Sick of Data

• Date: 01 January 2018
• DOI: http://doi.org/10.1002/jcaf.22309
• Published date: 01 January 2018

137

© 2018 Wiley Periodicals, Inc.

Published online in Wiley Online Library (wileyonlinelibrary.com).

DOI 10.1002/jcaf.22309

C

o

m

m

e

n

t

a

r

y

Sick of Data

Tim Chartier

Winter is coming. Wor-

ried about catching the

flu? At one time, the

Centers for Disease Control

and Prevention (CDC) was the

hub of information on possible

outbreaks of influenza. Their

warnings came from doctors

reporting new flu cases. When

the CDC’s analysis indicated an

emerging pandemic, the news

was outdated by a week or two.

Today, we can access nearly

real-time information by, as we

often say, Googling it.

Google’s work in predict-

ing flu outbreaks began when

the company examined the 50

million most common search

terms that Americans type.

They also looked at CDC

data between 2003 and 2008.

They searched for correlations

between frequencies in search

queries and the spread of flu

over time and space. A combi-

nation of 45 search terms led

to a strong predictive ability.

Like the CDC, Google could

pinpoint locations of flu out-

breaks. Google’s results were,

however, considerably faster

since their model could begin

to see a pattern before some

search engine users walked

into a doctor’s office or got the

results of a mouth swab. This

epidemiology work by Google

was published in Nature maga-

zine in 2009 claiming they

could “nowcast” the flu. Soon

after, the country entered an

H1N1 crisis. Google’s results

were more accurate than the

CDC and occurred essentially

in real time.

Google has since offered

the tool online, which has

been extended to more than

two dozen other countries.

Simply visit Google Flu

Trends. The site estimates

current outbreak levels in

the United States and can be

refined to examine data at the

state and some city levels.

Google’s ability to essen-

tially don an electronic lab

coat and diagnose H1N1 out-

break levels in 2009 garnered

well-deserved attention. The

2012–2013 flu season sent

Google down a different path,

one of sharply overestimat-

ing the amount of flu. Google

Flu Trends warned that nearly

11% of the population were

infected, when follow-up infor-

mation from the CDC found

no more than 6% had caught

the flu. What went wrong?

Media reports about influenza,

including a state of emergency

declared in New York, led to

more searches that year by

people who weren’t sick.

Google had already antici-

pated how media coverage of

the flu could lead to spikes

in searches by healthy people

for three to seven days. But

this time, flu season actually

was worse than the previous

year, and the searches actu-

ally continued at a higher level

throughout the season. More

media coverage and more flu

meant that even healthy people

did a lot more searching for

“flu” throughout that particu-

lar season. The assumptions

that successfully aided during

the 2009 H1N1 crisis lay at the

core of its failure in 2013.

It’s insightful to focus on

Google response. Seeing the

challenges of their epidemiol-

ogy predictions in 2013, Google

adjusted their algorithm. They

first adjusted for spikes in

searches after media cover-

age. They also refined how

their linear regression throws

out extreme values. With these

changes, their revised method

once again gave an accurate

estimate of flu, within 1% of

what CDC data reported.

The data analytics insight

for this article is less about

Google’s initial success and

much more about its response

to the method’s struggle. Algo-

rithms, like individuals, must be