Thread by @UntergrundmannG

So, I've managed to get a bit of time on my hands now and am taking a deeper look at this much quoted @jonatanpallesen piece on "debunking" @nntaleb's claims on IQ ...

There's this section which looks rather off ... +

I think the contention is ill-understood. Part of what I understand from @nntaleb's position is that performance is NOT "normally" distributed. Whereas IQ is taken to be so.

And so, @jonatanpallesen's claim here of "disproving" NNT looks off-target.

More as I process this ...

@nntaleb @jonatanpallesen And this entire section is built off a misunderstood position/irrelevant model. One can eliminate this out of hand w/o processing it at all.

+

@nntaleb @jonatanpallesen Another section that has an issue.

Average income is meaningless in case starting income is $40k in Y1, growing to $100k in Y6 going down to income ZERO in Y9 and Y10.

If your prediction is that IQ maps to some income at the END of Y10, use just that single value.

@nntaleb @jonatanpallesen Before logging off for today, highighting this image that a lot of the IQ folks share around as "proof" of the value of IQ.

I'm curious how this visualisation looks with the axes swapped.

Will pick this up later +

@nntaleb @jonatanpallesen Spent a bit of time with the WLS Longform Data - identifying the field labels of interest and otherwise.

Should be able to work with the IQ - income/success relations over the weekend.

@nntaleb @jonatanpallesen Did not get much time, but managed to get some basic data into a worksheet, filtered the IQ and a couple other NAs. Ran a rudimentary EXCEL correlation check ... and this is what it looks like. (The blue font includes spouse contributions as well)

+

@nntaleb @jonatanpallesen Regressed the earliest recorded parent income of the respondents to their average parental income ... +

@nntaleb @jonatanpallesen Regressed the IQ against both - earliest recorded parent income and average parent income to get this ... +

@nntaleb @jonatanpallesen Instead of going straight for income, I figured it may make sense to see how the Net Worth of the WLS respondent regresses against IQ only and then a combination of IQ and Avg Parent Income ... +

@nntaleb @jonatanpallesen And just to close that bit off, regressed Net Worth against Average Parental Income ... +

@nntaleb @jonatanpallesen The above makes sense with the CORR table in the sense that the Avg Parent Income correlates higher with Net Worth than does IQ +

@nntaleb @jonatanpallesen And this time I regressed Total Personal Income against Average Parent Income and a combination of IQ and Avg Parent Income ... +

@nntaleb @jonatanpallesen Stopping here for now ... closed the last step regressing Total Personal Income against IQ. Not interpreting anything at the moment. Noting the R^2 value ~ 3%.

This exercise needs to be repeated with clipping the IQ at 70/100 ("competence"/"mean") levels - for changes.

+

@nntaleb @jonatanpallesen Separated the data into three bins - IQ <70, 70<Q<100 and IQ>100. Ran the correlations again ... +

@nntaleb @jonatanpallesen Even if one is willing to treat this exercise as rudimentary, there isn't a lot that we can infer from this ... except acknowledge that the negative CORRs between IQ and TPI for IQ<70 make sense in line with what we already know - there's serious cognitive impairment at play.
+

@nntaleb @jonatanpallesen Running through the cycle again ... Regressed IQ (100+) against Parent Income and Earliest Known Parent Income. Make a note of the intercept and the p-Values ... +

@nntaleb @jonatanpallesen Ditto as above for IQ 70-100

+

@nntaleb @jonatanpallesen Ditto for IQ below 70. The cutoff at 60 explains itself as well as the small sample size.

+

@nntaleb @jonatanpallesen Regressed the Total Personal Income against IQ (below 70) and Avg Parent Income and Earliest Recorded Parent Income ... +

@nntaleb @jonatanpallesen Ditto as above for Net Worth against the same three factors (IQ below 70) ... +

@nntaleb @jonatanpallesen As above for Net Worth (70<IQ<100)
+

@nntaleb @jonatanpallesen As above for Total Personal Income (70<IQ<100)

+

@nntaleb @jonatanpallesen Net Worth for the above average IQ case (>100) ... +

@nntaleb @jonatanpallesen TPI for the above average IQ case (>100) ... +

@nntaleb @jonatanpallesen So, where are we now? There isn't much to infer here and the explanatory power (via R_Sq value) is rather low to take seriously.

What I want to do now is ignore the data for IQ below 70 and check for just IQ and parental income - singly - regressed for Net Worth and TPI +

Separately for below and above 100. Data points in either case are near equal. The idea is to check if there's any striking difference between the two cases.

Next would be to check for IQ and NW/TPI for cases which started/own a business and otherwise. And lastly, use grades.
+

Regresions for the IQ > 100 data. If you choose Total Personal Income as a measure of 'success', you find that IQ and parental income are comparable as indicators. If you use the estimated Net Worth at the same point of time, parental income 'explains' 3x of what IQ does.
+

Caveat: The professor who took our introductory business analytics class made it a point to emphasize that most of this is meaningless at such low R_Sq values. The residuals also indicate the amount of data that is UNEXPLAINED by these parameters

+

This is the case with 70<IQ<100. There isn't much to go by - towards predicting any kind of "success". And arguably, Net Worth at any given point of time, is a more valuable indicator of success.

+

So, I figured if it may make sense to see how IQ is distributed for different income intervals in the WLS data. This is what it looks like, eliminating the NAs etc.

The stability of the distribution, esp the STDEV and MODE, is rather striking...

+

...which raises the Q if it is sensible to assert that IQ predicts "income success" when the drift in MEAN/MEDIAN IQ remains within the range of IQ test measurement error.

I doubt any serious scientist would make any assertive claim on the predictive validity of IQ in this case.