Transcript
0:05 All right. Let's get started. So, welcome everyone. we're really pleased to have Emma Pearson with us today. So, Emma actually comes from Stanford. Well, I mean not originally, but she spent her undergrad and grad school at Stanford. And she actually took one of my classes when I first started at at Stanford. So, Emma is has done a lot of great work at in machine learning in particular addressing fairness and my group has done some of work in fairness and we always go and ask Emma because she's kind of the go-to resident expert on the topic.
0:47 So, she's since graduated she's spending a year at MSR New England before starting as an assistant professor in Cornell next year. So, I'm sure she'll have a lot of interesting things and important things to say along with the general theme that we've been trying to go for in these classes which is how AI really matters and affects people's lives. So, please take it away Emma. Thank you. Thank you for this invitation. It's a pleasure to be here to be back at Stanford if only virtually and actually to be back specifically in CS 221 which was actually the first computer science class I ever took at Stanford. So, it brings back fond memories and I'm not just saying that to to suck up to the professors.
1:31 Okay. So, today I'm going to be talking basically giving a two-part talk. In the first part of the talk I'm going to give an overview of some of the recent projects that I've worked on discussing sort of social implications of AI and trying to use it to improve people's lives. And then I'm going to give a little bit of story about how I got here just in case it's useful to you as you're sort of trying to unravel your own professional choices.
1:57 So, at a high level as Percy said I AI in data science and for a practical applications. And the specific applications I focus on are reducing inequality and improving healthcare. Today I'm going to be talking about using AI to study inequality in three areas. First I'm going to tell you a story about policing and how we can use AI to study inequality in policing. Then I'll talk about using AI to study inequality in pain. And then finally I'll talk about using it to study inequality in COVID-19.
2:30 So let's jump right into it. Let's talk about policing. This is joint work with a number of excellent co-authors whose names I will now attempt to rattle off. Camelia, Jan, Sam, Dan, Amy, Vignesh, Cheryl, Phoebe, sorry, Ravi, and Sharad. So it was quite quite a large project and and the efforts of a ton of people. So why is policing something we care about? I think this year that point doesn't really need to be explained, right? It's obvious that policing has a tremendous impact on communities across the United States. And in fact, it's one of the major leading causes of death for young men, particularly for young African American men.
3:09 And today I'm going to be talking to you about police traffic stops. Why do we care about police traffic stops? Well, they're one of the most common ways we interact with police. Tens of millions of Americans are stopped every year. And there's concern that traffic stops may be racially discriminatory. To be clear about what I mean by racial discrimination, and I'll make this more precise in a couple slides, this is when someone is treated more negatively because of their race. So someone is stopped by police because they're black.
3:36 They wouldn't have been stopped by police had they been driving the same way in the same car but they'd been white, for example. Now this is obviously very bad if it's happening, but it's hard to statistically test for. Let's talk about why. So first challenge we confronted when we embarked on this project is that there was no unified data set tracking every stop made by the police. Rather, the way data is stored is that each department stores data in its own little in its own idiosyncratic format.
4:05 And so we set about creating this data set, and we did so in two stages. In the first stage, our journalist collaborators submitted data requests to more than 150 police departments over the course of five years. This was like a colossal amount of work for them. Journalists are amazing collaborators. Now, of course, now that data comes pouring in, and you have this like nightmarish data standardization task, right? Where every every single data set is in a different format. So we put in thousands of hours to clean up the data and put it into a standard format.
4:34 Now, the good news for you is that we've made all this data available for you. So if you're looking for interesting data sets on inequality or on policing, this is a publicly available resource which is easy to download. The full data set tracks some 227 million stops made across 56 city agencies. So that's stuff like the San Francisco Police Department and 33 state agencies. That would be like the California Highway Patrol. And in the main analysis I'll be talking to you about today, we're going to be analyzing 95 million stops. The reason that number is somewhat smaller is, for example, we have to filter for departments that have enough data to even do this analysis at all. If a department doesn't track the race of stopped drivers, it's very hard to analyze racial discrimination.
5:18 So in our analysis, we look at three different questions. We look at whether the police discriminate in whom they stop in the first place. We look at whether they discriminate in whom they search after stopping them. And then we look at how policy changes affect these things. Today I'm only going to be talking to you about the second question, and I'll do so because it's particularly interesting from sort of a data science AI method standpoint, but also because the methods I'll be describing to you are applicable to studying bias in many other human decisions, as I'll describe.
5:47 So, are police searches discriminatory? A little bit of context on police searches. So after the police stop a driver, they're allowed to conduct a search in order to find contraband. Contraband here means things you're not supposed to be carrying like illegal drugs, weapons, etc. The The of a search is to find contraband. They're not supposed to search you just because they're curious or because they're trying to harass you or whatever. So because the purpose of a search is to find contraband, we're going to test whether minorities are searched when they are less likely to have contraband at a lower threshold of evidence. So if police are searching white drivers, for example, when they're only 40% likely to carry contraband, but they're searching black drivers when they're only 20% likely to carry contraband, those different different thresholds, that would be discrimination under our definition of discrimination.
6:34 Importantly, this is only one way the police can discriminate. There are a lot of other problematic things the police can do, as we've seen this year, of course. we're testing for a very specific type of police discrimination. This is not comprehensive. So the first simple test of whether the police are discriminating in whom they search is to look at the search rates. In other words, how likely is someone to be searched after a stop? And the results of this analysis are shown for our data in the graph at right. State patrol stops on the left, city stops on the right. We're plotting the average search rate across locations on the y-axis.
7:09 And you can see that there are like these very big gaps in this plot with black and Hispanic drivers much more likely to be searched after a stop than are white drivers. But this by itself does not prove that the police are being discriminatory, i.e., applying different thresholds on the basis of race. It's possible that some races are more likely to carry contraband, drugs, weapons, whatever. The purpose of a search is to find contraband. So if some groups are more likely to carry it, police may be more likely to search them even in the absence of applying different thresholds on the basis of race.
7:42 So a second simple test that's been proposed to get around this problem is to look not at the rates of searches, but at the outcomes of those searches. This is called an outcome test, and the idea is you look at how likely the search is to find contraband. We call that the hit rate. This was proposed by Becker and other economists. It's decades old. It's a very frequent test in the economics literature. The intuition behind this test is like, look, if searches of white drivers are finding contraband 90% of the time, but searches of black drivers are finding contraband only 10% of the time, it suggests that police are searching white drivers only when they're very likely to carry contraband, but they're searching black drivers on the basis of relatively little evidence indicative of discrimination.
8:25 So if there are differences in the hit rates by race, that's discrimination under the outcome test. And when you do this analysis on our data, you do indeed see that hit rates are lower for black and Hispanic drivers in both state stops and city stops than they are for white drivers, suggesting discrimination against minority groups. But it turns out that there's a flaw in the outcome test as well, and this is called infra-marginality, and I'm going to illustrate it with a simple hypothetical example, totally hypothetical, these numbers are made up.
8:54 Imagine there are two races, black drivers and white drivers, and imagine among each race there are two groups, those who are very likely to carry contraband and those who are quite unlikely. And these groups are easy to tell apart, you know, maybe one of them is wearing blue hats. So among the likely group, 50% of black drivers carry contraband and 75% of white drivers carry contraband. Among the unlikely group, 5% carry contraband regardless of their race.
