The Unintended Consequences of Internet Diffusion: Evidence from Malaysia


Can the introduction of the Internet undermine incumbent power in a semi-authoritarian regime? I examine this question using evidence from Malaysia, where the incumbent coalition lost its 40-year monopoly on power in 2008. I develop a novel methodology for measuring Internet penetration, matching IP addresses with physical locations, and apply it to the 2004 to 2008 period in Malaysia. Using distance to the backbone to instrument for endogenous Internet penetration, I find that Internet exposure accounts for 6.6 points, nearly half the swing against the incumbent party in 2008. I find limited evidence of increased turnover, and no evidence of an effect on turnout.

[Latest Draft]


The Political Impact of the Internet on US Presidential Elections


What effect did the Internet have on the 2008 U.S. presidential elections? According to anecdotal evidence, the internet is said to have played a key role: the Obama campaign’s online fundraising arm brought in a record $500 million in small individual donations; and the campaign’s heavy use of social media purportedly contributed to the highest rate of youth turnout since voting was extended to 18-year-olds. We test these assertions exploiting geographic discontinuities along state borders with different right-of-way laws, determining the cost of building new infrastructure. We find that areas with higher internet growth are more likely to swing to the Democratic presidential nominee and are more likely to provide small donations to the Democratic Party.


To Whom Do We Pay Attention when Following the Crowd?


When observing the behavior of players in a sequence, from whom do we learn the most? To answer this question I draw on a meta-dataset of 13 laboratory experiments on social learning. I find that people behave contrary to the predictions of the prevailing models used to understand social learning and information cascades. Under Bayes Nash Equilibrium, the most informative signals come from the first player in the sequence. Under logit QRE a cascade break should be particularly informative, revealing the private signal. However, I find that agents learn most not from the first player but from the decision directly preceding their own. Furthermore, I find no evidence of learning from cascade breaks.