www.sciencemag.org SCIENCE VOL 343 14 MARCH 2014 1203
POLICYFORUM
In February 2013, Google Flu Trends (GFT) made headlines but not for a reason that Google
executives or the creators of the fl u
tracking system would have hoped.
Nature reported that GFT was predicting more than double the proportion of doctor visits for influenza-like illness (ILI) than the Centers for Disease Control and Prevention (CDC), which bases its estimates on surveillance reports from
laboratories across the United States
(1, 2). This happened despite the fact
that GFT was built to predict CDC
reports. Given that GFT is often held
up as an exemplary use of big data
(3, 4), what lessons can we draw
from this error?
The problems we identify are
not limited to GFT. Research on
whether search or social media can
predict x has become commonplace (5–7) and is often put in sharp contrast
with traditional methods and hypotheses.
Although these studies have shown the
value of these data, we are far from a place
where they can supplant more traditional
methods or theories (8). We explore two
issues that contributed to GFT’s mistakes—
big data hubris and algorithm dynamics—
and offer lessons for moving forward in the
big data age.
Big Data Hubris
“Big data hubris” is the often implicit
assumption that big data are a substitute
for, rather than a supplement to, traditional
data collection and analysis. Elsewhere, we
have asserted that there are enormous scientifi c possibilities in big data (9–11). However, quantity of data does not mean that
one can ignore foundational issues of measurement and construct validity and reliability and dependencies among data (12).
The core challenge is that most big data that
have received popular attention are not the
output of instruments designed to produce
valid and reliable data amenable for scientifi c analysis.
The initial version of GFT was a particularly problematic marriage of big and
small data. Essentially, the methodology
was to fi nd the best matches among 50 million search terms to fit 1152 data points
(13). The odds of fi nding search terms that
match the propensity of the fl u but are structurally unrelated, and so do not predict the
future, were quite high. GFT developers,
in fact, report weeding out seasonal search
terms unrelated to the fl u but strongly correlated to the CDC data, such as those regarding high school basketball (13). This should
have been a warning that the big data were
overfi tting the small number of cases—a
standard concern in data analysis. This ad
hoc method of throwing out peculiar search
terms failed when GFT completely missed
the nonseasonal 2009 infl uenza A–H1N1
pandemic (2, 14). In short, the initial version of GFT was part flu detector, part
winter detector. GFT engineers updated
the algorithm in 2009, and this
model has run ever since, with a
few changes announced in October
2013 (10, 15).
Although not widely reported
until 2013, the new GFT has been
persistently overestimating flu
prevalence for a much longer time.
GFT also missed by a very large
margin in the 2011–2012 fl u season and has missed high for 100 out
of 108 weeks starting with August
2011 (see the graph ). These errors
are not randomly distributed. For
example, last week’s errors predict
this week’s errors (temporal autocorrelation), and the direction and
magnitude of error varies with the
time of year (seasonality). These
patterns mean that GFT overlooks
considerable information that
could be extracted by traditional
statistical methods.
Even after GFT was updated in 2009,
the comparative value of the algorithm as a
stand-alone fl u monitor is questionable. A
study in 2010 demonstrated that GFT accuracy was not much better than a fairly simple projection forward using already available (typically on a 2-week lag) CDC data
(4). The comparison has become even worse
since that time, with lagged models signifi –
cantly outperforming GFT (see the graph).
Even 3-week-old CDC data do a better job
of projecting current flu prevalence than
GFT [see supplementary materials (SM)].
Considering the large number of
approaches that provide inference on infl uenza activity (16–19), does this mean that
the current version of GFT is not useful?
No, greater value can be obtained by combining GFT with other near–real-time
health data (2, 20). For example, by combining GFT and lagged CDC data, as well
as dynamically recalibrating GFT, we can
substantially improve on the performance
of GFT or the CDC alone (see the chart).