9:20 And importantly, imagine in this hypothetical example that the police are not being discriminatory. They search everyone who is more than 10% likely to carry contraband. They apply the same threshold irrespective of driver race. What are the hit rates for white and black drivers going to be in this hypothetical example? Well, the police are going to search all the likely drivers and they're going to end up with a hit rate of 50% for black drivers and 75% for white drivers. So from that difference in hit rates, we're going to conclude that there's discrimination in this hypothetical example, but that's a misleading conclusion because by assumption we're applying the same threshold to both groups.
9:59 So why is this happening? Why are we getting this misleading result? Well, it's happening because the statistic we're looking at, the probability of carrying contraband conditional on being above the threshold, is not the same as what we actually care about, which is the rate itself. These are simply different quantities. Threshold itself is hard to infer. It's not directly measurable from the data, the way the hit rate is. So the solution that's been proposed is to use a Bayesian latent variable model to try to infer this threshold. So I'll tell you about that now. Before I do though, are there any pressing questions and also am I talking at an appropriate volume?
10:39 So the threshold test proposes a stylized model of a police stop. And when I say stylized, what I mean is you can never capture all aspects of the real world in math, right? Your hope is that you capture sort of enough relevant aspects of the real world to enable you to measure the quantities of interest. In this case, the thing we want to measure is that threshold at which the search is being conducted. So the goal of this model is to estimate the search thresholds which are consistent with the observed data, namely the search rates and the hit rates.
11:07 And discrimination, just as before, is if lower search thresholds are being applied in searches of minority drivers. So here's how the threshold test models a police stop. We imagine that when the officer stops someone, they estimate the probability P that that person carries contraband. P captures, you know, contextual factors like the age and the gender of the driver, how nervous they're acting, etc. And it's drawn from a risk distribution which is shown graphically at right. So the risk distribution is a probability distribution on the unit interval. So it ranges from 0 to 1.
11:44 For example, if the police, you know, pull over a bus driver, P is probably quite low, right? Because he's like driving kids around. Hopefully he's not also like burying weapons or drugs, and so P would be pretty low. On the other hand, if they pull over a driver and he's like acting woozy and drinking out of a bottle, like that's pretty sketchy, P is probably higher. Now, in order to fit this model at all, you have to make some assumption about what the risk distributions look like.
12:10 You can't fit arbitrary probability distributions because then you would have infinite degrees of freedom. So the parametric assumption that the model makes is that the risk distributions are beta distributions, which is a very standard distribution on the unit interval. Now, if p is greater than some threshold, the officer searches the person. And if they search the person, they find contraband with probability p. So in the case of the bus driver, he'd be below the threshold and so the officer wouldn't search him, wouldn't find contraband.
12:41 In the case of the woozy acting driver, he would be above the threshold, so the officer would search him and would find contraband with a 75% probability. The model allows the thresholds in the risk distributions to vary by race and location, and discrimination as before is if lower thresholds are being applied in searches of minority drivers. Now, in order to fit this model at all, this being a Bayesian model, you have to specify how you go from the unobserved objects to the observed data. So what are the unobserved objects and the observed data here?
13:15 Well, the unobserved objects are the thresholds, which are the main thing we care about, and the risk distributions. So graphically, that's the dotted line and the blue line in the figure to the right. The observed data are the search rates and the hit rates for each race and location. For example, the search rate for black drivers in Alameda County is 30% and the hit rate is 40%. So how do we go from unobserved to observed? Well, I've shown this graphically at right.
13:41 The search rate is the amount of the risk distribution that lies above the threshold. So graphically, it's the amount of gray mass. You can also, you know, express it as 1 minus the CDF of the risk distribution. This is intuitive, right? It's how much of the risk distribution lies above this threshold. The hit rate is what is the expected value of the risk distribution conditional on drawing from the gray mass. So conditional on drawing from the portion of the risk distribution which lies above the threshold, what's your expected value?
14:10 So that's how we go from these unobserved objects to the observed data. That's sort of the likelihood portion of the Bayesian model. To complete the Bayesian model specification, you also need a prior. You need to place priors on your parameters. I'm not going to describe that in detail. But basically, in order to complete the specification, you place priors on the thresholds and the risk distribution parameters. Now, by combining those two things, the likelihood and the prior, you can use standard Bayesian inference to infer the posterior over the parameters.
14:39 And the specific thing we care about is what is our best estimate of what those thresholds are given our observed data. Now, unfortunately, it turns out the story I told you is a little too simple. And it turns out that fitting a model on a data set of our size is much much much too slow. And the reason goes back to the fact that the risk distributions are beta distributions. In order to compute the search rate and the hit rate, you have to compute the CDF and conditional mean of the beta distribution. And it turns out that that is very slow, especially when you have to compute their gradients as what you have to do to use the MCMC that you do.
15:19 The exact mathematical details of why I'm not going to get into, but the TL;DR is that fitting the entire national data set is impossible. And perhaps more importantly, the test can't be used by people who really need it, journalists, police departments, anyone who doesn't have sort of a ton of compute and a ton of grad students. So what we had to do was replace the beta distributions with a new family of probability distributions called discriminant distributions.
15:46 And describing those distributions in details beyond the scope of this talk, although I'm happy to chat with people afterwards if they're specifically interested in probability distributions. But it turns out that this new family of probability distributions makes the test run two orders of magnitude faster. And that makes it feasible to run on a data set of our size. I guess a high-level takeaway here is that like probability distributions are not just something you learn in CS 109 so you can pass CS 109. They're actually quite practically important and it's worth paying attention to them and and thinking about what their drawbacks are.
16:19 For now though, I'm just going to show you the results, which is, you know, now we can actually take this fast threshold test and we can apply it to our national data set. And so here what I'm showing you is the output of this model. It's the average estimated threshold. And again, we're averaging across locations. And you can see that the average threshold is lower for black and Hispanic drivers than it is for white drivers, suggesting that they're being searched on the basis of less evidence.
16:45 So to summarize what I've shown you from this search analysis, I've shown you three results. I've shown you that search rates are higher for minorities, that hit rates are lower, and that thresholds are lower. This is sort of a characteristic pattern for discriminatory searches. You'll see the same pattern, for example, if you look at stop and frisk data in New York City, which is a very, very obviously discriminatory policy. All three tests here are suggesting discrimination against minorities, but the threshold test is doing so in a way which is robust to the statistical flaws of simpler tests like infer marginality.
17:19 I mentioned that the same methods can be applied in other data sets where you have a binary decision and a binary outcome. So I just want to give you some quick examples of this. For example, we can apply it in the medical domain to COVID testing, for example, where the binary decision is, does someone get tested for COVID? And the binary outcome is, do they test positive for COVID? And if you see, for example, that minorities who get tested for COVID are much more likely to test positive, then it's a worrisome sign because it suggests that they're only getting tested at higher thresholds of evidence. They may be being under tested for COVID. And in fact, we we do see some evidence that that is the case.
17:56 So this is sort of a more broadly applicable methodology. Finally, to close on the public policy impact of this work, I mentioned one benefit of using this different probability distribution is your test runs 100 times faster and this makes it easier for journalists to use. And in fact, that was exactly what we saw. The Los Angeles Times was able to take our faster test with some assistance from our team and use it to show that black and Hispanic drivers in Los Angeles were being searched on the basis of less evidence.
