The Economist - USA (2022-02-26)

TheEconomistFebruary26th 2022

Graphic detail Fishy covid-19 data

85

More equal than

others

S

ometimesthenumbersaresimplytoo

tidy  to  be  believed.  Irregular  statistical

variation  has  proven  a  powerful  forensic

tool  for  detecting  possible  fraud  in  aca­

demic  research,  accounting  statements

and  election  tallies.  Now  similar  tech­

niques are helping to find a new subgenre

of faked numbers: covid­19 death tolls.

That is the conclusion of a new study to

be  published  in  Significance,  a  statistics

magazine, by the researcher Dmitry Kobak.

Mr  Kobak  has  a  penchant  for  such  stud­

ies—he  previously  demonstrated  fraud  in

Russian elections based on anomalous tal­

lies from polling stations. His latest study

examines  how  reported  death  tolls  vary

overtime.  He  finds  that  this  variance  is

suspiciously low in a clutch of countries—

almost  exclusively  those  without  a  func­

tioning democracy or a free press.

Mr Kobak uses atestbasedonthe“Pois­

son  distribution”.  This  is  named  after  a

French  statistician  who  first  noticed  that

when  modelling  certain  kinds  of  counts,

such as the number of people who enter a

railway station in an hour, the distribution

takes on a specific shape with one mathe­

matically  pleasing  property:  the  mean  of

the distribution is equal to its variance.

This idea can be useful in modelling the

number  of  covid  deaths,  but  requires  one

extension.  Unlike  a  typical  Poisson  pro­

cess,  the  number  of  people  who  die  of

covidcan be correlated from one day to the

next—superspreader  events,  for  example,

lead to spikes in deaths. As a result, the dis­

tribution of deaths should be what statisti­

cians  call  “overdispersed”—the  variance

should  be  greater  than  the  mean.  Jonas

Schöley, a demographer not involved with

Mr  Kobak’s  research,  says  he  has  never  in

his  career  encountered  death  tallies  that

would fail this test.

That  should  make  it  easy to  pass.  And

the  vast  majority  of  countries  reporting

datato the World Health Organisation do.

This does not mean that their death tallies

were  necessarily  accurate—undercount­

ing  still  plagues  many  countries  with  in­

sufficienttesting(whichiswhyTheEcono-

mistestimates  the  pandemic’s  death  toll

using  excess  deaths).  But  it  does  suggest

that  the  numbers  reported  are  not  being

deliberately tampered with.

Yet data from 17 countries had the oppo­

site  pattern.  In  many  weeks,  the  variance

of  each  distribution  was  less  than  the

mean. This is a statistical smoking gun. “It

seems  reasonable  to  conclude  that  there’s

no  way  these  are  independent  observa­

tions,” says David Steinsaltz, a professor of

statistics at the University of Oxford.

Imputing  motives  is  harder.  A  benign

explanation would be bureaucratic bottle­

necks in processing death certificates. Yet

there  are  other  irregularities:  the  usual

drop­off in weekend reporting is often ab­

sent.  According  to  Mr  Kobak,  the  likelier

explanation is cackhanded tampering.

The Russian numbers offer an example

of  abnormal  neatness.  In  August  2021

dailydeath tallies went no lower than 746

and no higher than 799. Russia’s invariant

numbers  continued  into  the  first  week  of

September,  ranging  from  792  to  799.  A

back­of­the­envelope  calculation  shows

that  such  a  low­variation  week  would

occurby chance once every 2,747 years.n

Abnormal tallies suggest some

countries reported false covid-19 data

→ Reported covid-1 death tallies have an expected amount of variance. But in some countries, variance is abnormally low

New confirmed covid-19 deaths per million people

*Lessthan1%ofexpectedvariation Sources:“Underdispersioninthereportedcovid-19caseand death

numbersmaysuggestdatamanipulations”,byD.Kobak,workingpaper,2022;OurWorldinData;JHU CSSE

Mar

2020

Apr May Jun Jul Aug Sep Oct Nov Dec Jan

2021

Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan

2022

0

5

10

15

20

Brazil

United

States

Russia

Tu r ke y

Genuine death counts vary

from day to day, as people

do not fall ill uniformly

Recent data from Russia are

also suspiciously smooth

Turkish data during the first

year of the pandemic have a

dubious lack of variation

60

65

86

86

29

79

81

67

54

26

28

92

90

69

58

62

80

% abnormal

↓

Mongolia

Cambodia

Russia

Lebanon

Tu r ke y

Albania

Uzbekistan

Algeria

Kyrgyzstan

Syria

Egypt

Venezuela

ElSalvador

Azerbaijan

Belarus

Serbia

Saudi Arabia

Very low*

Weeks with normal variation

Weeks with abnormally low variation