This is no substitute for ongoing evaluation
and improvement, but, by incorporating this
information, GFT could have largely healed
itself and would have likely remained out of
the headlines.
The Parable of Google Flu:
Traps in Big Data Analysis
BIG DATA
David Lazer,1, 2* Ryan Kennedy,1, 3, 4 Gary King,3 Alessandro Vespignani3,5,6
Large errors in fl u prediction were largely
avoidable, which offers lessons for the use
of big data.
CREDIT: ADAPTED FROM AXEL KORES/DESIGN & ART DIRECTION/ISTOCKPHOTO.COM
1Lazer Laboratory, Northeastern University, Boston, MA
02115, USA. 2Harvard Kennedy School, Harvard University,
Cambridge, MA 02138, USA. 3Institute for Quantitative Social
Science, Harvard University, Cambridge, MA 02138, USA.
4University of Houston, Houston, TX 77204, USA. 5Laboratory
for the Modeling of Biological and Sociotechnical Systems,
Northeastern University, Boston, MA 02115, USA. 6Institute
for Scientifi c Interchange Foundation, Turin, Italy.
FINAL FINAL FINAL FINAL
*Corresponding author. E-mail: d.lazer@neu.edu.
1204 14 MARCH 2014 VOL 343 SCIENCE www.sciencemag.org
POLICYFORUM
Algorithm Dynamics
All empirical research stands on a foundation of measurement. Is the instrumentation
actually capturing the theoretical construct of
interest? Is measurement stable and comparable across cases and over time? Are measurement errors systematic? At a minimum,
it is quite likely that GFT was an unstable
refl ection of the prevalence of the fl u because
of algorithm dynamics affecting Google’s
search algorithm. Algorithm dynamics are
the changes made by engineers to improve
the commercial service and by consumers in using that service. Several changes in
Google’s search algorithm and user behavior likely affected GFT’s tracking. The most
common explanation for GFT’s error is a
media-stoked panic last fl u season (1, 15).
Although this may have been a factor, it cannot explain why GFT has been missing high
by wide margins for more than 2 years. The
2009 version of GFT has weathered other
media panics related to the fl u, including the
2005–2006 influenza A/H5N1 (“bird flu”)
outbreak and the 2009 A/H1N1 (“swine fl u”)
pandemic. A more likely culprit is changes
made by Google’s search algorithm itself.
The Google search algorithm is not a
static entity—the company is constantly
testing and improving search. For example,
the offi cial Google search blog reported 86
changes in June and July 2012 alone (SM).
Search patterns are the result of thousands of
decisions made by the company’s programmers in various subunits and by millions of
consumers worldwide.
There are multiple challenges to replicating GFT’s original algorithm. GFT has never
documented the 45 search terms used, and
the examples that have been released appear
misleading (14) (SM). Google does provide
a service, Google Correlate, which allows
the user to identify search data that correlate
with a given time series; however, it is limited to national level data, whereas GFT was
developed using correlations at the regional
level (13). The service also fails to return any
of the sample search terms reported in GFTrelated publications (13, 14).
Nonetheless, using Google Correlate to
compare correlated search terms for the GFT
time series to those returned by the CDC’s
data revealed some interesting differences. In
particular, searches for treatments for the fl u
and searches for information on differentiating the cold from the fl u track closely with
GFT’s errors (SM). This points to the possibility that the explanation for changes in relative search behavior is “blue team” dynamics—where the algorithm producing the data
(and thus user utilization) has been modifi ed by the service provider in accordance
with their business model. Google reported
in June 2011 that it had modifi ed its search
results to provide suggested additional search
terms and reported again in February 2012
that it was now returning potential diagnoses
for searches including physical symptoms
like “fever” and “cough” (21, 22). The former recommends searching for treatments
of the fl u in response to general fl u inquiries, and the latter may explain the increase
in some searches to distinguish the fl u from
the common cold. We document several other
changes that may have affected GFT (SM).