18:23 And in response to that, within about a week, the LAPD announced that they were going to cut back on police searches in response to these concerns over racial bias. This is why working with journalists and other real-world actors is nice because they help you translate your sort of research findings into real-world impact. Okay. So before I go on to the second story, are there any questions I should answer? Yeah. So we have one question from a student asking in India, police harass the poor based on how someone is dressed for or two-seated drivers, for example.
18:59 So can this model that you've been describing be applied based on economic status instead of Yeah, that's that's super You know, I've given this talk like 50 times and no one has ever asked that question. That's super interesting. I would be curious to hear more. I There is nothing in principle which precludes applying it on the basis of economic status. Okay. Last only question for now. Yes. Okay. Cool. All right. So let's move to our second story, which is about using AI to study inequality in in pain.
19:30 And this is joint work with David Uray, Sendal, and Seattle. Uray is a is a professor here and he also prefers black and white photos, it would appear. Okay. Oh, he's also my academic advisor, I guess is is a relevant point. Okay. So a general fact about pain is that disadvantaged groups experience more of it. You see this for socioeconomic disadvantage across a variety of types of pain, across multiple continents, across multiple samples. It's quite a robust finding.
19:56 And you see it for racially disadvantaged groups as well. And this is also true in the condition I'll be talking about today, knee osteoarthritis, which is one of the most common causes of disabling pain in older adults. So mechanically, what's happening is that with the wear and tear of time, sort of the padding between your knee bones erodes, the bones grind together, and this causes a lot of pain. And it's like very common. You know, odds are good that like multiple people listening to this talk will develop it.
20:22 So in osteoarthritis, as in other conditions, disadvantaged groups experience worse pain. A natural explanation is like, oh, maybe they just have worse osteoarthritis. But here's the interesting thing. Here's the fact we're going to try and explain. It turns out these groups have worse pain even when we control for how severe the doctor thinks their disease is. So I want to explain to you what I mean by that, but in order for that to make sense, I have to explain how we measure severity and pain.
20:51 So how do we measure severity? Basically, a doctor looks at an X-ray of the knee, grades it on a bunch of factors, and says this is a summary score. So like specifically, you know, they'll look at an X-ray of the knee and say stuff like, oh, you definitely have an osteophyte, a bone spur, and you have these other features like the joint space between your knee bones has reduced. and so I'm going to give it a score called a Kellgren-Lawrence grade that ranges from 0 to 4, and it's sort of a categorical summary measure where higher scores indicate worse pain.
21:21 How do we measure pain? Well, you ask the patient a bunch of questions like, how much pain do you feel when you're bending your knee? And then we take the answers to these questions and we aggregate it into a single score called KOOS pain score. So it's the result of a survey. The data we're going to be using comes from the osteoarthritis initiative. It's publicly available data. All the results I'm going to be presenting are on about 1,300 people.
21:44 And we're going to be comparing pain by three binary groupings. We're going to be comparing black to non-black patients, and almost all the non-black patients in the data set are white. And we're going to be comparing lower and higher income patients, lower and higher education patients. So what do I mean when I say disadvantaged patients have more pain? So here what I'm showing you is a vertical histogram with pain on the Y-axis. So lower scores indicate worse pain.
22:09 And I'm showing you the histograms for black versus non-black patients. And you can see that there's a big visual difference in the histogram where black patients have worse pain. If you want to summarize it in a single measure, you can just take the difference in means for the two groups, and it's about 10.6 points on the KOOS scale, which is about two thirds of a standard deviation. So it's a big gap. And the results for income and education are somewhat smaller, but still substantively large and statistically significant.
22:37 The things I'm showing in parentheses are the confidence intervals. So what happens when we control for severity? Does the pain gap go away? It turns out that it doesn't. So now the graph I'm showing you right has severity on the x-axis, that KL G score I was telling you about before, and pain is on the y-axis as before. And the important point from this graph is that the orange and blue lines are not on top of each other. Even conditional on severity, there's a gap in pain between black and non-black patients.
23:07 And if we want to summarize the size of that gap in a single number, the standard way to do so is with a linear regression. Specifically, we do a regression on pain of pain on race and KL G. And that tells us basically the size of the pain gap when we control for that severity score KL G. And I've shown those numerical results in the second numerical column. You can see that for race, for example, the pain gap shrinks from 10.6 points when we don't control for anything to 9.7 points when we do control for KL G.
23:38 The important point being it really doesn't get all that much smaller, right? 10.6 is almost, you know, as big as 9.7 and only gets 9% smaller. And results for income and education are similar. So the high level takeaway is controlling for severity doesn't do very much to narrow the pain gap. This isn't our unique finding, by the way. Other studies find this as well. The goal of our paper is to explain why. Why is there a pain gap even conditional on severity?
24:05 Specifically, we're going to try and differentiate between two theories. The first theory we call the outside their knees theory, namely that there are non-knee related factors which are causing disadvantaged patients to report higher pain even when their knee disease is no more severe. And this isn't just some like crazy theory we plucked out of thin air. A bunch of prior work points to some factors that might, you know, might cause higher pain in disadvantaged groups. Maybe higher life stress, differences in access to pain medication, differences in how different groups report pain. You know, there're there're a whole bunch of possibilities.
24:38 The commonality here, though, is that whatever the factor is, it isn't anything that can be seen in a knee x-ray. It's something outside the knee. But there's a second possibility, right? And we call this the in their knees theory. Namely, that there are pain-related ailments in the knee x-ray which KL G isn't capturing. And if we could capture these physical features, we would be able to explain more of the pain gap. So under the first theory, there's nothing to be seen in the knee x-ray that would explain this gap. And under the second theory, there is something to be seen that KL G isn't picking up on.
25:16 So why is the second hypothesis plausible? Here are two reasons. The first is that we don't understand pain all that well. This is true generally. It's also true in osteoarthritis specifically. KL G just doesn't explain all that much of the variation in pain. And a possible reason for this is that KL G was developed decades ago in heavily white British populations. And so it's plausible that it's not capturing all the environmental or occupational features that may be relevant to pain in modern and more diverse populations that may live and work very differently.
25:48 So we're going to try and test whether there are overlooked physical features in the knee which would explain the higher pain levels in disadvantaged groups. This isn't just an academically interesting question. It's also a question with concrete clinical implications. And the reason is that whether you get knee surgery depends on whether the source of your pain is in your knee. If you go to the doctor in a lot of pain and she looks at your knee and she says, I'm sorry, I can't see what's wrong with it, she's unlikely to give you knee surgery for an apparently healthy knee.
26:14 She's more likely to prescribe non-specific therapies like opioids or other painkillers. In contrast, if you go to the doctor in a lot of pain and she says, aha, I know exactly what's wrong with you, you very severe radiographic arthritis, you know, you're a four on the Kellgren Lawrence scale, then it's much more likely under clinical guidelines you'll get some kind of surgical intervention. Consequently, if KLGs is missing true sources of pain within the knee in disadvantaged groups, these groups may be under referred for surgery.
26:43 Okay, so we're going to try and test this and and methodologically what we're going to do is we're going to train a convolutional neural network. This is how you know this is sophisticated because we're using deep learning to search for additional signal in the knee x-ray which would explain the higher pain levels in disadvantaged groups. So what does that actually mean? How are you going to search for additional signal in the knee x-ray? Well, the standard approach to searching for signal in a medical image is to train a model to replicate the doctor's clinical judgment, to train it to predict KLG.
