Joshua Loftus

A letter, signed by over 800 scientists and published in Nature called for an end to using p-values to decide whether data refutes or supports a scientific hypothesis. The letter has received widespread coverage and reignited an old debate. […] Most of the objections to p-values or the p < 0.05 threshold in these articles can be summarized into two categories: […] Banning p-values or “p < 0.05” thresholds wouldn’t address these objections. We will still have to make decisions, we can’t just report a Bayes factor (or a p-value) and refuse to decide whether a drug trial should continue or not. So our decisions will still sometimes be wrong, and in both directions. […] The last kind of objection is more sensible–though less often the … [Read More]

A brief post describing counterfactual privilege, a notion we use in a new algorithmic fairness paper to design optimal fair interventions. Links the paper, slides, and so on are included, along with a simple example to illustrate the basic idea. [Read More]

(Note: links don’t work in this preview, click on the post to view). I’m happy to be speaking at 1pm EST today at Columbia University on the topics of causal inference and selection bias in algorithmic fairness. I believe video will be available at the webinar link, and here are my slides. The talk is based on work described in this survey with my coauthors Matt Kusner, Chris Russell, and Ricardo Silva. See here for the video for Matt’s oral presentation of our first paper in this line of work at NIPS 2017. [Read More]

You may have seen stories about Twitter accounts operated by Russians attempting to influence the 2016 election in the United States. Much of the reporting that I’ve seen described a Simple Narrative: Russians tried to help trump and hurt Clinton, even supporting Bernie Sanders in order to attack Clinton. I’ve also seen plenty of Democrats on Twitter attacking Sanders over this. I have not seen any stories reporting the fact that many of these bots also attacked Sanders. (If you’re aware of such stories, I’d be happy to hear about them). […] Recently, FiveThirtyEight published a repository with roughly 3 million tweets from about 2,800 accounts that Twitter concluded were associated with the Russian effort. About the data: […] The data set is the work of two professors at … [Read More]

Model selection can invalidate inference, such as significance tests, but statisticians have recently made progress developing methods to adjust for this bias. One approach uses conditional probability, adjusting inferences by conditioning on selecting the chosen model. This post motivates the conditional approach with a simple screening rule example and introduces the selectiveInference R package that can compute adjusted significance tests after popular model selection methods like forward stepwise and LASSO. [Read More]

This post describes a simple example that illustrates why algorithmic fairness is a hard problem. I claim it is at least as hard as doing causal inference from observational data, i.e. distinguishing between mere association and actual causation. In the process, we will also see that SCOTUS Chief Justice Roberts has a mathematically incorrect theory on how to stop discrimination. Unfortunately, that theory persists as one of the most common constraints on fairness. [Read More]

People often do regression model selection, either by hand or using algorithms like forward stepwise or the lasso. Sometimes they also report significance tests for the variables in the chosen model. After all, a significant p-value means they’ve found something real. But there’s a problem: the reason for that significant p-value may just be something called model selection bias. This bias can invalidate inferences done after model selection, and may be one of the contributors to the reproducibility crisis in science. Adjusting inference methods to account for model selection is an area of ongoing research where I have done some work. [Read More]

A letter, signed by over 800 scientists and published in Nature called for an end to using p-values to decide whether data refutes or supports a scientific hypothesis. The letter has received widespread coverage and reignited an old debate. […] Most of the objections to p-values or the p < 0.05 threshold in these articles can be summarized into two categories: […] Banning p-values or “p < 0.05” thresholds wouldn’t address these objections. We will still have to make decisions, we can’t just report a Bayes factor (or a p-value) and refuse to decide whether a drug trial should continue or not. So our decisions will still sometimes be wrong, and in both directions. […] The last kind of objection is more sensible–though less often the focal point of this debate–and I … [Read More]