In improving its service to customers,
Google is also changing the data-generating
process. Modifications to the search algorithm are presumably implemented so as to
support Google’s business model—for example, in part, by providing users useful information quickly and, in part, to promote more
advertising revenue. Recommended searches,
usually based on what others have searched,
will increase the relative magnitude of certain
searches. Because GFT uses the relative prevalence of search terms in its model, improvements in the search algorithm can adversely
affect GFT’s estimates. Oddly, GFT bakes in
an assumption that relative search volume for
certain terms is statically related to external
events, but search behavior is not just exogenously determined, it is also endogenously
cultivated by the service provider.
Blue team issues are not limited to
Google. Platforms such as Twitter and Facebook are always being re-engineered, and
whether studies conducted even a year ago
on data collected from these platforms can
be replicated in later or earlier periods is an
open question.
Although it does not appear to be an issue
in GFT, scholars should also be aware of the
potential for “red team” attacks on the systems we monitor. Red team dynamics occur
when research subjects (in this case Web
searchers) attempt to manipulate the datagenerating process to meet their own goals,
such as economic or political gain. Twitter
polling is a clear example of these tactics.
Campaigns and companies, aware that news
media are monitoring Twitter, have used
numerous tactics to make sure their candidate
or product is trending (23, 24).
Similar use has been made of Twitter
and Facebook to spread rumors about stock
prices and markets. Ironically, the more successful we become at monitoring the behavior of people using these open sources of
information, the more tempting it will be to
manipulate those signals.
8 6 4 2 0
10
07/01/09 07/01/10 07/01/11
Data
07/01/12 07/01/13
Google Flu
Lagged CDC
Google Flu + CDC
CDC
–50
0
50
100
150
07/01/09 07/01/10 07/01/11 07/01/12 07/01/13
Google Flu Lagged CDC
Google Flu + CDC
Google estimates more
than double CDC estimates
Google starts estimating
high 100 out of 108 weeks
Error (% baseline) % ILI
GFT overestimation. GFT overestimated the prevalence of fl u in the 2012–2013 season and overshot the
actual level in 2011–2012 by more than 50%. From 21 August 2011 to 1 September 2013, GFT reported overly
high fl u prevalence 100 out of 108 weeks. (Top) Estimates of doctor visits for ILI. “Lagged CDC” incorporates
52-week seasonality variables with lagged CDC data. “Google Flu + CDC” combines GFT, lagged CDC estimates,
lagged error of GFT estimates, and 52-week seasonality variables. (Bottom) Error [as a percentage {[Non-CDC
estmate)(CDC estimate)]/(CDC) estimate)}. Both alternative models have much less error than GFT alone.
Mean absolute error (MAE) during the out-of-sample period is 0.486 for GFT, 0.311 for lagged CDC, and 0.232
for combined GFT and CDC. All of these differences are statistically signifi cant at P < 0.05. See SM.
www.sciencemag.org SCIENCE VOL 343 14 MARCH 2014 1205
POLICYFORUM
Transparency, Granularity, and All-Data
The GFT parable is important as a case study
where we can learn critical lessons as we
move forward in the age of big data analysis.
Transparency and Replicability. Replication is a growing concern across the academy. The supporting materials for the GFTrelated papers did not meet emerging community standards. Neither were core search
terms identifi ed nor larger search corpus provided. It is impossible for Google to make its
full arsenal of data available to outsiders, nor
would it be ethically acceptable, given privacy
issues. However, there is no such constraint
regarding the derivative, aggregated data.
Even if one had access to all of Google’s data,
it would be impossible to replicate the analyses of the original paper from the information
provided regarding the analysis. Although it is
laudable that Google developed Google Correlate ostensibly from the concept used for
GFT, the public technology cannot be utilized
to replicate their fi ndings. Clicking the link
titled “match the pattern of actual fl u activity
(this is how we built Google Flu Trends!)” will
not, ironically, produce a replication of the
GFT search terms (14). Oddly, the few search
terms offered in the papers (14) do not seem
to be strongly related with either GFT or the
CDC data (SM)—we surmise that the authors
felt an unarticulated need to cloak the actual
search terms identifi ed.