27:14 The problem though is that if KLG doesn't capture all the pain relevant features, we don't want to just replicate it. We don't want to set a ceiling of clinical knowledge when by hypothesis that clinical knowledge might be biased or incomplete. So instead what we're going to do is train the model to learn from the patient by predicting KOOS pain score. So to be very clear, the input to the model is an x-ray of the knees and the output is a knee specific pain prediction called alg p for algorithmic severity measure.
27:43 And if when we control for this algorithmic severity measure alg p it narrows the pain gap more than does controlling for this clinical severity measure KLG, it implies that the clinical severity score is overlooking knee features which might explain disadvantaged patients higher pain levels. Before I go to the results, any any questions about sort of the setup? No. Nice. there's one more. Okay. it's in terms of comparing the pain gaps between different factors like income and race, do we have to consider overlap between the groups?
28:20 Yeah, that's a great question. there is overlap between the groups. There's correlation between all three of these binary variables. each of the individual pain gaps remains statistically significant even when you control for all three at once. you could probably do an analysis where you sort of controlled for all three at once and that might be an interesting thing to do here to kind of keep the exposition as as clear as possible we looked at each group separately but yeah it's it's a good point they're they're definitely correlated.
28:50 Great I think that's it for now. Okay so our first result is that the algorithm does in fact find additional signal for pain in the knee x-ray. The algorithmic severity score alg p predicts pain better than the clinician severity score kl g the r squared is higher the difference is statistically significant and you see similar results for other predictive measures. But like those r squared are really not that high right r squared ranges from zero to one if we're at zero point one six that's not all that high and it's not the central question of our analysis anyway which is does controlling for the algorithmic severity score reduce the pain gap?
29:26 And it turns out that the answer to that second and more important question is also yes. So here the first column is just what I showed you before it says when you control for kl g the pain gap doesn't get that much smaller. But the second column is new it says when you control for the algorithm severity score alg p how much smaller does the pain gap get? The final column gives the ratio of the two columns so for race for example you can see that the algorithm explains 43 percent of the pain gap while kl g explains only nine percent the ratio of those two point two numbers is four point seven.
29:58 The overall implication is that yes there is overlooked signal in the knee x-ray which helps explain disadvantaged patients higher pain so this supports the in the knees hypothesis. yes you should never fit a neural net without doing a lot of robustness checks whatever current computer science practice may be and so we do a lot of them I'm not going to talk about them now but happy to talk about them more later if people have specific questions.
30:26 I do though just want to talk about two accessory results the first is that a diverse data set improves performance spe specifically we compare training the model on a non-diverse train set from which we've removed all black patients to a diverse train set from which we've removed the same number of non- black patients. and so the size of the train set remains the same. We've just altered the racial diversity of it. And what we find is that while both models beat KLG, using a diverse train set further boosts performance. You get a better R squared, you get a bigger reduction in the pain gap. You see similar results for income and education as well.
31:01 So to put this within broader context of sort of AI in medicine, there's been a lot of concern that the training data sets may not be sufficiently diverse, you know, and and this is actually more broadly true than AI in medicine. This is true in medicine full stop. and and this sort of testifies to to the importance of collecting diverse diverse data. And then finally, to speak about the clinical implications. As I said, one of the clinical implications of having good severity scores is that it influences the way surgery is allocated. So we decide to test how would using algorithmic pain scores affect the way surgery is allocated.
31:37 Now, to test the way surgery is allocated, we replicate a previous study and we say we're going to assume knee surgery is given to patients with high pain and severe disease. So two, you have to satisfy two criteria. And we try measuring severity in two different ways using KLG, the clinician severity score, and using AlgP, the algorithm severity score. And we find that because AlgP gives disadvantaged patients higher severity scores, it's in turn more likely to recommend them for surgery. For example, among black patients, roughly twice as many knees were eligible for surgery when using the algorithm severity measure as opposed to KLGs.
32:12 So to summarize, we trained a deep learning algorithm to predict pain from knee X-rays. Our algorithm finds overlooked signal in the knee X-ray, which helps explain disadvantaged patients' higher pain. And a clinical implication is that these disadvantaged groups may be under referred for surgery. To put this within broader context of sort of AI and medicine and AI fairness, there's been a lot of previous and very important work on how machine learning methods can potentially increase disparities in medicine and in other high stakes domains. And that's super important.
32:42 But we should also keep the more optimistic flip side in mind that machine learning and AI, you know, they give us predictive superpowers and they shouldn't inherently be a bad thing if we're wise enough to apply them properly. here we show how machine learning methods can also reduce disparities by detecting signal that humans miss. Key to our results here, key to reducing rather than to increasing disparities is first the choice of the prediction test. So we didn't just try and replicate clinical knowledge.
33:07 And second, we train the model on a diverse data set and we show that that contributes to our results. Any questions about this before I go to the third and final story? Yeah, so we have a question from the first first section of your slides. so can the Bayesian threshold test be applied where the observed data is the output of an algorithm? The observed? I mean you would have to give you more details, but I'm intrigued. I mean there's nothing it's designed to assess bias in in decision making. So whether the decision maker is human or algorithmic, there's, you know, you could apply it to both. I would say in the case of an algorithm, you know, it's likely that you know like at least in principle someone knows the threshold, right? So it might be easier to just like figure out the the actual source code or or procedure behind the algorithm rather than attempting to infer it. But there still might be some algorithmic settings where you don't know that threshold. For example, it's some third party company and they won't tell you what they're doing. and then in principle you might want to apply it here. Yeah.
34:13 Mhm. And then on the on the line of kind of determining whether or not something is discriminatory or biased, what metric would you use to test for testing if something like compass is discriminatory? So how do you know if an algorithm's biased? that that's a big question. I you know, I I would say it is highly context dependent. you know, if you observe large disparities in things like you know, in the case of compass, you see these big disparities in like FPR and TPR, false positive rate, true positive rate, that should certainly be a red flag that you want to dig deeper on. But then you want to try and understand like why are these things arising and how can I ameliorate the situation? I don't think I I would not say like in all cases use AUC and that is your golden answer, you know? No, I don't think so.
35:07 Cool. Should I go? Yeah, you can go ahead and go. Okay, cool. So now I'm going to move to our final story on inequality. this is joint work with Serena and with Pang Wei, so I'm a little nervous cuz cuz Pang Wei will actually know if the details are wrong here. and then with so so Serena is a computer science PhD in Urai's lab and then we also worked with Jaline, who's an epidemiologist at Northwestern, and then we also worked with Beth and David who are sociologists and then with Urai who's a computer scientist. So it's very interdisciplinary work cuz we're sort of studying inequality and COVID-19. So intuitively it sort of draws on on people in a bunch of different domains.
35:46 Okay. So as you know, viruses like COVID-19 spread through human contact. That's why I'm giving this talk remotely rather than in person. Which is to say there is an underlying contact network which modulates the spread of the virus. So under a simple epidemiological model, an infected person can infect anyone she comes into contact with with some probability. Those people then infect their contacts and then you get this you know, sort of incredible spread of the disease across the network.
