Hillary won the election by nearly 3M votes.
She lost the election due to Pootin/Assange strategy of dribbling out emails over many weeks, and
above due to Comey trashing her last summer on a false document and
then again days before the election
which Silver says caused a clear shift to Trash in three key states and their large electoral votes, where she lost each by tiny fractions of a percent.
As Dems always do, Hillary won the women's and black vote by a huge margin, while Trash won the men's vote. All are "identity politics".Looks like Google influenced the election much more than "Russia"
A Method for Detecting Bias in Search Rankings, with Evidence
of Systematic Bias Related to the 2016 Presidential Election
Robert Epstein (
[email protected])
American Institute for Behavioral Research and Technology
Ronald E. Robertson
Northeastern University
Samantha J. Shepherd & Shu Zhang
American Institute for Behavioral Research and Technology
Summary
In a report published in the Proceedings of the National Academy of Sciences summarizing the
results of five randomized, controlled experiments with 4,556 participants in two countries,
Epstein and Robertson (2015) showed that search results favoring one political candidate can
shift the voting preferences of undecided voters substantially—up to 80% in some demographic
groups—and that this effect can easily be masked so that few or no people are aware they are
viewing biased search results. These and other experiments demonstrating the impact of the
Search Engine Manipulation Effect (SEME) beg the question, however, of whether real search
rankings are actually biased.
We now describe a method for creating a Nielsen-ratings-type network of confidants whose daily
online searches are automatically collected and pooled in a manner that allows for the detection
of actual bias in search rankings. The method was implemented in the spring of 2016 in
anticipation of the November presidential election. Ninety-five people from 24 U.S. states (mean
age 39.9) were recruited, 21 of whom identified themselves as “undecided.” In the months
leading up to Election Day, daily election-related searches automatically provided us with the
first page of results from a total of 13,207 searches conducted using Google, Bing and Yahoo
through the Firefox browser, allowing us to preserve 98,044 election-related web pages. Those
pages, in turn, were rated by people recruited from a crowdsourcing website, and the ratings
allowed us to compute the average bias per search position, as well as the overall average bias.
We are still analyzing this wealth of data—all of which will be made public by the Internet
Archive (http://archive.org) later in 2017—but so far we have found that between May and
November 2016, search results displayed in response to a wide range of election-related search
terms were, on average, biased in Mrs. Clinton’s favor in all 10 search-result positions. This bias
could not be accounted for by the bias in the search terms themselves. We also found different
levels of bias in different search engines, as well as evidence of demographically-targeted bias.
We don’t know what caused these patterns of bias, but no matter what the cause or causes, given
the power of search rankings to shift votes and opinions without people’s awareness
(http://bit.ly/1REqzEY), they are a matter for concern.
We believe our new tracking system can serve as a prototype for the development of a
worldwide ecosystem of software that passively monitors search results, search suggestions, news feeds, advertisements and other normally ephemeral internet content. Such a system might
someday play an important role in protecting the integrity of the free-and-fair election. It might
also bring, for the first time, some degree of accountability to the Big Tech companies that
control ephemeral content.
Selected Supporting Data
1) Issue: Were search results provided by search engines in the U.S. biased toward one
candidate or the other? Yes. Based on a sample of 4,045 election-related searches conducted
during a 25-day period from October 15 to November 8 (Election Day) using the Google and
Yahoo search engines through the Firefox browser, we found that search results were, on
average, biased to favor Hillary Clinton on all of those days. (Note: In the graph below, values
above 0 show a Clinton bias, and values below 0 show a Trump bias.)
http://aibrt.org/downloads/EPSTEIN_et_al_2017-SUMMARY-WPA-A_Method_for_Detecting_Bias_in_Search_Rankings.pdf
2) Issue: On the first page of search results (10 positions), where was the bias expressed? We
found that the bias was expressed in all 10 search results:
http://aibrt.org/downloads/EPSTEIN_et_al_2017-SUMMARY-WPA-A_Method_for_Detecting_Bias_in_Search_Rankings.pdf
3) Issue: Was the bias the same for all search engines? No. The level of pro-Clinton bias we
found on Google (0.19) was more than twice as high as the level of pro-Clinton bias we found on
Yahoo (0.09). The difference between these values was highly statistically significant (p <
0.001).
