**Factchecking PolitiFact on Greenland’s Shrinking Ice**

--

[A related Wall Street Journal OpEd is here.]

A PolitiFact “fact check” written by Jon Greenberg has examined my statement

*Greenland’s ice sheet isn’t shrinking any more rapidly today than it was 80 years ago.*

and found it “mostly false.” Greenberg supports his finding by noting that “In the 1930s, the Greenland ice sheet lost about 206 gigatons of ice per year.” and “In the 2010s, the ice sheet lost about 247 gigatons per year.”

“Mostly false” is a grave accusation by an organization that claims to be “fact-checking journalism.” I have therefore taken it seriously; accurate statements of fact have been paramount for me during more than four decades as a practicing scientist and scientific advisor.

Scrutiny of the PolitiFact article reveals that Greenberg misleadingly portrayed the data, and his editors failed to do proper due diligence. It is unclear to me why PolitiFact undertook this analysis, which is far from its focus on politics. But for its own credibility and stated commitment to “truth”, PolitiFact should retract its “mostly false” rating. Unfortunately, after two weeks of pondering what I’ve written here, Greenberg responded that *we feel comfortable with our analysis, but that doesn’t mean there isn’t room for alternative views.*

PolitiFact’s incorrect statements and assessment stem from a flawed and incomplete analysis of the available data. A proper discussion of the one dataset Greenberg analyzed shows that the Greenland Ice Sheet (GrIS) lost an average of 241 Gt per year over the ten years 1928–1937, to be compared with the 212 Gt over the most recent ten years (2012–2021). That result, together with a comparable analysis of a second data set that PolitiFact incorrectly dismissed, validates my statement.[1]

Perhaps these errors should be forgiven since PolitiFact’s “fact check” was written by a non-scientist for a general audience. But judgements of statements about science by scientists should be held to scientific standards of completeness and rigor, which are quite unlike those for political discussions, where “alternative views” are the norm.

This post presents the issues in sufficient detail to allow readers to make their own informed judgement about how Greenland’s annual mass loss has changed during the past century. I first discuss the PolitiFact analysis and presentation, follow with an analysis of the dataset it ignores, then use both analyses to quantify the probability that my statement is correct, and conclude with a summary.

Along the way, I highlight (Figure 2) that Greenland’s average annual mass loss has varied greatly since 1900, including a decline by 18 percent during the most recent decade. That fact will no doubt surprise most non-experts and is illuminating context for any headline about Greenland’s accelerating ice loss.

**PolitiFact’s analysis and presentation**

PolitiFact focused on a dataset described in Mankoff *et al.* (2021) (hereafter “Mankoff”) that uses a variety of sources to compile a record of the GrIS annual mass balance (*i.e.*, the net amount of ice lost or gained during each calendar year) from 1840 thru 2021. Their regularly updated mass balance data (hereafter MB) are available here in file MB_SMB_D_BMB_ann.csv, and are plotted in Figure 1. My analysis uses the version of that file downloaded February 10, 2022.

**Figure 1** Annual total mass balance of the GrIS (from Mankoff); the uncertainties shown are 1σ.

Given the high year-to-year variability, some averaging is necessary, and Greenberg chose a 10-year averaging interval. Figure 2 therefore displays the 10-year trailing average of the MB data from 1900 to 2021 (the average annual mass loss is the negative of the average annual mass balance).

**Figure 2** Average annual GrIS mass loss from MB data (1900–2021) using a 10-year averaging interval. Points plotted correspond to trailing intervals; *e.g.*, the point plotted at 1980 corresponds to the average over the ten years 1971–1980. The 1σ uncertainties shown are computed assuming the annual values are statistically independent.

The ups and downs of Figure 2 will likely surprise most non-experts, as they appear to be little in sync with steadily growing human influences. The cause of this decadal variability is beyond the scope of this note, but it is thought to be related to the cyclic atmosphere/ocean dynamics of the North Atlantic (see, for example, here and here). Developments over the next decade or two will be revealing.

Figure 2 shows it would be a stretch to say the average annual losses during the 1930’s were any less than that in recent years. But since the time series remains quite variable even after the 10-year averaging, there are cherries to pick. Indeed, Greenberg selected his fruit carefully, writing that[2]

*The 1930s low point of 206 gigatons of ice sheet loss is not as large as the 247 gigatons average yearly loss between 2010 and 2020.*

The 206 Gt Greenberg quotes for the 1930’s corresponds to the average over the ten years 1930–1939 (it’s actually 207 Gt according to the Mankoff spreadsheet with 1σ uncertainty of ±46 Gt). But the largest average mass loss for 10-year intervals around the 1930’s is 241 ± 47 Gt for the ten years 1928–1937, 35 Gt larger and just two years earlier than the interval Greenberg selected. Greenberg then plucks a second cherry by quoting a 247 Gt average “between 2010 and 2020.” That value is actually the average of the ten years 2010–2019. He thus ignored the most recent ten-year interval that was available to him (2011–2020), which has a smaller value, 227 ± 30 Gt; the average over the most recent 10 years, 2012–2021, is even smaller, 213 ± 30 Gt.