What is at stake is twofold. First, science
is a cumulative endeavor, and to stand on the
shoulders of giants requires that scientists
be able to continually assess work on which
they are building (25). Second, accumulation of knowledge requires fuel in the form of
data. There is a network of researchers waiting to improve the value of big data projects
and to squeeze more actionable information
out of these types of data. The initial vision
regarding GFT—that producing a more accurate picture of the current prevalence of contagious diseases might allow for life-saving
interventions—is fundamentally correct, and
all analyses suggest that there is indeed valuable signal to be extracted.
Google is a business, but it also holds in
trust data on the desires, thoughts, and the
connections of humanity. Making money
“without doing evil” (paraphrasing Google’s
motto) is not enough when it is feasible to do
so much good. It is also incumbent upon academia to build institutional models to facilitate collaborations with such big data projects—something that is too often missing
now in universities (26).
Use Big Data to Understand the Unknown.
Because a simple lagged model for fl u prevalence will perform so well, there is little room
for improvement on the CDC data for model
projections [this does not apply to other
methods to directly measure fl u prevalence,
e.g., (20, 27, 28)]. If you are 90% of the way
there, at most, you can gain that last 10%.
What is more valuable is to understand the
prevalence of fl u at very local levels, which is
not practical for the CDC to widely produce,
but which, in principle, more fi nely granular
measures of GFT could provide. Such a fi nely
granular view, in turn, would provide powerful input into generative models of fl u propagation and more accurate prediction of the fl u
months ahead of time (29–33).
Study the Algorithm. Twitter, Facebook,
Google, and the Internet more generally are
constantly changing because of the actions
of millions of engineers and consumers.
Researchers need a better understanding of
how these changes occur over time. Scientists need to replicate findings using these
data sources across time and using other data
sources to ensure that they are observing
robust patterns and not evanescent trends. For
example, it is eminently feasible to do controlled experiments with Google, e.g., looking
at how Google search results will differ based
on location and past searches (34). More generally, studying the evolution of socio-technical systems embedded in our societies is
intrinsically important and worthy of study.
The algorithms underlying Google, Twitter,
and Facebook help determine what we fi nd
out about our health, politics, and friends.
It’s Not Just About Size of the Data. There
is a tendency for big data research and more
traditional applied statistics to live in two different realms—aware of each other’s existence but generally not very trusting of each
other. Big data offer enormous possibilities
for understanding human interactions at a
societal scale, with rich spatial and temporal dynamics, and for detecting complex
interactions and nonlinearities among variables. We contend that these are the most
exciting frontiers in studying human behavior. However, traditional “small data” often
offer information that is not contained (or
containable) in big data, and the very factors that have enabled big data are enabling
more traditional data collection. The Internet
has opened the way for improving standard
surveys, experiments, and health reporting
(35). Instead of focusing on a “big data revolution,” perhaps it is time we were focused
on an “all data revolution,” where we recognize that the critical change in the world has
been innovative analytics, using data from all
traditional and new sources, and providing a
deeper, clearer understanding of our world.
References and Notes
1. D. Butler, Nature 494, 155 (2013).
2. D. R. Olson et al., PLOS Comput. Biol. 9, e1003256
(2013).
3. A. McAfee, E. Brynjolfsson, Harv. Bus. Rev. 90, 60 (2012).
4. S. Goel et al., Proc. Natl. Acad. Sci. U.S.A. 107, 17486
(2010).
5. A. Tumasjan et al., in Proceedings of the 4th International
AAAI Conference on Weblogs and Social Media, Atlanta,
Georgia, 11 to 15 July 2010 (Association for Advancement
of Artifi cial Intelligence, 2010), pp. 178.
