Monthly Archives: October 2017

Recommended: When the revolution came for Amy Cuddy

This page is moving to a new website.

This is one of the best articles I have ever read in the popular press about the complexities of the research process.

This article by Susan Dominus covers some high profile research by Amy Cuddy. She and two co-authors found that your body language not only influences how others view you, but it influences how you view yourself. Striking a “power pose” meaning something like a “legs astride or feet up on a desk” can improve your sense of power and control and these subjective feelings are matched by physiological changes, Your testosterone goes up and your cortisol goes down. Both of these, apparently, are good things.

The research team publishes these findings in Psychological Science, a prominent journal in this field. The article receives a lot of press coverage. Dr. Cuddy becomes the public face of this research, most notably by garnering an invitation to give a TED talk and does a bang-up job. Her talk becomes the second most viewed TED talk of all time.

But there’s a problem. The results of the Psychological Science publication do not get replicated. One of the other two authors expresses doubt about the original research findings. Another research team reviews the data analysis and labels the work “p-hacking”.

The term “p-hacking” is fairly new, but other terms, like “data dredging” and “fishing expedition” have been around for a lot longer. There’s a quote attributed to the economist Robert Coase that is commonly cited in this context, “If you torture the data long enough, it will confess to anything.” I have described it as “running ten tests and then picking the one with the smallest p-value.” Also relevant is this XKCD cartoon.

If p-hacking is a real thing (and there’s some debate about that), then it is a lot more subtle than the quotes and cartoon mentioned above. You can find serious and detailed explanations at a FiveThirtyEight web article by Christie Aschwanden and this 2015 PLOS article by Megan Head et al.

If p-hacking is a problem, then how do you fix it? It turns out that there is a movement in the research world to critically examine existing research findings and to see if the data truly supports the conclusions that have been made. Are the people leading this movement noble warriors for truth or are they shameless bullies who tear down peer-reviewed research in non-peer-reviewed blogs?

I vote for “noble warriors” but read the article and decide for yourself what you think. It’s a complicated area and every perspective has more than one side to it.

One of the noble warriors/shameless bullies is Andrew Gelman, a popular statistician and social scientist. He comments extensively about the New York Times article on his blog, which is also worth reading as well as many comments that others have made on his blog post. It’s also worth digging up some of his earlier commentary about Dr. Cuddy. Continue reading

PMean: The unthinking approach to borderline p-values

I ran across a nice discussion of how to write the results section of a research paper, but it has one comment about the phrase “trend towards significance” that I had to disagree with. So I wrote a comment that they may or may not end up publishing (note: it did look like the published my comment, but it’s a bit tricky to find).

Here’s what I submitted. Continue reading

PMean: 100 interview questions? A big oops on the very first one.

I shouldn’t do this, because we’ve all made mistakes, especially me. But I took a peek at a website with the intriguing title “100+ commonly asked data science interview questions” with the thought “Maybe I could be a data scientist”. But the author of this list choked on the very first question. It’s interesting to examining why the question is bad. Continue reading

Recommended: Search for unpublished data by systematic reviewers: an audit

The authors looked at all systematic reviews (excluding methodological reviews) published in a few key journals as well as a random sample of Cochrane reviews to see how often the authors tried to search for unpublised data. The answer is not often enough (64% or 130/203). The article also describes the success rate in getting unpublished data when the attempt was made (89% or 116/130) and how often authors found evidence of publication bias when they did such an assessment (40% or 27/68). Although some people have argued that it is not that important to search for unpublised data, this is still a big concern. A closely related article is Searching for unpublished data for Cochrane reviews: cross sectional study. Continue reading

PMean: What does large mean when talking about negative values?

Dear Professor Mean, I saw a paper where the authors said that they wanted a diagnostic test with a large negative likelihood ratio, because it was important to rule out a condition. False negatives mean leaving a high risk condition untreated. But don’t they mean that they want a diagnostic test with a small likelihood ratio?

Okay, I agree with you, but it’s an understandable mistake. Let’s quickly review the idea of likelihood ratios. A positive likelihood ratio is defined at Sn / (1-Sp) where Sn is the sensitivity of the diagnostic test and Sp is the specificity. For a diagnostic test with a very high specificity, you get a very large ratio, because you are putting a really small value in the denominator. For Sp=0.99, for example, you would end up getting a positive likelihood ratio of 50 or more (assuming that Sn is at least 0.5).

The positive likelihood ratio is a measure of how much the odds of disease are increased if the diagnostic test is positive.

A negative likelihood ratio is defined as as (1-Sn) / Sp. For a diagnostic test with a very large sensitivity, the negative likelihood ratio is very close to zero. For Sn=0.99, the likelihood ratio is going to be 0.02 or smaller, assuming that Sp is at least 0.5.

The negative likelihood ratio is a measure of how much the odds of disease are decreased if the diagnostic test is negative.

The two likelihood ratios should remind you of the acronyms SpIn and SnOut. SpIn means that if specificity is large, then a positive diagnostic test is good at ruling in the disease. This isn’t always the case, sadly, and for many diagnostic tests, the next step after a positive test is not to treat the disease, but to double check things using a more expensive or more invasive test.

SnNout means that if the sensitivity is large, then a negative diagnostic test is good at ruling out the disease. You can safely send the patient home in some settings, or start looking for other diseases in different settings.

That sounds great, but sometimes you are very concerned about false negatives, and you don’t want to send someone home if they actually have the disease. If you are worried about a cervical fracture, ruling out the fracture and sending someone home might lead to paralysis or death if you have a false negative. So you want to be very sure of yourself in this setting.

Now with regard to the comment above, I think it is just a case of careless language. When the authors say “large negative likelihood ratio”, they should have said “extreme negative likelihood ratio” meaning a likelihood ratio much much smaller than one. I’ve done it myself when I talk about a correlation of -0.8 as being a “big” correlation because it is very far away from zero.

We tend to shy away from words like “small” when we talk about a negative likelihood ratio being much less than 1, because “small” in some people’s minds means “inconsequential” when the opposite is true. When I am careful in my language, I try to use the word “extreme” to mean very far away from the null value (1 for a likelihood ratio or 0 for a correlation) rather than “large” or “small”.