Stephanie Shipp is a research professor at the Biocomplexity Institute, University of Virginia. She co-founded and led the Social and Decision Analytics Division in 2013, starting at Virginia Tech and moving to the University of Virginia in 2018. Dr. Shipp’s work spans topics related to using all data to advance policy, the science of data science, community analytics, and innovation. She leads and engages in local, state, and federal projects to assess data quality and the ethical use of new and traditional data sources. She is leading the development of the Curated Data Enterprise (CDE) that aligns with the Census Bureau’s modernization and transformation and their Statistical Products First approach.
Donna LaLonde is the Associate Executive Director of the American Statistical Association (ASA) where she works with talented colleagues to advance the vision and mission of the ASA. Prior to joining the ASA in 2015, she was a faculty member at Washburn University where she enjoyed teaching and learning with colleagues and students; she also served in various administrative positions including interim chair of the Education Department and Associate Vice President for Academic Affairs. At the ASA, she supports activities associated with presidential initiatives, accreditation, education, and professional development. She also is a cohost of the Practical Significance podcast which John and Rosemary appeared on last year.
Episode Description
What fundamental values should data scientists and statisticians bring to their work? What principles should guide the work of data scientists and statisticians? What does right and wrong mean in the context of an analysis? That’s the topic of today's stats and stories episode with guests Stephanie Shipp and Donna LeLonde.
+Full Transcript
John Bailer
What fundamental values should data scientists and statisticians bring to their work? What principles should guide the work of data scientists and statisticians? What does right and wrong mean in the context of an analysis? Today's Stats and Stories episode will be a conversation about ethics and data science. I'm John Bailer. Stats and Stories is a production of Miami University's Department of Statistics and media, journalism and film, as well as the American Statistical Association. Rosemary Pennington is away. Our guests today are Dr. Stephanie Shipp and Donna LaLonde. Shipp is a research professor at the Biocomplexity Institute at the University of Virginia and a member of the American Statistical Association’s Committee on Professional Ethics, Symposium on data science and statistics Committee, and the professional issues and visibility Council. LaLonde is the Associate Executive Director of the American Statistical Association, where she supports activities associated with presidential initiatives, accreditation, education, and professional development. She's also a co-host of the practical significance podcast, Stephanie and Donna, thank you so much for being here today.
Stephanie Shipp
Well, thank you for having us. I'm delighted to be here.
Donna LaLonde
Thanks, John. It's always fun to have a conversation on Stats and Stories.
John Bailer
Oh, boy, I love that. I love getting that love from another podcaster. So thank you so much.
Donna LaLonde
Absolutely.
John Bailer
So your recent Chance article had a title ending in an exclamation mark Making Ethical Decisions is Hard! Well, I'd like to start our conversation with a little bit of unpacking of that title by having you describe an example or two, where data scientists encounter decisions that need to be informed by ethics.
Stephanie Shipp
I might start with that, because I'm the one that's always saying making ethical decisions is hard. And Donna seized on that and said, that will be the title of our article for Chance. And I'm like, Okay, that's great. So I don't have examples, but I want to just start by saying that I'm always on the hunt for tools to incorporate ethical thinking into our work. And I find conversations about ethics, especially with my staff primarily, who are young, a lot of postdocs and assistant research professors and students. But these conversations often go flat. So when we try to have conversations about our projects in the context of ethics, their reaction is well, I'm ethical, do you think I'm not ethical, or we only use publicly available data? So what's the big deal? And so we do a lot of the things like the traditional implicit bias training, and that's helpful. But that's actually more individually focused. It does translate to projects, because implicit bias is one of the areas of looking at ethics and projects. But it's not the entire answer. And so the focus of my work throughout my career has always been on: how do we benefit society? And thanks to Donna, if you notice that I'm participating in three AASA activities, I didn't actually realize that until they were listed, and I'm like, that's why I'm always so busy. Okay. I digress. Is that one of the first activities that I got involved in? Because I asked Donna if I would join the Committee on Professional Ethics. And there was a spot at that time because it's a committee of nine members, although they do have a lot of friends. And I was fortunate to join in the year that they were revising, they have to revise them every five years, the HSA guidelines. And I got to watch with awe, as a subgroup, every two weeks met and talked about how they would broaden those guidelines to incorporate data science and statistical practice across disciplines. I then also, at about the same time, was invited to be part of the Academic Data Science Alliance, and they were coming up with their own guidelines. And the group decided we had enough guidelines as good, the American for computing scientists are good. So why don't we create a tool which happened to me, I was like, This is great. And then I also became very involved in the history focused on societal benefit. So that's not really answering the ethical dilemmas I faced in my career, but sort of why I find making ethical decisions hard and what I've set out to try to do to maybe make it easier for not only me but others as well.