36:18 So because this network is so important to the spread of the disease, current models often attempt to estimate it in some way so they can simulate the spread of the virus. But they often have to use simplistic estimates of the underlying contact networks because intuitively it's very hard to know who everyone comes into contact with unless you're living in some kind of surveillance state. So people do this in various ways. They might assume for example that anyone can infect anyone, so the network is fully connected.
36:44 Or you might some kind of network which captures trends at a very macro level. For example, an airline network which connects city to city but doesn't tell you anything about the network within a city. Or you might use historical data and say I'm just going to assume that what patterns looked like in 2016 are what they look like now. Intuitively though, having really crude estimates of the contact network is not enough for a couple reasons. The first is that we're undergoing an incredibly dramatic change in human mobility, you know, probably in any of our lifetimes, hopefully in any of our future lifetimes also, right? We have these stay-at-home orders, reopening policies, like everything is crazy.
37:21 And the second is that we often want to find ask very fine-grained questions that depend on mobility in a very fine-grained way. For example, we might want to know the impact of fine-grained reopening policies. Like what happens if I open I open restaurants, you know, from 3 to 4 p.m. on Saturdays but not on Wednesdays or something like this. We also might want to understand inequality in infections by race or by socioeconomic status due to mobility patterns. And intuitively, if we want to do that, we need to understand mobility at a fine-grained level. Simply understanding how New York is connected to LA won't be very useful to helping me understand disparities in infection rates within New York, for example, between rich and poor New York neighborhoods.
38:07 So because we have to understand this mobility network in a fine-grained way, our approach is a two-step approach. In the first step, we're going to try and estimate the human contact mobility network. And then we're going to try and build a model to capture transmission on this network. So let's talk about each of these steps in turn. So how do we estimate this network? Well, we're going to use cell phone mobility data from a company called SafeGraph. Specifically, that data is going to tell us how many hourly visits there are from a neighborhood to a place. What do I mean by neighborhood?
38:40 This is like a census block group, which you can think of as a fairly fine-grained census area with a couple hundred to a couple thousand people. A place, which I'll refer to as a POI throughout the talk, is a point of interest, like a restaurant or a cafe or a religious establishment. You can think of them broadly as places people go when they're not at home. So our cell phone mobility data set basically gives us some sense of the number of hourly visits from a neighborhood to a place.
39:08 So mathematically what we're going to try and estimate is a neighborhood is a network that links CBGs, neighborhoods, to POIs, places. So you can think of this in various ways. You could think of it as a list of matrices, a list of networks where each network each network represents sort of traffic at one hour, or you could think of it as like a three-dimensional cube where the dimensions are sort of neighborhoods, places, and and time slices. But that's the object we're going to try to estimate.
39:37 The problem we run into though is that the cell phone data that SafeGraph provides doesn't actually provide us with an exact estimate of that hourly network. The data they give us for the number of visits from CBGs to POIs is only at a weekly or monthly level because of the way they aggregate their data, and they also censor it for privacy reasons. So in terms of the actual data that we have, we have the number of hourly people going to each POI, the number of hourly people leaving each CBG, and then we have a noisy estimate of the networks connecting POIs to CBGs. So you can sort of think of it as the number of people going out, the number of people coming in, and then a noisy estimate of the matrix linking going out to coming in.
40:21 Now it turns out luckily that there is a machine learning algorithm which is designed exactly for this scenario and which you will learn about if you're lucky enough to work with Pang Wei and the other people in Percy's lab. This is very much Pang Wei's work and this was very fundamental to this to this project. and it's called iterative proportional fitting. and basically it's designed for exactly this setting. It says let's imagine that you're trying to estimate some matrix and you know the row sums of that matrix and you know the column sums of that matrix and then you have a noisy estimate of the matrix itself.
40:51 IPF is an algorithm that will give you back a matrix which is consistent with those row sums and column sums, and subject to that constraint is as similar as possible in terms of KL divergence to the initial noisy matrix. And that's exactly the setting that we're operating in here. So we use IPF to estimate the true mobility networks from the noisy SafeGraph data. So that's a little mathy, a little abstract. Let's sort of give you a picture. so here what we're showing you is an example from the Chicago MSA and we're showing you from two slices, two time slices. the first time slice on the left comes from early March, and the second comes from early April after social distancing measures have started to take effect.
41:33 And the gray lines here represent the number of hourly visits from a CBG to a POI. So you can see two things from this this visualization. first is that the density of the gray lines decreases, indicating that total mobility has decreased from March till April. And the second is that most of the lines are vertical, indicating that people mostly hang around their own homes, and that makes sense. Okay, so now we got our network. Honestly, if you didn't understand any of the math, you that that's fine. The the main point is we have a network linking POIs to CBGs at an hourly level.
42:06 Now we have to put a disease transmission model on top of this network, and this relies on a pretty simple epidemiological model, and I'm going to give you a 30-second crash course in epidemiology, and then you'll know about as much as I do about epidemiology. So let's describe the model now. So a very standard model in epidemiology is called an SEIR model, and probably some of you have heard of this if you've been reading the news.
42:28 and the basic idea is that people move through four states in that order, SEIR. You can't go in any other order. You can't go back in loops. So how does this work? You start at the begin before a disease has entered a population, you start in the susceptible state, which is to say you don't have the disease, you've never had the disease, but you're susceptible to it. Now if you come into contact with someone who's infectious, you you move to the exposed state, which is to say, you know have the virus, but you're not infectious yourself yet. So it's sort of in your body but at low levels.
42:58 Now, after some period of time, you move from exposed to infectious, meaning you have it and you can infect other people. And then after some further period of time, you move to the removed state, which is to say, you no longer have the disease, you can't catch the disease, maybe you've recovered, maybe you've died, but at any point, in any case, you can't catch it again. So what we're going to do is we're going to say we're at each hour of our simulation, for each neighborhood, each CPG in our simulation, we're going to model the fraction of people in each of these four states. So we might say in neighborhood five at hour four, 90% of people are in the susceptible state, 7% are in the exposed state, 1% are in the infectious state, and 2% are in the removed state.
43:39 And then we're going to update that hour by hour. So we have to model transitions between these four states. Two of the transitions, the last two transitions are pretty straightforward and boring and don't depend on mobility. We just say at each time step, you have some constant chance of transitioning to the next step. But intuitively, the first transition, that S to E transition is going to depend a lot on mobility because whether or not you get sick depends on whom you come into contact with.
44:12 So how do we model this critical S to E transition? We assume that infections can occur in two ways, at CBGs and at POIs. You can think of CBG infections as like you're just hanging around your house, but unfortunately someone in your house is sick and so now you're sick. You can think of POI infections as you went out to a bar, there was someone in the bar who was sick and now you yourself are sick.
44:35 So we assume that the CBG infection rate is just proportional to the fraction of a CBG which is infected. Intuitively, if more people are in your neighborhood are sick, it's more likely that you yourself will get sick. The POI infection rate is a little bit more complicated. We assume that the probability of getting infected at a POI is proportional to the fraction of the POI which is infected times a POI specific factor which is capturing sort of specific features about the POI, like how big it is, how long people stay there.