In sum, Greenberg judged my statement by arbitrarily choosing averaging intervals that minimize the past rate and maximize the recent one. A proper assessment compares the greatest ten-year average annual mass loss around 80 years ago, 241 ± 47 Gt, with the average over the most recent decade, 213 ± 30 Gt. As I show below in a probabilistic analysis that considers the uncertainties, the earlier value is likely the greater of the two.

Beyond a misleading choice of analysis intervals, the PolitiFact piece both errs and misleads when it says

*The latest **report card for the Greenland ice sheet** from the National Oceanic and Atmospheric Administration said that in 2012 and 2019, the ice sheet saw its greatest losses since regular monitoring began in the 1950s. In 2012, it shed about 464 gigatons, and in 2019, it lost about 532 gigatons.*

without telling the reader that these numbers cannot be compared with the Mankoff data that Greenberg analyzed. Those report card values are for the “biennial year” that runs from September 1 thru August 31, rather than for the calendar year. More importantly, they are systematically larger than the MB values because they include peripheral glaciers, while Mankoff does not. The MB data show smaller annual losses, 428 ± 110 Gt in calendar year 2012 and 411 ± 99 Gt in 2019. These losses are still large relative to other recent calendar years, but other years since 2010 are comparatively small (for example, only 69 ± 83 Gt in calendar year 2018 and 101 ± 91 Gt in 2017).[3] *In fact, as seen in Figure 2, the 10-year average annual mass loss has declined during the most recent decade*.

**Analysis of a second dataset**

My statement was originally based on work by Frederikse *et al.* (2020) (hereafter “Frederikse”). That paper synthesized a variety of observations to identify the components of global sea level rise from 1900 through 2018. One of those components is mass loss from the GrIS. Their sea level data (hereafter SL) are available here in the file global_basin_timeseries.xlsx. My analysis uses the most recent version of that file, dated April 10, 2020. The relevant data are plotted in Figure 3.

**Figure 3** Global sea level rise associated with mass loss of the GrIS (from Frederikse); the uncertainties shown are 90%.

The SL data can be directly converted to GrIS mass loss by noting that 1 mm of sea level rise is equivalent to 361 Gt of mass loss, since the area of the world’s oceans 361 x 10^6 km^2. An averaging interval also needs to be chosen because the SL data integrate (accumulate) the highly variable annual MB data. I first adopt the standard climatological averaging interval of thirty years, as defined by the World Meteorological Organization, as used by Frederikse in calculating trends, and as I used in my book *Unsettled*’s Figure 8.6.

The upper panel of Figure 4 shows the average annual mass loss calculated from the 30-year rise in the SL data (*i.e.*, from the difference in SL values 30 years apart). The average annual mass loss in the early 20th century clearly exceeds that of recent decades.

**Figure 4** Average annual GrIS mass loss from the SL data (upper, 1929–2018) and MB data (lower, 1900–2021) using a 30-year averaging interval. The graphs plot the trailing interval and extend to the most recent year available for each data set. In both cases, the 1σ uncertainties shown are computed assuming the annual values are statistically independent. Under the assumption of a normal distribution, the 90% uncertainties in the SL data from Frederikse were symmetrized about the mean and reduced by a factor of 1.64 to correspond to ±1σ.

Although the PolitiFact article acknowledges that the SL data show a larger mass loss in the 1930’s than in recent decades, Greenberg excuses the SL data from any detailed discussion through the incorrect statements

*The main problem is that study only went through 2010, and ice losses have accelerated since then.*

and

*Koonin relied on a study that had data through 2010.*

In fact, the SL data extend to 2018, as stated in Frederikse and as given in its data spreadsheet, so average rates for intervals ending in that year can be calculated. *And, contrary to Greenberg’s assertion that ice losses have accelerated since 2010, Figure 2 shows that average losses have decelerated since then.*

For completeness, I show the 30-year averages of the MB data in the lower panel of Figure 4. They agree reasonably well with the SL values (upper panel) for intervals ending after about 1980, but disagree significantly before then, with the SL average annual mass loss peaking at just over 250 Gt for intervals ending in the early 1950s, while the MB average annual losses are 150 Gt during the same period. Disagreements might stem from different scopes (SL includes Greenland’s peripheral glaciers, MB does not, which Mankoff states amount to some 45 Gt per year) and from different weightings and adjustments of various data sources.

Finally, Figure 5 shows the SL data averaged with the 10-year interval Greenberg used.