John Bailer
So Donna, you want to jump in with some sort of your sense of some cases or places where data scientists encounter decisions that need to be informed by ethics?
Donna LaLonde
Yeah, actually, we probably could have just titled the article Making Decisions is Hard. And I think that one of the reasons that I was so excited to see the ad work is the Academic Data Science Alliance, because I thought their focus on case studies aligned really nicely with the ethical guidelines for professional practice that the Committee on Professional Ethics had been involved in, in revising. And then obviously, the HSA board approved. And I think the reason that making ethical decisions is hard is, or maybe the two top reasons in my way of thinking, one is, is that there's often a power differential. And it's really hard to navigate that power differential just in your day to day work, right? If you're a junior investigator, and there's a more senior investigator, it can be difficult not to say that all of the conversation is too difficult, but it can be difficult to navigate, a concern or a potential place for disagreement about what's the best practice. And so that's, that's a part of where we're, I think the melding of case studies, and the ethical guidelines are really powerful, because it lets you practice before you're actually confronted with having to deal with a potential issue. I think the other issue that I became more aware of, as I was sitting in on the deliberations of the Committee on Professional Ethics, is there are a lot of stakeholders, and all of those stakeholders bring different perspectives and have different context. And so just navigating that landscape that is really complicated, also takes practice. So not specific examples of ethical decision making being hard, but sort of the bigger picture, which I think the ads tool and the ethical guidelines, help support.
John Bailer
You know, one of the things that I find interesting about discussions of professional ethics, ethics and data analysis, is that it's something that has evolved over time, you know, that you have this history. And you mentioned that in your article as well, going back to the late 1940s. So I was wondering if you could give a little bit of a sort of a lead into what was some of the history of research ethics, that then led to kind of this latest layering of considering data science issues?
Stephanie Shipp
I started with the Belmont Commission, which works only because that is the foundation for the IRB. So the Institutional Review Board processes that at least in the social sciences, we have to file an IRB protocol for every project that we undertake. Amazingly, there's a lot of disciplines that don't have to do that, although at UVA, that's somewhat different. But the Belmont Commission started because of the ethical failures of researchers primarily in the United States that were coming to the surface. Perhaps the most famous is the Tuskegee syphilis study that was conducted for a period of over 40 years in which African American men were subjected to a study of watching the progression of syphilis, even after penicillin had been discovered. And they were not told about the treatment, violating every ethical principle by today's standards. Because of that, I sort of wanted to say, Okay, how far back does this go and it actually, it's not that ethical discussions haven't gone back for a long time. But the first written one that I could find was the Nuremberg Code, which was a result of the atrocities of World War Two. And they had 10 ethical principles, and they were really clearly written but tense a lot to remember. And so 30 years later, when the Belmont commission formed around 1979, I think they realized that and they came up with the three principles of respect for people, which means you must be able to volunteer for the study, and you must be able to withdraw from the study. And that goes to the point that Donna made about the power differentials. You know, if there's somebody in authority telling you, you have to be part of that study, you may feel you have no choice, but that's not true. And then beneficence, understanding the risk of benefits of the study, but you have to weigh that with doing no harm and maximizing the benefits over possible harms. And then justice decides on the risks and benefits, so the research is distributed fairly. I think these are really important. But I also think their language is a bit hard to deal with, sometimes grab your, you know, wrap your arms around, and that's why I would advocate that you do need new tools and new ways of thinking. So that's a little bit of the history, but I think Donna's perspective was also really insightful when we looked at that and how we might be expanding our look at what the Mulla report did as well.