4) Issue: Did everyone see the same level of bias? No. We found evidence of demographic
targeting. For example:
a. Decided vs. undecided. Search results seen by people who said they had decided how to vote
were nearly twice as biased in favor of Clinton (0.21) as search results seen by people who said
they were undecided (0.11, p < 0.001).
b. Men vs. women. Search results seen by men were twice as biased in favor of Clinton (0.24) as
search results seen by women (0.12, p < 0.001).
c. Blue vs. red vs. swing states. People in blue states generally saw the highest level of proClinton
bias in search rankings (0.24); people in red states saw the next highest level of proClinton
bias (0.12); and people in swing states saw the lowest level of pro-Clinton bias (0.10, p <
0.001).
d. Young vs. old. Search results seen by people under age 35 were more than twice as biased in
favor of Clinton (0.21) as search results seen by people 35 and over (0.10, p < 0.001).
Interpretation: Our study doesn’t look at this issue directly, but the pattern of demographic
differences we found is somewhat consistent with the idea that Big Tech companies show people
what they want to see. Note, however, that people in blue (that is, pro-Clinton) states weren’t the
only ones to see pro-Clinton search results; people in red (that is, pro-Trump) and swing states
did too.
5) Issue: Could search results have been biased simply because people were selecting biased
search terms? On a scale from -5 (pro-Trump) to +5 (pro-Clinton), the average bias in people’s
search terms was slightly pro-Trump (-0.08). The search terms people used should therefore
have yielded a pro-Trump bias in search results, but they did not.
6) Issue: Was there bias before October 15? Yes, although we typically received data from only
a few searches per day before that date, so we are less certain of the numbers. Looking at data
from 1,050 searches conducted between May 19 and October 14, 2016, we found, on average, a
pro-Clinton bias throughout this period (0.17), as well as a pro-Clinton bias in all 10 search
positions on the first page of search results. To put this another way, we found evidence of a proClinton
bias in search rankings over a period of nearly six months before the election.
7) Issue: Did we use all the data we collected? No, for two reasons:
a. First, although we started collecting data on May 19, 2016, it took time for us to recruit
confidants for our new network, as well as to refine our new software and procedures. Beginning
on October 15, we began reliably to receive more than 50 searches per day, reaching a peak of
483 searches on November 8 (Election Day). We therefore have focused our analysis on this 25-
day period.
b. Second, we ultimately had to discard all the data we had received from confidants who
communicated with us using gmail, the email service owned by Google. Some of these data were
transmitted to us in spurts, with dozens of searches being conducted in a few seconds by the
same person; such searches were presumably automated. Some of the statistics these data yielded
were also suspect. Compare the pattern of bias values we obtained from non-gmail users of the
Google search engine (graph on the left) to the pattern of bias we obtained from gmail users of
the Google search engine (graph on the right):
http://aibrt.org/downloads/EPSTEIN_et_al_2017-SUMMARY-WPA-A_Method_for_Detecting_Bias_in_Search_Rankings.pdf
As you can see, the search results seen by non-gmail users were far more biased (0.19) than the
results seen by gmail users (0.03, p < 0.001). Perhaps Google identified our confidants through
its gmail system and targeted them to receive unbiased results; we have no way to confirm this at
present, but it is a plausible explanation for the pattern of results we found. (Note: We are
missing data on October 25, 26, and 27 because our server was unable to keep up with the
rapidly increasing influx of data. We may be able to recover some of the missing data, but it is
unlikely. We made adjustments to our software during this period so our server could
accommodate a much faster data flow.)
8) Issue: What happened to the bias after the election? It appeared to decrease. The graph below
compares the bias observed during the 10 days leading up to the election to the bias observed
during the 10 days following the election:
http://aibrt.org/downloads/EPSTEIN_et_al_2017-SUMMARY-WPA-A_Method_for_Detecting_Bias_in_Search_Rankings.pdf
9) Issue: Could the pro-Clinton bias in search results have shifted votes to Mrs. Clinton? A
comprehensive study published in 2015 in the Proceedings of the National Academy of Sciences
found that biased search rankings can easily shift the voting preferences of undecided voters by
20% or more – up to 80% in some demographic groups. Extrapolating from the mathematics
introduced in this report, in articles published in February 2016 and thereafter, the lead author of
the PNAS study predicted that a pro-Clinton bias in Google’s search results would, over time,
shift at least 2.6 million votes to Clinton. She won the popular vote in the November election by
2,864,974 votes. Without the pro-Clinton bias in Google’s search results, her win margin in the
popular vote might have been negligible.