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

Do wearable healthcare devices work?

A report came out from Stanford School of Medicine about a study of Apple Watch's health monitoring features. Some headline writers are proclaiming that "finally, there is proof that these watches benefit our health!" For example, Apple Watch Stanford Study Shows How It Can Save Lives (link).

When you read the official story, you will learn the following facts about the study:

• The research is funded by Apple
• It was a purely observational study in which they follow (400,000) people who wear Apple Watches
• Participants must own both an Apple Watch and an iPhone to be eligible (plus meeting other criteria)
• There was no "control" group - they did not follow anyone who did not use Apple Watch or use any other health monitoring wearables
• Every participant is self-selected
• The device issued warnings to only 0.5 percent of the participants (~ 2,160)
• Those who received a warning were directed to a video consultation; and the doctor decided whether or not to send the participant an ECG patch, which is used to establish the "ground truth". Only about 30 percent were sent patches, and of those, 70 percent (450) returned the patches for analysis.
• Only those who had ECG data were analyzed. One third of these were shown to have experienced "atrial fibrillation" (irregular heartbeat). This means that two-thirds got false alarms. But if we include the 70% who were not sent patches after the video consultation as false alarms as well, then out of every 100 warnings, only 7 were validated.
• There is no discussion of false negatives: did any of the 99.5 percent who did not receive warnings experienced irregular heartbeats?
• We do know that if there were significant false negatives, then more warnings would have to be sent, which pushes up false alarms.
• Despite the headlines, any lives saved were extrapolated.

There are some major methodological limitations about this study.

Firstly, the study design prevents drawing conclusions of the type "People wearing Apple Watches .... compared to those who did not wear Apple Watches." It does not include anyone not wearing Apple Watches.

Secondly, it's difficult to interpret the accuracy metrics. Is 20 percent false alarms a good or bad number? Is 0.5 percent receiving warnings a reasonable proportion given the demographic and health characteristics of the study population?

Hopefully, this study is just the beginning, and more rigorous studies are being planned.

 

Too much information as bad as no information

I haven't read Kartik's book but it looks like something I'd enjoy.

In this interview with Verge, he cited the following experiment:

One problem with grading in college courses is that different TAs are more or less lenient. So they used an algorithm to normalize or modify the grade so that the level of leniency was consistent. Then, one group received minimal information about how the algorithm worked, a second group got some high-level data, and the third got all of the information on how the algorithm worked and the raw data and all the changes made. The result was that the level of trust in the third group was back down to the same level as the group that didn’t receive any information. So it goes to show that if you reveal that much information, it’s as if you reveal nothing.

This discussion is about transparency when it comes to algorithms. His recommendation is to provide "just enough" information but not "too much". The test group that received details  about the algorithm trusted it less than the group that received high-level information, and about the same as the group that was kept in the dark.

Kartik cites this experiment in the context of explaining to end-users how algorithms work.

The above commentary is equally valid when explaining algorithms to decision-makers, such as senior managers at an organization. Most people experienced at this quickly realize the futility of explaining minute details of an algorithm. Decision-makers are satisified if they gain an intuitive understanding of how the machine functions, and if they observe that the outputs "make sense."

***

One important issue that isn't addressed in the interview is gaming. More transparency is more fodder for those who want to game the system. I discuss this in the chapter about credit scoring in Numbers Rule Your World (link): we can now correct our credit history data, but this will clearly introduce a bias because people will remove incorrect, negative data but not incorrect, positive data.

 

[PS. After I wrote this, the current controversy around Boeing 737 MAX provides a counter-point. The current theory about what caused the crashes (at least the Lion Air one) is that an automation software misdiagnosed the problem. Apparently, Boeing decided not to inundate pilots with too much information and did not disclose the software prior to the Lion Air crash.]

 

Excel error, but could happen in any tool

The most famous Excel spreadsheet error in recent memory is the one that asserts that countries with high debt-to-GDP ratios experience slow growth.

The most recent case of Excel error is the National Highway Traffic Safety Administration's (NHTSA) analysis using data supplied by Tesla to conclude that the "autosteer" feature reduces crash rates by 40%. NHTSA no longer stands behind that analysis, after it was debunked by a consulting firm called Quality Control Systems which spent two years to force the data to be released.

The error can happen not just in Excel but any analytical tool because it relates to how missing data are treated. First, the analyst has to notice the existence of missings. Then, the analyst has to recognize - through some further analysis - that the data with missing values are not like the data without missing values. Finally, the analyst decides to treat the missing values in the appropriate manner - sometimes, they can be dropped; other times, imputed.

The original analysis leads to a simple conclusion. Auto-steering was a feature that was included in Tesla cars since 2014 but the feature was not enabled until October 2015. So, the analytical plan is to compare the same cars before and after the feature was enabled. (There might be a state of enabled but not utilized but it's not in the analysis.) The outcome metric - crash rate - is a measure of airbag activation, based on what one sees in the spreadsheet. Notably the airbag activation has a TRUE/FALSE value so if the airbag activated in the same car more than once, it will only count once.