John Bailer
So Donna, did the AASA have sort of guidelines for professional ethics?
Donna LaLonde
It was informed by some of these discussions of this Menlo report. Well, actually the most recent revision was approved prior to the work that Stephanie Wendy Martinez and I have been doing and then it's since been joined by an ethicist colleague of Stephanie's. Although Stephanie mentioned she was on the Committee on Professional Ethics at the time that the working group was working on the revisions, and so certainly acknowledged the existence of the Menlo report. And obviously that's the Belmont, the Belmont Report. I think I'm excited about the opportunity and feel it's really critical that the HSA play a role moving forward. Now we're talking about artificial intelligence technologies, and how those technologies are going to impact science, but also society. I read, and I think I'll get this, this is close to correct, if not a direct quote, I read that Tim Berners Lee has said recently, that in 30 years, we'll all have a personal AI assistant. And it's up to us to work to make sure that it's the kind of assistance that we want. And I think that that's a really important conversation, that that needs to be informed by the American Statistical Association, obviously, that at the ad set group is really important as well, the Association for Computing Machinery, it has to be collaborative, because data science and AI is is collaborative, but we have to be focused on it right. And so I'm kind of excited that we might be able to use this Chance article as a jumping off point to figure out how to move that conversation forward and how to build some consensus. I'll just share one other reading. I don't know if you all, because I've just started reading the book, The Worlds That I See by Fei Fei Li, who, I guess now is being called the godmother of artificial intelligence, right. But anyway, in one of the chapters of the book, she says something like, we're moving into a world where from Ai being in vitro, to AI being in vivo, and I thought that is spot on. And we have to be paying attention.
John Bailer
Well, you're listening to Stats and Stories. Our guests today are Stephanie Shipp, and Donna LaLonde. Ethical uses of data have been legislated in parts of the world, including the European General Data Protection Regulation rules, are similar laws starting to emerge in the United States?
Donna LaLonde
Well, I'm not an expert on the laws, I would say similar conversations are happening. And I know that NIST, the National Institute for Standards, is leading the way by having framework conversations. Obviously, the White House issued a memo on artificial intelligence. So I don't, I'm not aware of laws. But I think certainly we're talking about how AI needs to be legislated.
John Bailer
So my question in part was sort of thinking about what are some of these rules of practice, and in your article, you talk about the importance of ethical decisions throughout the entire process, this whole investigative process. And one aspect of that was kind of the data security and data, you know, kind of how you deal with the data, and sort of this is a matter of trust. And that immediately got me thinking about things like this GDPR rules that were really kind of codifying, and forcing this idea. So that was an example of kind of saying, Look, if you're there certain information, informed uses of your data. So this is tying on some of those issues that you mentioned about informed consent, risks, benefits, and otherwise. Can you talk about some of the other components of an analysis where ethical decisions are coming into play? I mean, you know, Stephanie, you kind of hinted at it kind of with where you were talking about this idea of implicit bias, that might be part of an analysis. Maybe you could sort of expand on that a little bit for us.