45:06 and so intuitively places that are sort of smaller and more crowded are more dangerous and that's what this this part of the simulation is capturing. A nice thing about this model is that it's relatively simple. For each city we're only going to have three free parameters which remain fixed over time in spite of the dramatic changes in human mobility. Those three free parameters are going to scale those two types of infections, infections at CBGs and infections at POIs. And then we're also going to have a parameter which scales sort of the initial conditions in the model, what fraction of people start infected. The rest of the parameters we're just going to take from the prior literature. We're not going to estimate them at all and this is important because it minimizes concerns about overfitting.
45:53 Okay. How are we actually going to choose those three free parameters? We're going to do what's called grid search. We're going to look over all possible parameter combinations of those three free parameters for each city. How are we going to choose which one is best? Well, we're going to take real COVID case data, the number of COVID cases every day from the New York Times, and we're going to keep the parameter combination which gives us the best fit to real cases in terms of RMSE.
46:18 Now, in order to capture uncertainty in the parameters, we're actually not just going to show results from that best fit set of parameters. We're also going to use all parameter settings which yield an RMSE within 20% of that best fit RMSE. And that kind of captures the idea that like look, our parameters are somewhat uncertain here and you know, we we want to capture that uncertainty. Some of you might be thinking like oh, you know, I think Bayesian inference or something might be a more principled way to do this. Totally agree. please figure it out and write to us. I think that would be awesome.
46:48 and in terms of the time period we're going to model oh, but like why why didn't we do that? Because it was it's computationally difficult as it is to fit this model and so we just weren't, you know, that that would have been a further computational difficulty, but I think it's an interesting direction for future work. Bayesian and Frances only taught next week, so we're safe. Oh, nice. Oh, so hopefully they did they understand the at all?
47:09 Whatever. Okay, well, you'll understand it even better next week. That'll be great. Anyway, but yeah, but next week you can figure this out for us. That sounds good. okay, cool. Anyway, we model early March to early May. and the reason we choose that time period is that's what was available when we were doing this analysis. Cool. okay, so to make things a little more concrete, I want to show you just a sort of video of how how this model looks over time. Is this actually going to work? Praying. Okay.
47:38 Okay, so what's going on here? I'll just talk you through the three graphs in turn. The graph at right is showing you mobility over time. This is not from our model, it's from the raw data. The y-axis is the number of visits to POIs, so you can think of it as a measure of overall mobility. and you can see that it drops pretty dramatically about three weeks into the simulation, so three weeks into March, and that's as social distancing takes effect.
48:01 The middle graph is model output. It's showing you the fraction of people the model thinks are in each of the four states, and it's a logarithmic graph. and you can see that the fraction of people in the EIR states, i.e. those who have had the disease, rises over time. and you can also see sort of how mobility is feeding into the model. For example, if you look at that E state, you can see sort of a very high frequency wiggle, just like there is a very high frequency wiggle in the mobility patterns. Those are daily changes in mobility, so over the course of the day. And that's basically telling you, look, people are more likely to get sick when they're going out in the middle of the day than in the middle of the night.
48:38 And finally, that graph at the left is showing you by it sort of spatially, geographically, where does the model think people are most likely to get sick? Redder indicates that a larger fraction of the population is in one of the infected states. and you can see sort of that especially red segment in the middle of the city. And I'll return to that point in a bit. So, okay, anyone can make pretty graphs. Does this actually fit the data? Yes, it turns out it does fit observed case count data reasonably well. Here the orange X's are reported cases and the blue is the model prediction. And that you you can see that it fits the observed data reasonably well even if as in the left plot you only fit the model on data prior to April and then see how well it performs on data from April to May, it continues to fit the data reasonably well.
49:31 This isn't just true in Chicago. That's not a cherry picked example. It fits data pretty well across cities. And it turns out that it also fits the data better than two baselines that we try comparing to. But I think the high-level point here is not that we have some super duper predictive model. The high-level point is like, look, we have this model that fits the data reasonably well, but also enables you to ask very fine-grained questions. So let's talk about some of those fine-grained questions now. What are some of the questions you can ask with this model?
50:04 So this model is of what would have happened if we had done something differently? What if we had started distancing a week later? What if we had distanced only 50% as much as we actually had? It can help you ask stuff like what are the riskiest locations, the riskiest POIs? Are there POIs which are likely to be super spreader locations because they have a ton of people? It can help you answer questions like what's the impact of different reopening strategies? What happens if you reopen POIs only halfway, for example, like only to half of their maximum capacity?
50:34 How do infection rates look like under that scenario? And finally, it can help you understand why socioeconomic and racial disparities arise. And today I'm actually only going to talk about that fourth question. The rest of the answers to the other questions you can find in our paper. And I think there are probably other interesting questions you can ask as well. The basic point though is because the model sort of, you know, it models mobility from neighborhoods to individual places in such a fine-grained way, you can ask a lot of questions that naturally flow from that fine-grained mobility network.
51:08 Okay, so let's talk briefly about disparities. So we know that socially disadvantaged racial and socioeconomic groups were hit harder by COVID-19. Higher case rates, higher death rates, disparities are very dramatic. That's not our work. That's prior work. It's very very clear, very striking. So there are a bunch of reasons for this, right? It's not all mobility. You know, it's stuff like pre-existing conditions, differences in access to care, worse care when they do, you know, get get into the hospital, etc.
51:36 But mobility is probably part of it, too. We know, for example, that if you are of lower socioeconomic status, it's harder for you to work from home, more likely that you're an essential worker, more likely that you have to go out and do these dangerous jobs and you expose yourself to risk of infection. So it's interesting to ask first, does our model learn that disparities, you know, flow in part from mobility? Like can the model naturally predict the emergence of these disparities? And second, if it does, can it expose the mechanisms via which these disparities arise?
52:06 In order to study this, we don't actually have data on individual people. So what we do is we compare neighborhoods. We compare higher and lower income neighborhoods, for example, and we look at how infection rates vary. So a first result is yes, the model does predict the emergence of these disparities based on mobility patterns alone. Here, the left graph is showing you disparities by income and the right graph is showing you disparities by race. On the x-axis, what we're plotting is how much likelier are people to get infected. So for the left graph, if you're in a lower income CBG, and on the right graph, if you're from a less white CBG.
52:44 And you can see basically that all those boxes, all the blue boxes are to the right of one, indicating the people are likelier to get infected under the simulation if they're from lower income or a less white CBG. So the model is predicting these SES and racial disparities, socioeconomic and racial disparities on the basis of mobility alone. And because the disparities by socioeconomic status are particularly dramatic, I'll focus on those for the rest of the talk, but you can see all the results for both in the paper.
53:16 So, why is this happening? Well, we show two two mechanisms via which it arises. The first you probably already guessed. It's that people from lower income and less white CBGs weren't able to reduce their mobility as much. They had to go out more, and this is probably in part because of stuff like differences in occupation. They're more likely to be essential workers. But the second mechanism is a little subtler. And it's this. It's that when they do go out, they go to places which are smaller and more crowded and therefore more dangerous. And this is true even within the same type of POI. So even conditional on I went out, I went to a restaurant, the people coming from lower income CBGs tend to go to restaurants that are smaller and more crowded and more dangerous. And that's the second thing contributing to these infection rate disparities.
54:06 So I want to show an example of this for Philadelphia, which is the place where we see the most striking disparities. Let's see if I can get this to play. Yeah. Okay. So here this graph on the left is showing you Philadelphia and it's showing you results over time. And what you can see over time is that, you know, this big red spot emerges in the middle of Philadelphia. And where is that? Well, it turns out to be the place with the highest population density. So that's the top right plot.
