Big Data, Plainly Spoken (aka Numbers Rule Your World)

Wage inequality and its sources

In the sister blog, I featured a data graphic showing the difference in pay levels between white male workers in the U.S. and women workers of various races. That post discusses purely visual matters, in effect accepting the analysis as given. In this post, I take a deeper look at the underlying data analysis.

Let us review the analysis strategy behind the chart, and then discuss why this simple strategy is not particularly insightful.

The starting point of the analyst is income data collected by the Census Bureau. Even here, income is measured in several ways. It appears that using personal income (as opposed to household or family income) is the most appropriate here. There is an additional complication of how to handle "missing values", which may arise in this context because someone is not employed and thus earn zero income. When one says "median income", does one include or exclude the zero earners? What about those who only worked part of the year?

After those questions are addressed, one can work with median incomes, computed at the right level of aggregation. The report linked to by Business Insider only contains aggregation by race, and aggregation by gender, each calculated separately. What is needed is a cross-tabulation. It is not possible to obtain the median income of Asian women from the median income of Asians and the median income of women - unless the analyst makes unwarranted assumptions.

So the analyst learns that white men made $46,000 in 2017 while Asian women made $27,000 and Hispanic women made $21,000. This data can be plotted directly, or after computing the race-gender discount (off the white male wage level).

***

The analyst wants to make this data come alive by using a different unit, the number of days worked.

One way to achieve this is by converting the annual salary to daily salary, then computing how many days the median Asian woman must work in order to earn $46,000, the median for the white man. This is roughly 611 days, which suggests that the Asian woman must work 246 extra days.

In the above description, I have seamlessly papered over several annoying details. I have assumed - without checking with reality - that everyone works 365 days a year. In fact, no one worked 365 days a year. And even if I obtained the correct value of the average number of days worked, call it X, it will also be the case that X will vary between race and gender. Thus, I made a further implicit assumption that such variance is not large enough to bother about.

I justify this lack of care because I rounded the median salaries to the nearest thousand. The questions I raised above ought to concern those analysts who insist on printing estimates in decimals. Is it possible to attain that level of precision while making simplifying assumptions?

***

Also notice that the analysis strategy is counterfactual in nature - it requires conjuring up a hypothetical scenario. It's a comparison of a white man who works exactly one year, and a woman (of any race) who works till she earns $46,000 or whatever is the current median wage for white men.

The notion of "extra days" is an invention since there are only 365 days in a year, and the women can never catch up unless the white guys stop working.

***

When comparing white men and Asian women, both gender and race are shifted. The Business Insider story leads readers to attribute the wage gap to primarily gender but unless we see the median salaries for Asian men, Hispanic men and black men, one can't be sure.

Most likely, there is a gender "effect" as well as a race "effect". The gender effect may even be differently sized for each race. This is known as an "interaction" effect.

***

Finally, there are even more factors to be considered. It is well known that at least some of the wage gap is explained by the difference in the mix of jobs and industries that men and women tend to be employed in. So one can't conclude discrimination without further investigation.

Unequal pay for equal work is discrimination but unequal pay for unequal work need not be.

***

Now, check out my comments about the calendar visualization of the wage gaps by Business Insider.

 

Questionable reviews on Amazon

Isn't it rather suspicious when you're shopping for standard white notepads that you run into thousands of reviews of them on Amazon?

This is what I'm talking about:

This popular item on Amazon has amassed over 1,000 reviews. What could these people be saying? What could be worth saying about stacks of blank white paper with lines on them? A low-cost, disposable item that will be thrown away during consumption? When you learn that there are already 200 essays about this one notepad, do you feel a need to add to the pile? Is there anything original left to be said?

The natural question is how many of these reviews are fake - paid for by sellers? In my latest video, I talk about how fake reviews are made on Amazon, including why reviews labeled as "verified purchases" are just as fake.

 

 

What's wrong with this data?

I've talked about "fake data" before. A lot of fake data come from people trying to game algorithms or skew metrics, and oftentimes, automated bots are involved. Attempts to obscure these tactics typically involve creating layers of complexity so it is not easy to connect the dots.

I come across suspicious data all the time, and it's not always clear what's going on or why. So I thought I'd feature some of them here and see if anyone can figure it out.

***

This account on Yelp caught my attention because this user apparently uploaded a photo of her desktop to one of the restaurant pages. She has uploaded a total of four photos, all of them are unrelated to food. The four photos were uploaded to two New York City restaurants while she indicated that she lives in San Francisco. She did not review either of those NYC restaurants but she has written one review for a cafe in Long Island City. The review seems genuine (although it's hard to tell unless you've been to that cafe).