Stephanie Shipp
Sure. I'll go back to your GDPR question for a second. I mean, that's primarily on the commercial side, and making sure that companies aren't misusing the data in ways unintended that could cause unintended consequences. Claire McKay Bowen has written a book, Protecting Your Privacy in a Data Driven World, and I highly recommend that and maybe highly recommend her. Maybe she's already been on Stats and Stories. Okay, and so she would be the expert to talk about that specific legislation. But definitely in terms of implicit bias, that's probably one of the hardest parts. Because we all think we're ethical. We all think we're very objective. When we're doing our work primarily as statisticians or economists or any, anyone in a quantitative field. I think it's because of constant conversations and training. And I'll just give a really simple example from work that we were doing a few years ago, where we were bringing data in science to inform or promote or support economic mobility in rural areas. And it was a three state project, we were working with colleagues in Virginia, Iowa, and Oregon. And one of the professors was just in this is what I find with ethics, when you see solutions. They're deceivingly simple and elegant. But, you know, thinking of those ahead of time, it's not always so easy. But anyway, this professor, they were just starting out with working with a project in rural areas. So he used a Mentimeter. It's a tool that collects data or answers from a team or a group, anonymously. And then it provides some analysis. In this case, he did a word cloud. So he just asked them a really simple question, what is life in rural America like? And so these students, you know, they quickly started putting in a lot of just words and keywords in their thoughts. But when the word cloud showed up, they immediately recognized their implicit bias. So there were a lot of positives or neutrals that they talked about rural areas being quiet, hardworking, healthy, small towns, crops, or farming, and also had a lot of negatives for the uneducated, ignorant, isolated, forgotten, non optimal. Well, they now went into their project working in a rural area with their eyes wide open, they now understood Oh, now when I'm looking at the research questions, we're going to be asking for problems that they mutually identified with the community. They could now address, am I being biased? When they're looking at the data sources they were using? You know, are these data sources? Will they have unintended consequences? What about my analysis? What are the results? Will they harm a particular group over another group, you know, maybe at the benefit of another group? So I thought that was just a very simple but excellent way to teach implicit bias specifically in the context of a research project. And that got me excited.
John Bailer
So would you think about the kind of workflow in a data analysis project? There's also analysis that occurs, there's modeling, there's prediction, and you mentioned some of their ethical issues, even in how you train a model, how you build a model to make predictions for other cases? Could you talk a little bit about how that might play out in terms of an ethical concern?
Donna LaLonde
Well, I'll just jump in and say, I think we started to appropriately pay more attention to vulnerable populations, right. And so that if the data set isn't reflective of the population, then the model is going to be flawed. And I think, you know, we all are probably familiar with some of the facial recognition, the concerns about facial recognition, right, and where white faces are more likely to be recognized than people of color. So I think it starts with the data that's being collected, then it also is, I think, we talked about models or being black box. Right? Really, do we really understand what the model is doing? Or do we just sort of trust and I think that many in our community are moving us to be more aware that we need to have interpretable machine learning, right, we need to understand what the model is doing. Because otherwise we're, we're likely to make flawed decisions. And I guess, John, I'll just say one thing, I think I left the tee out of NIST. So I want to make sure I give a shout out to the National Institute of Standards and Technology, right.
John Bailer
Nailed that answer to a tee. Perfect. Yeah. So it's interesting, when I was looking at some of your discussion in that pit in the paper, you talked about the idea that some of these rules like the ADSA ethos, talks about different lenses to think about, you know, work that's being done. Could you give a couple of examples of such lenses and why they're important?
Stephanie Shipp
I'm happy to jump in on that one. So I think in their case studies, they gave good examples, and one of the simplest ones and they say it was the simplest way to get the sort of story or get people thinking about this was using cell phone data to conduct a census and they just focused on the life cycle stage of data discovery. And of course, data discovery led them to say using cell phone data. And so what are the kinds of questions you might ask? It would be like, What was the motivation of the company for sharing their data? And are they sharing a complete set of data? What are the challenges with the data? Are they willing to be forthright about that? Or is it again, a black box? And if it's a black box, maybe you can validate those data using other data sources. But really going through that sort of the whole lifecycle and asking those questions, but how important first, the problem identification is to identifying those data sources that are relevant. And then really questioning? How are the data born? What's the motivation for providing them? What's missing in those data? And what kind of biases might be implicit in the data as well? And then again, always the ultimate question, how might this harm one group at the risk of benefiting another? And so the cellphone data in some countries, that may be all they have, they may not have the resources to conduct a census, but then how might you validate that if you are using it, so it's always weighing the pros and cons of the limitations and the caveats, with the benefits?
