyzing their results (glad to see others doing so!) and so any input is rather uninformed and just thoughts... by weighted average, you're meaning you're taking into account the spatial distribution, yes? Wonder if perhaps they used a more limited set of sites.... one thing that would seem somewhat important to such a graph would be station existence... it'd seem best to only use stations existing since the start of the graph's timeperiod to ensure the change in weighting/data availability doesn't alter the graph over time? $\endgroup$ – JeopardyTempest Dec 11 '18 at 10:42
  • 1
    $\begingroup$ Thanks. There are about 425 stations out of 12,000 with 100+ years of data. Looking only at those (or higher thresholds), my spike in the 2010s goes down to be closer to the original (a lot of stations are relatively new), but the whole chart is still not a great match -- 1940s is still barely above 1.0. Progress... $\endgroup$ – xan Dec 12 '18 at 2:22
  • 1
    $\begingroup$ I added the latest chart to the question body. The number of stations per year doesn't vary much once I filter on stations with so many years present. I have emailed the author of the chart and am in the mean time collecting data from another source (ACIS) to compare. I have worked on my own gridding, but I have to figure out a way to fairly compare grid cells with different station counts. $\endgroup$ – xan Dec 12 '18 at 23:28
  • 1
    $\begingroup$ Indeed, still work to be done: combined overlay of your lines with the original graphic to recreate $\endgroup$ – JeopardyTempest Dec 13 '18 at 5:19
  • 1
    $\begingroup$ I got some more details from the author and am even closer. Will post his response as an answer if I get permission, and hopefully a couple clarifications. In short: stations with 90% nonmissing data and years with 300+ nonmissing days. $\endgroup$ – xan Dec 14 '18 at 2:07
2
$\begingroup$

After contacting the author of the chart, I believe the following defines the process behind the chart.

  1. Data source: NOAA's Global Historical Climatology Network Daily station data for "PCPT".
  2. Use only stations in the Southeast states which have data for 90% of days from 1895 to 2011, to match stations used previous edition of the climate report.
  3. Use only years with at least 300 days of data.
  4. Remove data values that have been flagged in GHND for quality issues.
  5. Count the number of days of heavy rainfall (3 or more inches) for each station and year.
  6. Using a 1° by 1° lat/lon grid, average counts of stations in the same grid cell.
  7. Report the average of all grid cell-years for a given decade.

I have still not been able to reproduce the chart because I have had trouble getting clean data. Will update with a chart when that happens.

$\endgroup$
  • $\begingroup$ The 1 by 1 seems a bit quirky... for example a lot of partial cells in places like Florida counting equitable to whole cells, so it would over-represent Florida perhaps (depending on how many cells even have stations with that length of years!) And I wonder "only use years with 300 days"... do they normalize the final result numbers if some grid cells end up empty because a year was thrown out? Lots of little questions... but probably not that crazy of differences I guess. $\endgroup$ – JeopardyTempest Dec 21 '18 at 19:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.