6. J. Bollen et al., J. Comput. Sci. 2, 1 (2011).
7. F. Ciulla et al., EPJ Data Sci. 1, 8 (2012).
8. P. T. Metaxas et al., in Proceedings of PASSAT—IEEE
Third International Conference on Social Computing,
Boston, MA, 9 to 11 October 2011 (IEEE, 2011), pp. 165;
doi:10.1109/PASSAT/SocialCom.2011.98.
9. D. Lazer et al., Science 323, 721 (2009).
10. A. Vespignani, Science 325, 425 (2009).
11. G. King, Science 331, 719 (2011).
12. D. Boyd, K. Crawford Inform. Commun. Soc. 15, 662
(2012).
13. J. Ginsberg et al., Nature 457, 1012 (2009).
14. S. Cook et al., PLOS ONE 6, e23610 (2011).
15. P. Copeland et al., Int. Soc. Negl. Trop. Dis. 2013, 3
(2013).
16. C. Viboud et al., Am. J. Epidemiol. 158, 996 (2003).
17. W. W. Thompson et al., J. Infect. Dis. 194 (Suppl. 2),
S82–S91 (2006).
18. I. M. Hall et al., Epidemiol. Infect. 135, 372 (2007).
19. J. B. S. Ong et al., PLOS ONE 5, e10036 (2010).
20. J. R. Ortiz et al., PLOS ONE 6, e18687 (2011).
21. Organizing lists of related searches, Google; http://
insidesearch.blogspot.com/2011/06/organizing-lists-ofrelated-searches_16.html.
22. Improving health searches, because your health matters, Google; http://insidesearch.blogspot.com/2012/02/
improving-health-searches-because-your.html.
23. E. Mustafaraj, P. Metaxas, in Proceedings of the WebSci10, Raleigh, NC, 26 and 27 April 2010 (Web Science
Trust, 2010); http://journal.webscience.org/317/.
24. J. Ratkiewicz et al., in Proceedings of 5th International
AAAI Conference on Weblogs and Social Media, San
Francisco, CA, 7 to 11 August 2011 (AAAI, 2011),
p. 297–304.
25. G. King, PS Polit. Sci. Polit. 28, 443 (1995).
26. P. Voosen, Chronicle of Higher Education, 13 September
2013; http://chronicle.com/article/Researchers-Struggleto-Secure/141591/.
27. R. Lazarus et al., BMC Public Health 1, 9 (2001).
28. R. Chunara et al., Online J. Public Health Inform. 5, e133
(2013).
29. D. Balcan et al., Proc. Natl. Acad. Sci. U.S.A. 106, 21484
(2009).
30. D. L. Chao et al., PLOS Comput. Biol. 6, e1000656
(2010).
31. J. Shaman, A. Karspeck, Proc. Natl. Acad. Sci. U.S.A. 109,
20425 (2012).
32. J. Shaman et al., Nat. Commun. 4, 2837 (2013).
33. E. O. Nsoesie et al., PLOS ONE 8, e67164 (2013).
34. A. Hannak et al., in Proceedings of the 22nd International World Wide Web Conference, Rio de Janeiro, 13
to 17 May 2013 (Association for Computing Machinery,
New York, 2013), pp. 527–538.
35. A. J. Berinsky et al., Polit. Anal. 20, 351–368 (2012).
Acknowledgments: This research was funded, in part, by
NSF grant no. 1125095 Army Research Offi ce (ARO) grant
no. W911NF-12-1-0556, and, in part, by the Intelligence
Advanced Research Projects Activity (IARPA) via Department
of Interior National Business Center (DoI/NBC) contract
D12PC00285. The views and conclusions contained herein are
those of the authors and should not be interpreted as necessarily representing the offi cial policies or endorsements, either
expressed or implied, of the NSF ARO/IARPA, DoI/NBE, or the
U.S. government. See SM for data and methods.
Supplementary Materials
www.sciencemag.org/content/343/6176/page/suppl/DC1
10.1126/science.1248506