54:35 And it's also the place with the lowest income. So this sort of very high density, low income area has higher predicted infection rates in our model. And that's happening because the POIs that people are going out to are smaller and more crowded and more dangerous. A final implication is that, you know, the model can look at sort of the predicted impact of reopening plans for people in lower income deciles as opposed to the population as a whole. And it basically what we show is that often reopening plans have larger predicted impacts for people in lower income deciles sort of for people in poorer neighborhoods than for the population as a whole. So when you do consider a reopening plan, it's important not just to consider the overall impact, but also the impact on poorer neighborhoods.
55:22 And in fact, California is starting to consider doing stuff like this. Like you have to look at racial disparities in reopening and racial disparities in impact. You can't just look at the impact on the population as a whole. This is also good practice, by the way, when you're evaluating the impact of an algorithm. You shouldn't just look at how it performs on the population as a whole. You need to also look at how it performs on different subgroups.
55:45 So takeaways. This approach sort of showcases the power of fine grained mobility networks. We show that even a simple model leads to accurate fits in 10 different American cities, metropolitan statistical areas. We show that it can scale even to sort of large networks with lots of places and lots of people. We show that you can because you can capture these very micro trends down to neighborhoods and locations by the hour. You can perform these detailed analyses that can potentially inform more equitable analyses to COVID-19.
56:16 I think a general question that I would have for people, and I don't know if we want to talk about this now or at the end or not at all, but like what are other questions you might want to answer with this model? Because I think there are a lot of other things you can potentially ask beyond what we have asked, and I'm curious as to your thoughts. Should we return to that point at the end though? I don't know how we're doing on time.
56:36 Yeah, we have 20 minutes left, so maybe we can take some questions now and then move move on. Sounds great. Cool. So actually if you guys have working mics, do you want to read out your own questions? So my question was if you are able to take into account the percentage of people that wear masks. We are not. That's a great question. You're your reviewer two, three, one. I don't know. Many reviewers had that question. We do not attempt to take into account the fraction of people wearing masks, and I think that's an interesting direction for future work.
57:11 Thank you. Yeah, hi. I suppose this sort of goes along with your question that you put at the end of the slide, which is what other questions you might answer with this model. But I was wondering like could this model of mobility be used to analyze like other mobility issues that don't revolve around like health or epidemiology, such as how different types of zoning codes or access or use of public transportation and different CBGs affect mobility of those neighborhoods?
57:37 yeah, absolutely. I mean SafeGraph data is very broadly relevant to social science and other questions of mobility and we're using it for other projects as well. It's definitely a gold mine for for other other yes and yes. I was wondering do you think it's possible that we can make connections between like physical mobility between CBGs and POIs and whether that somehow correlates to the degree of socio-economic mobilities with CBGs?
58:11 It's happening. Yeah. Yeah, and you might look at yeah, you might look at Susan Athee and Gensekow. Athee and Gensekow would be the names to look up, but like they look at sort of socio-economic segregation. Sorry, they look at racial segregation using SafeGraph data, but then they correlate it with other measures of sort of economic opportunity using work from Raj Chetty, I think in their paper. So that's absolutely something you can do. I mean causal claims are hard, but it's still interesting.
58:39 And I think that's it. Cool. All right, let's go on then. Great. Okay, cool. So so I I was I was asked to speak briefly about sort of how I ended up on on this path and doing this kind of work just in case it was helpful to people, so I I attempted to to write this down. Okay, so you know I liked math and physics and other similarly nerdy stuff ever since I was a little kid. this is a picture of me dressing up as a chessboard for Halloween, so you can tell I was super cool and definitely had a ton of friends. and I took my first AI class in high school, but I was the only girl in the class and I had a lot of less experience than the boys.
59:19 and some of them made fun of my lack of experience and told me I was the worst in the class. So by the time I got to Stanford, I actually decided I was not particularly good at computer science and I came to Stanford as a physics major and I did not even take any CS classes my first year at Stanford. But in my second year at Stanford, I decided I should give computer science another try. And so I actually enrolled in this class, which at that point was not taught by Percy. I don't know who Who were the teachers?
59:48 Andrew? No. Mm-mm. It's the best teacher. Seba I think so and some other I don't know. I don't even remember who taught the class. I do remember the class was awesome. and and and and honestly like I I this is not propaganda. Like it was actually true. I thought I thought that computer science was super cool. and that summer I started doing computer science research in a physics lab. I was developing algorithms to identify certain types of galaxies. but I really something was missing that I thought, you know, AI was amazing, but I didn't want to use it to study galaxies that were, you know, millions of light years away. There were too many problems that were closer to home.
60:24 And this was really driven home for me a few months later when I got a genetic test which told me I carried a mutation that gave me a very high risk of getting cancer. It's called a BRCA mutation. Some of you might have heard of it. and as you can imagine, this was pretty difficult news to receive as a 20-year-old and I spent the next few months being pretty upset about it. And during this period when I was trying to come to terms with the news, I came across this paper out of Daphne Koller's lab. She was an AI professor at Stanford at the time. I'm sure many of you have heard of her. These days she's in industry.
60:55 And in her paper, they take images of cells from cancer patients and they apply a computer vision model to try to predict whether the patients will survive. And these days you would probably do this using deep learning. Back then they were using old school computer vision. I thought it was the most amazing thing. and more importantly, it gave me hope. I thought, fine, I I this cancer problem. I'll work on this problem with AI. I knew that I wasn't going to cure cancer, but I thought that working on it and learning about it would make me feel better, understanding the things that frighten you often does.
61:25 And so I wrote to Daphne and she wrote back on New Year's Day and she offered me a spot doing research in her lab and I took it. There's a lesson I take from this kind of tough period in my life, which is that crying is underrated or as Gandalf puts it in Lord of the Rings, not all tears are an evil. I I found tough times like this one to be very useful in crystallizing what matters to me. From that point on, I became much more intensely focused on what I wanted to do in life.
61:51 So then I graduated Stanford with a bachelor's in physics and a master's in computer science and I decided to take a job at 23andMe, which was a genetics company that offered very cheap testing for the BRCA mutation which I carried. Their test was so cheap that my little sister, who was then only a teenager, could afford to get tested and learn that she did not carry the same mutation that I did. And I thought that was amazing, you know, expanding access to genetic testing in that way. So I accepted a job at 23andMe.
62:19 But about a month after I arrived at 23andMe, the US government sent them a letter ordering them to stop selling their health-related tests because they hadn't gotten basically the proper regulatory approval. So they could no longer sell their BRCA test, which was the whole reason I had gone to the company in the first place. And for the entire year I was there, I basically didn't do any BRCA research at all, which I think is another important lesson. Like even if you start out on a path with the best of intentions, it's very easy to get derailed, at least for me, and it's very hard to predict what projects will pan out.
62:51 Then I went back to school to start my graduate research. I was still very motivated to do cancer research and over the next couple of years I wrote a half dozen papers developing AI papers for computational biology methods. But I began to feel my work was unsatisfying because I was still too far away from real people's lives. Early in my graduate research, my grandpa, who carried the same genetic mutation that I do, died of brain cancer. We We very close. That's us playing chess up there when I was little. And I wrote a fair bit of my master's thesis next to his hospital bed.
