Hacker News Reader

Comment by fchollet | Hacker News Reader

92 volunteers, 43.5% were infected -> exactly 40 volunteers were infected, and 52 weren't. Then they claim to have found that the infected sample performed "10% worse" on intelligence tests.

Intellectual performance follows a gaussian distribution. If you take two random samples, one of 40 people, and one of 52 people, and you average the IQs in each sample, the probability that you will find a deviation 10% or greater one way to another is over 50%... in this particular case, to compute the probability of the null hypothesis precisely we would need the exact result distribution on the tests they used, but in any case, take these results with a pinch of salt.

View on HN

Where in the world did you learn stats from?

You would be correct if we were talking about a standard normal distribution with mean 0 and standard deviation 1. Very different.

In the case of mean 100 and standard deviation ~15 or so, the probability of two samples being at least 10% different is very small like under 1%.

Very true. I wish the article had the statistics behind this claim:

"The team carried out further tests, in which they injected uninfected and infected green algae into the mouths of mice and put them through a series of lab tests. The results revealed that infected animals took 10% longer to find their way out of mazes and spent 20% less time exploring new objects than uninfected mice."

Which is the much stronger experimental manipulation.

Furthermore 'ATCV-1 exposure in mice also resulted in the altered expression of genes within the hippocampus. These genes comprised pathways related to synaptic plasticity, learning, memory formation, and the immune response to viral exposure.'

http://www.pnas.org/content/111/45/16106.abstract

So it seems like they have opened up multiple avenues of scientific inquiry - this strikes me as a much stronger result than presented in the Newsweek article. Assuming for the sake of argument that their speculation is correct, it's a rather frightening thought that some ~40% of the population could be suffering a significant virally-induced cognitive impairment. Also, I wonder if elimination of this virus could result in an improvement in cognitive and immune function.

The economic and public health implications of this research are potentially quite significant.

That 40% likely applies fairly specifically geographically-speaking. There are other pathologies that degrade cognitive ability which are almost omni-present in developing nations. Hookworm, for instance, causes permanent IQ loss in children who are infected with it and it is nearly universal in Africa and many other places with inferior sanitation. The American US south had a big hookworm problem in the past, but intentional steps were taken to eliminate hookworm. The result was climbing average IQ scores across the south for nearly a century. Malaria also causes cognitive problems, not the least of which is causing missed school in children. Malaria, obviously, sees an awful lot of attention paid to it, but hookworm is not nearly as popular. If I could choose something to focus on, hookworm would probably come even before malaria in my mind. The negative cognitive effects of hookworm hamper everything that can be done to fight malaria and other things, preventing people from realizing how important hygeine and the like is.

There's a comment below with the full text of the study: https://news.ycombinator.com/item?id=8608989

I thought: E[IQ] = 100, SD[IQ] = 10 then SD (for n = 37) = (10 / sqrt(36)) =~ 2

So a 10 point deviation between groups is 5 sd's, p <.001 ?

Guys, we're programmers here not statisticians:

def test():

... n1 = numpy.average(numpy.random.normal(100,15,52))

... n2 = numpy.average(numpy.random.normal(100,15,40))

... return abs(n1-n2) / ((n1+n2)/2.0)

...

>>> vals = [test() for i in xrange(0,1000)]

>>> numOver10 = len([v for v in vals if v > .1])

>>> numOver10

Crucial in your code is the standard deviation of 15. Where is that coming from?

Ah, I see. Standard IQ normalizations.

http://en.wikipedia.org/wiki/Intelligence_quotient

The standard deviation of IQ scores is 15 by definition. (That is, raw scores on the item content of IQ test batteries are generally not reported at all, but rather what is reported for each test-taker is a standard score defined as performance at the norming population median level being an IQ of 100, and performance one standard deviation up from that being an IQ of 115, and so on. All mainstream individual IQ tests have been scored this way for just about all of my lifetime.)

P.S. I would take four standard deviations up or down (the IQ range of 40 to 160) as the outer limit of validated IQ scores, and indeed few modern IQ test batteries purport to report scores outside this range. See

https://en.wikipedia.org/wiki/IQ_classification

and the published reference books cited there for more background.

P.P.S. I should have mentioned when I first posted this comment that any IQ test has a standard error of measurement even if it is faultlessly administered and scored. Sometimes small-n studies like the study reported here have IQ score differences between two groups that are simply the result of the test errors showing a difference where no actual difference exists. IQ tests are often also very frequently administered incorrectly or scored incorrectly even if correctly administered, magnifying the size of the error band around the obtained score. For this study, I'm not sure if the human subjects tested were tested by test-givers who were "blind" to what group the test-takers were in, which could also be an issue.

> All mainstream individual IQ tests have been scored this way for just about all of my lifetime.

Correct me if I'm wrong but I thought the Cattell tests were obviously scored in the Cattell scale (24 s.d.) and the Stanford-Binet tests in the Binet scale (16 s.d.).

When I was into reading about psychometrics it seemed to me that the cut off percentile was a more straightforward way to talk about IQ than using either of those three scales (Wechsler, Binet, Cattell).

The current Stanford-Binet test (and the next one, which is in preparation) use the typical 15 points per standard deviation scoring. The Cattell test is not a "mainstream" IQ test and correlates rather poorly with expected variables that ought to correlate well with IQ like academic achievement. Standard scores by the current definition are readily converted to percentile ranks by look-up tables, by the definition of standard deviation assuming a roughly normal distribution. As above, it's important to note that there is error in any IQ test. I've now had a chance to read the underlying papers related to this thread (which go back at least to 2008), and the IQ tests used in this study are even less mainstream than I feared. "A total of 240 individuals were evaluated with the Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) and the Wisconsin Card Sorting Test (WCST)."

doi:10.1016/j.bbi.2008.04.156

That is true, but we can't translate 10% longer time taken on performing some task into 10% less effective IQ.

No, but there aren't really any good parameters for IQ, and time taken is a damn good, reliable, objective measure.

Yes, but without knowing standard deviations and mean time taken for both groups we can't do math and claim any likelihood of observed result occurring in a random distribution.

Reply on news.ycombinator.com