**Figure 5** Ten-year trailing averages of the GrIS annual mass loss (1909–2021) derived from the SL data. The uncertainties shown are ±1σ

There is reasonable agreement with the MB data (Figure 2) post-1950 (including the deceleration during the most recent decade) after accounting for the different scopes of the SL and MB datasets. Uncertainties in the SL averages become large going back into the 20th century, but these are likely overestimated, since I have not distinguished between systematic and statistical uncertainties. [4]

**Probabilistic analysis**

The veracity of my original statement can be quantified by accounting for the uncertainties when comparing recent mass loss rates with those in the past. For each dataset (SL or MB) and for each averaging interval (30 years or 10 years), I compute the difference *D* between the average annual mass loss in an earlier averaging interval and that in the most recent averaging interval (*i.e.*, that ending in 2018 for SL and in 2021 for MB). Under the usual assumption of independent normal distributions, the uncertainty in that difference, σ, can also be computed, and hence the probability

that the earlier value was greater than the recent value. Figure 6 shows those probabilities for each dataset for the 10-year averaging interval (upper) and for the 30-year averaging interval (lower). In discussing them, I use IPCC’s formal uncertainty terminology: *Unlikely* (10–33%), *As likely as not* (33–66%), *Likely* (66–90%), *Very likely* (90–99%).

For the 10-year averaging interval used by Greenberg, the upper panel of Figure 6 shows that for the SL dataset it is *likely* that mass loss in the earlier interval exceed that in the most recent interval during the six years 1934–1939 and *about as likely as not* that it did during the 14 years 1930–1943. For the MB dataset, it was *about as likely as not* during the nine years 1932–1940. For the 30-year averaging interval (lower panel of Figure 6), there are eleven years (1949–1959) in the SL dataset where it is *very likely* that the earlier rate exceeds the recent rate, and that those years are embedded within a longer 28-year period (1936–1963) where it is at least *likely*. The MB dataset is less definitive, with only the two years 1952–1953 having 30-year rates *as likely as not* greater than the most recent year. Although the yearly values in these graphs are not statistically independent, these periods are long enough compared to the averaging interval to be significant. The decline in the average mass loss during the past decade is also evident in the upper panel.

I can now consider various statements that accurately express this quantitative analysis. For the 10-year averaging interval, one can say with *high confidence* that

*It is likely that Greenland’s ice sheet isn’t shrinking any more rapidly today than it was eight decades ago*

since the SL dataset shows *likely* during the 1930’s and the MB data shows *as likely as not* in that time.[5] For the 30-year interval, it can be said with medium confidence that

*It is likely that Greenland’s ice sheet isn’t shrinking any more rapidly today than it was seven decades ago*

since the SL dataset shows *very* *likely* during the 1950s, in disagreement with an *unlikely* conclusion from the MB dataset.

**Figure 6** (*Upper*) The probability that the average annual mass loss during an earlier 10-year averaging interval exceeds that in the most recent interval. Values of *P* for both the SL and MB datasets are plotted at the final year of each interval. Also shown are the probability ranges of the IPCC formal uncertainty language. (*Lower*) Similar to the upper figure, but for 30-year averaging interval.

**Summary**

A thorough and consistent analysis of two independent datasets pertaining to the annual mass loss of Greenland’s ice sheet does not support PolitiFact’s finding that the statement

*Greenland’s ice sheet isn’t shrinking any more rapidly today than it was 80 years ago.*

is “mostly false.” Rather the analysis shows that it’s “mostly true” and would be absolutely true with a minor modification to

*It is likely that Greenland’s ice sheet isn’t shrinking any more rapidly today than it was seven or eight decades ago.*

Readers can judge for themselves whether dropping “It is likely” and saying “80 years” instead of “seven or eight decades” significantly changes the meaning.

_________________________________________________________________

[1] To set a scale for these numbers, if the average annual loss were 250 Gt, it would take some 12,000 years for the GrIS to vanish.

[2] Greenberg’s use of “low” here is incorrect and confusing, as he’s referring to an alleged high in the loss.

[3] Even the “report card” mass loss, derived from GRACE observations, shows high interannual variability. It was indeed high during 2019 (532 ± 58 Gt) but was unusually low in 2018 (98 ± 26 Gt) and 2021 (85 ± 16 Gt).

[4] For simplicity, I have assumed statistical independence of the underlying SL and MB annual values in these analyses. Any systematic errors (*i.e.*, time-independent biases) or positive correlations between random errors will not change the means of the differences between the most recent and past averages but will reduce their uncertainties. This overestimate of the uncertainties is likely most important for the 10-year analysis of the SL data shown in Figure 5, where I’ve obtained average annual mass loss from the difference of sea levels separated by only a decade.

[5]The overestimate of uncertainties in the SL analysis discussed in the previous footnote would increase those probabilities greater than 0.5 and hence the confidence in this statement.