63:23 The thesis develops new dimensionality reduction methods, like a fancy PCA or factor analysis basically, if you've heard of those things, for a certain type of biological data, which is important in cancer and many other settings. And work like that, work like what I was doing just felt very far away, decades away from helping people like my grandpa, which is not to say that no one should do it. I think it's super important that you have people doing that kind of fundamental research, even if it doesn't touch real people's lives for a long time.
63:52 But I began to feel that it wasn't for me, that I wanted something that was going to help people in the short term if I was going to be happy for the with the sort of research that I was doing. So based on sort of that understanding, I started working on data sets where each row was not a cell or a gene or something very abstract, but a person. I kept working on healthcare problems and I also started working more on inequality. That took me forward to some of the problems that I've told you about today, studying things like inequality and pain and policing and COVID, which feel very concrete to me.
64:31 Looking back at my research, I see a lot of failures and wrong turns. I went to 23andMe to research BRCA and I failed to do that. I went to grad school to study cancer and I mostly failed to do that. I've spent more than 10,000 hours of my life getting a PhD and I think it's fair to say that many of those hours have not made anyone's lives better. There's a lot of time running down blind alleys and even when you do have a good idea, there's a lot of time polishing it and repolishing it to get it published.
64:57 And even when you do publish it, often very few people read it. And even when people do read it, does it actually change their minds? A few months ago, a man contacted me because he was writing a New York Times piece about our work on policing that I was just telling you about. And I went back and forth with him meticulously, trying to make sure he stated our conclusions accurately. I'm sure he was very sick of me. And when the piece finally came out, I read the New York Times comment section, and it was obvious that none of the commenters were actually reading our research. They were just spouting what they already believed. And that is probably the project I've gotten to work on, which has been most impactful.
65:36 But even though I've spent so much of my life failing to do good, I still think it's important to try, and that's the final topic I want to discuss. I outlined this part of the talk the night Ruth Bader Ginsburg died. I heard the news, and I knew I wasn't going to be able to write any more code that day, so I decided I'd just walk until I felt better. But unfortunately, it got very dark and cold before that happened. So you'll forgive me if this comes across as a little moralistic or mawkish, but it wasn't the best evening.
66:02 But before I tell you why I think you should try to do good rather than just making a lot of money, I want to acknowledge that there are students watching this talk who really do need to make a lot of money when you graduate. You have families to support, you have huge student loans. These are frightening economic times, and if that describes you, I'm not going to lecture you, and you should please feel free to ignore this bit.
66:24 Still, I can almost promise that for some of you listening to this talk, there will come a point not too far not too long from now where you will have a choice between multiple jobs which are both fun and both interesting and both pay you more money than you could possibly need as a young person. That's what the Stanford Computer Science Salary Survey shows for the last six years I have data. And when that moment comes, I'm asking you to choose a job that makes the world better, and not just in some trivial way, and not just for the very richest people.
66:53 I'm not asking you to donate a kidney or storm the beaches at Normandy or risk your lives treating COVID patients. I'm asking you to choose to make a large amount of money as opposed to an obscene amount of money. It's just not that big a sacrifice in a world with such desperate problems where we've gotten so lucky. And I also think you'll find that you'll get more enjoyment out of whatever money you do make if you feel like you earned it doing something meaningful.
67:19 The other reason I think we're compelled to fight for good is that there are a lot of people doing the opposite. I don't want to get political about this, but I think we've all seen just how catastrophic the consequences of that can be. So if we, who are given the most power to push the world in the right direction, take morally neutral or morally harmful careers instead, the world will slide in the wrong direction. I am only here giving this lecture. Many of you are only here listening to it because people like Ruth Bader Ginsburg woke up every day for decades vowing to push the world in the right direction, to expand the circle of people allowed to be in classes like this one. She could have gone into corporate law instead. Apparently she had a taste for Armani suits. She could have bought a lot more of them.
68:02 I think, ultimately, a lot of us take high-paying, socially neutral or socially harmful jobs, not because we really need all that money, but because we've internalized the implicit and insidious claim that if we make a lot of money, we're good engineers, we've made it in life, we're worthy of respect. We need to break that link. We need to redefine what it means to good be a good engineer. A few weeks ago I got an email from a recruiter from a big finance firm, and I responded the way I typically do. I told him I don't work for finance firms, and he asked me why I didn't want to work with the best engineers in the world.
68:35 And I thought, the best engineers in the world think about the social implications of their work. The biggest factor determining your impact will not be whether you understand all the variance of gradient-based optimization, although you like should put some effort into learning those, both because they are very useful and so Punk way like won't kill me. the biggest factor determining your impact will be the problems you choose to work on. That's what makes a great engineer.
69:02 I'll close with a quote from Ta-Nehisi Coates, Between the World and Me, which is a letter he writes to his son who is about your age. He writes, I would have you be a conscious citizen of this terrible and beautiful world. This is what I would wish for my child and for my students and for myself and for you as well. Thanks very much for listening and I'm happy to take any further questions. Yeah, hi. Well, thank you very much. for this was impressive listening to you.
69:29 Well, I was thinking while you were talking about this medical impact and how to study that especially with regards to to the bias and the privileged population. So there are quite a lot of actually biases, cognitive biases that doctors can can show while taking important medical decisions.
70:03 Are we able to study somehow how they how this happens so just to help them avoid and eliminate or reduce this this kind of risk. Yeah, totally. I mean, I think this sort of like behavioral economics approach to like let's understand doctors biases and put them in in terms of sort of these common heuristics that people use, these common biases that people have is a is a broad and promising line of research. I'm not a behavioral economist. For one example of this type of work, I would point you to it's a recent paper. It's called like who gets tested for heart attack and who should be.
70:44 And it it sort of studies, you know, how do these cognitive heuristics that people use play into decisions like this. The authors are you should look for Sento Milan Ethan and Ziad Obermeyer and there may be some other authors as well. But but the broad answer to your question is yes, you know, you can you can absolutely study doctors biases in terms of sort of things we know cognitively about people and they make decisions. Thank you.
71:10 Yeah, we like a question. Oh, sure. Yeah, just about the the difference between racism and I'm not sure if any study actually about this if any difference between you know IQ is what actually caused the difference or there's a difference as a guess between you know either correlation or consequence and is there any study looking deeper on this to understand the difference?
71:46 I mean there are a lot of differences looking in a lot of studies looking at differences by race and ethnicity. This is a fraught topic. It you know some of the studies have not been good studies like so like in particular studies of racial differences in IQ I think is a very fraught topic and then there's stuff which is you know not at all fraught like let's look at racial differences and I don't know incidence of breast cancer deaths from breast cancer. So yes I mean there's there's there's a lot of research in this area of varying quality but a lot of it is super important.
72:20 It is extremely difficult to figure out causality here and most studies that claim they kind of pushing a particular political agenda and should be treated with a lot of skepticism. Okay thank you. Okay as since we are about time perhaps we shall be wrapping up. Yeah sounds good. So if there's no other questions let's if you can unmute and clap that would be great. I'll count to three so we can all give Emma a really big round of applause. That was an amazing talk.
73:03 One two three. Thanks so much for speaking. Thank you. It's a pleasure. Thank you for the great questions.