Published on Development Impact

Weekly links September 18: better multiple testing, surveying women, fast and slow p-hacking, unexpected effects of inequality, and more…

This page in:

·       Julian Reif has updated his wyoung Stata command to allow for multiple treatments, and so now this command covers all the key issues that I had outlined in my earlier blog post comparing multiple hypothesis testing commands. I’ve now updated the blog post to reflect this.

·       On the World Bank’s Data Blog, Erwin Knippenberg and Moritz Meyer summarize ongoing work using call detail record (CDR) data to track internal mobility in the Gambia at a weekly level during different stages of the Covid-19 lockdown and Ramadan. Also on the Data Blog, Legovini et al. on their program of work to improve road safety data in Nairobi (with lots of cool maps).

·       The IDInsight blog discusses lessons from trying to survey women by phone in rural India: asking a few questions to the male household head first before asking to speak to a women made it more likely they would pass the phone over than asking to speak to the woman straight away; female surveyors were more likely to be able to complete surveys with women than male surveyors.

·       Oxford’s Mind and Behavior group has a set of “Methods matter” notes covering topics like how to measure risk attitudes, time preferences, and income expectations.

·       On the Stata blog, part 5 of using python in Stata illustrates how to do three-dimensional surface plots of marginal predictions.

·       Datacolada has an interesting post on “fast” and “slow” p-hacking, and how much we expect it to change p-values. The claim is that observational data is more likely to be subject to “slow” p-hacking, where researchers make a series of choices over observations to trim, transformations, controls to include, etc. that slowly change a p-value, so that if it gets under 0.05, it will just get under this; whereas (and this is the unsupported part) experiments may be subject to “fast” p-hacking, where the p-value changes dramatically with choices (e.g. by deciding on which outcomes to look at, or looking at subgroups, or bigger changes – so that if the p-value gets under 0.05, it need not be clumped at 0.049).

·       Finally, for your weekend fun, the IgNobel prizes were awarded yesterday. As always, some superb entries. The economics prize goes to work showing a surprising “effect” of income inequality.


David McKenzie

Lead Economist, Development Research Group, World Bank

Join the Conversation

The content of this field is kept private and will not be shown publicly
Remaining characters: 1000