She has five friends. While she lives in San Francisco, these friends live in Manhattan, Brooklyn, Scottsdale and Oceanside. She has no friends in the Bay Area. None of these friends have ever written a single review and have no likes. However, each of these friends have 100-400 friends. It's not clear why one would be friends with someone on Yelp who has no reviews or likes.

***

Is this account fake? If so, why was it created? How did those photos get uploaded? How did they get placed in those particular restaurants? Who are these friends? Are they fake as well? If the account is fake, was that review also fake? Is it possible to predict that the review is fake?

So many questions, and so hard to get answers. What do you think is going on?

 

 

The unsung heroes of AI

In the latest video, I introduce the unsung heroes behind the AI tech that drives those popular smart home assistants, like Alexa and Google Home. These are the minimum-wage workers in India, China, and other places with cheap labor who are tasked with listening to the recordings from your home conversations.

This practice is raising a lot of concern and controversy. Amazon, Apple and Google have all announced that they are suspending the listening programs. Nevertheless, this hiatus is likely temporary because the machine learning methods behind such tech are extremely data-greedy, and must be trained with human input.

For more, watch the clip.

 

For further reading:

Apple contractors hear confidential details on Siri

Amazon holds on to the data even when you delete audio files

Google angry after contractor leaked recordings

Apple and Google temporarily stop listening

Amazon allows opt-out of human review

More Grubhub theatrics, with Yelp as supporting actor

Just recently, I made a short clip about how Grubhub is extracting lead-generation fees from unsuspecting restaurant owners by setting up a network of shadow websites (and phone numbers). Diners who thought they were ordering directly from restaurants were instead shuffled through the Grubhub toll booth. Click here to see how this works.

 

 

Now, Vice discovered yet another of Grubhub's toll booths. This time, the toll booth is set up on the Yelp superhighway. When a Yelp user clicks on the restaurant's phone number on its Yelp page, a pop-up shows up, with two options, one of which redirects the user to a shadow number owned by Grubhub. These numbers are not labelled "direct" and "Grubhub" but "general questions" and "deliveries and orders". So now both Yelp and Grubhub are making lead-generation money off the restaurant owner.

***

This dispute is about causality! And because causality is tough to establish, it creates a gray zone of disagreement.

Ideally, the restaurant owner pays for orders caused by Grubhub's marketing activities. By cause, we mean the orders would not have materialized without Grubhub's marketing. In the examples we saw, it's highly unlikely that Grubhub did anything to cause those orders. That's because those diners thought they were ordering directly from the restaurants - they are quietly re-routed to the shadow websites and phone numbers set up by Grubhub, sometimes without even the restaurants' knowledge.

It is the secrecy that gives the game away. If Grubhub's causal value is clear, it can form open partnerships with the restaurants without resorting to trickery.

As I pointed out in the video, the majority of the digital marketing industry relies on similar tactics. Search engines do not typically send you directly to the webpage you clicked on - you are often rerouted through the search engine's server, so that the search engine can "track" the click and use the paper trail to receive lead-generation money. Anyone passing through this toll booth is counted as "causal" but in reality, many of these users would have found their way to the webpage even if the search engine didn't exist. (Consider going directly to Macys.com instead of typing Macys into a search engine.)

***

This is a great example of how data, algorithms and software are silently running our lives, and often to the detriment of those who don't understand what's going on. Our video series is a small effort to help you stay in front of these data-driven technologies.

The more you know, the more you can leverage its power, and avoid its harms.

Where is the app that makes one look younger?

That was the question I asked when first hearing about the mobile app that ages you 30 years. The FaceApp went viral, and became even better known when some politicians proclaimed it unsafe because it was run by Russians.

It turns out FaceApp also allows users to make them look younger. It seems that the aged photos are more likely to be shared than the make-young photos. Why is that?

In my new 2-minute clip, I give an answer to this question.

***

The issue raised in the clip is important for any machine-learning developer. Depending on your application, the errors made by the algorithm may be more or less visible to the users.

I discussed this aspect of decision-making in Chapter 4 of Numbers Rule Your World (link), the chapter about steroids testing. When the lab makes a false positive error, the falsely accused athlete will complain very loudly. When the lab commits a false negative, the cheating athlete will not make a sound. This creates a situation in which the lab's false-positive errors are highly visible while almost all false-negative errors are hidden from view.

It was/is generally believed that the athletes and the testing labs are playing a cat-and-mouse game. Given the above, is that really true?

For the sports stuff, you can read my book. For how this applies to FaceApp, play my video:

 

 

If you like our videos, support us by subscribing to the channel and sharing.