According to the original NHTSA report, the cars in the dataset had a crash rate of 1.3 per million miles before the auto-steering and 0.8 per million miles after, which led to the claim of 40% reduction.

***

The consultant who obtained the dataset noticed two worrying problems in the data (supplied by Tesla):

• Two-thirds of the cars have empty cells for pre-autosteering miles but non-empty cells for post-autosteering miles. It's implausible that these cars were not driven until auto-steering was activated.
• 47% of the cars have different values of mileage just before auto-steering and just after. These two values should be the same if the moment of activating autosteer was clearly identified for each car. There is a gap in which 134 million miles were driven but not counted in either pre- or post-autosteering miles.

The NHTSA analysis used a zero imputation, meaning that it assumed that two-thirds of the cars had zero pre-autosteering miles. If it had used a mean imputation, that would have added 129 million miles to the pre-autosteering count (average before-autosteering miles for those cars with data was 4426 miles). Since all crashes have been counted, these additional miles drop the crash rate from the reported 1.3 per million miles to 0.4, wiping out the entire 40% reduction and reversing the effect!

Let's turn to the cars that show a gap in just before vs just after autosteering mileage. We can hypothesize that having such a gap or not is uncorrelated with crash rate. This means that if we segment the cars by those with a gap and those without, we should expect the crash propensity to be roughly the same. The data show that cars with gap accounted for 60% of the pre-autosteering crashes but 48% of the post-autosteering crashes. So these cars with unexplained gaps in mileage are pushing up the pre-autosteering rate relative to the post-autosteering rate. This calls for investigating why those gaps exist, and why they are associated with crash propensity.

We still can't decide where to place the 134 million miles driven. That requires understanding the data collection process. This leads us back to realizing that all the numbers in the spreadsheet were presented to NHTSA by Tesla and not audited.

 

 

 

 

Is A/B testing that scary?

Reader AR pointed me to this Fast Company article that examines the ethics of A/B testing.

The only way to comprehend this point of view is to think of A/B testing not as a scientific experiment but as a decision-making process that involves running an experiment. The researchers are unhappy that A/B tests could lend support to decisions that have undesirable impact on society.

Two such examples are described:

• Two images are tested for a job ad. During the test, site visitors were shown one of the two images, selected at random. The winner of the test is an image that disproportionately drives male applicants.
• Separate pricing tests are run in different zip codes. The "winning" prices at the conclusion of these tests are different for different zip codes. Because racial profiles differ by zip code, prices are in effect different for different races. Therefore, the test result leads to race-based discrimination.

There are two important questions to discuss here. First, what is the alternative to A/B testing? Is that method of decision-making better? Second, is the harm produced by the experiment itself, or by the decision made as a result?

 

Alternatives to A/B Testing

Consider the image test described above. Presumably, the test is run because someone believes that one of those two images might perform better at driving applicants. At most companies, a test sees the light of day after teams of people debate and prioritize testing ideas. If the test including a sexist image is run, then the team in charge of testing has approved it for some (possibly bad) reason.

If they didn't have A/B testing, how would they have decided which image to run? And if the image is not explicitly sexist - in other words, if the analyst had to analyze the data to learn that one image drove more male applicants - how would that insight be surfaced without running the other image? The alternative decision process may be even worse.

It is certainly true that automated A/B testing is risky - because no human beings are involved in turning test results into actions. The absence of humans is usually touted as a benefit by vendors of such testing tools. In this example, a human analyst reports on the test result, and includes the analysis by gender showing that while total applications increased, the winning image disproportionately attracted male applicants. The decision-makers can and should decide not to adopt the winning image based on that analysis. The A/B test revealed the bias but did not cause it.

Even without the gender issue, such analysis and discussion of results is necessary. For example, ad clicks can be generated by placing ads near scroll bars to stimulate accidental clicking. Human analysts can report that clicks increased but only through accidental clicking. The decision-makers can and should decide not to implement the winning design.

From where does the harm come?

The other example is more far-fetched. I am reverse-engineering the pricing test as described. Given that the test led to different prices for different zip codes, they would be running separate A/B tests stratified by zip code. Given the law of supply and demand, it might be the case that the winning price would be lower in poorer zip codes and higher in richer zip codes. This definitely results in price discrimination by zip code. If the design team did not want price discrimination by zip code, then such a test design would not have been approved so the test itself isn't creating harm.

Further, race-based price discrimination is accused because zip codes are correlated with race. Almost all variables are correlated with race. Age is correlated with race, so are income, education, what websites one visits, etc. So this standard leads to a banning of all segmentation and targeting policies. The only possible pricing policy would be one price for all.

***

In short, human supervision of A/B testing from design to interpretation is definitely needed. A/B tests provide a wealth of data to support decision-making. The biases highlighted by the Fast Company article are merely revealed by the testing - they are not caused by it.