John Bailer
You know, it's interesting, as you're talking about some of these applications, in certain places, you can't get other kinds of data, they're not even available. And I know that the existing datasets are becoming more and more important to our friends in the official statistics community. Just because you know others, they're a great supplement to these existing data sources they can find. But I'm curious about this idea of provenance of data, just sort of knowing where it comes from. And that's also something that makes me think a lot about the models that are being used, whether they're the generative AI models, or others that are being used for prediction. A lot of times the good examples that you've given, people have provided a lot of detail about where their data comes from, and their analyses and they share the models on GitHub or some other repo, there's sort of, it's almost this, this kind of let the light shine in. And you can see what I've done. So is this a sea change in terms of how people are being asked to think about when they're doing an analysis, and thinking about when I publish my results, I'm also publishing everything that goes into it.
Donna LaLonde
So, I hope so, John. I think that I think and I hope that we, the members of the American Statistical Association, are leading the way for that, you know, which obviously builds on lots of great work around reproducibility and replicability. But I wanted to come back to your data provenance question, and bring in another group of folks that I think we explicitly want to acknowledge, who need to be a part of the ethical decision making education process, and that is students and teachers. And I think that students and teachers, not just at the undergraduate level, not just as graduate students, but K-12. And I think a lot about this, because I don't know if we are doing a sufficient job of describing the data provenance of the secondary data sources that teachers might bring into their classrooms. And I think that's on us, right. So the work that we are doing at the research level, where we're asking researchers to make their code available, make their data available, I think we need to be thinking about how we're describing these data sets that might be part of an educational experience. So that students are practiced in recognizing the provenance and the ethical concerns that could, could arise. And so wanted to make that explicit. And I think that's the kind of nice compliment that the ethical guidelines and the ads ethos project bring to mind for us, right? Because the lenses are really interesting in terms of a socio-technical view. And then the guidelines are really focused on you as the individual statistical practitioner. And I think you take those two together, and we actually have a powerful way in which to both educate and make sure that in practice researchers and data scientists and statisticians and computer scientists are behaving ethically.
John Bailer
You know, one of the things that I'm really glad that you all have done this type of work, so I sort of, you know, I raised my cup of water to you and saluted because I think that it's so important to have these. When I taught data practicum classes, I would use these as an early assignment for the students to start thinking you know, you're using data from someone you have a responsible ability to treat that with respect. And we used to, we also used to bring people in to do the IRB training with these classes, just to get them thinking about it. But I really love this idea of how do we push a conversation of kind of where does data come from? And what is your responsibility to handle this appropriately? Not just thinking that you can mechanically process it. I'm curious now, just sort of as we're sort of sneaking up on a close here. What do you see as kind of some of the future issues or challenges thinking about ethics and practice of data science and statistics?
Stephanie Shipp
I think we've already discussed some of them with AI. And how do we go forward? Donna, Wendy, the other co-author on the paper, and I have been talking about, does there need to be a Menlo commission, version two or point to 2.0. And Donna brought up education at a young age. I remember when my daughters were learning statistics in first and second grade, I was so excited. But now how do you incorporate like, Okay, where did the data come from? And what are the ethical dimensions of this not you need to of course, make those words a little easier to look at. I also think from this article, what I learned the most was the benefit of looking across disciplines. And I have a colleague who likes to say statistics is the quintessential transdisciplinary science. And in this article, we brought together science and technology studies through these four lenses through the ad sub tool. I learned a lot from that. Again, a lot of the language around ethics, though, I think is very hard to grapple with. And I wish there were a way to simplify that language. But once you understand the concepts, that's also important. We also looked at the computer and the IT world through the Menlo report. But it's also just beginning to look through these from a cross disciplinary perspective, which is what statistics does, but encouraging even more of that, because I think how much we learned just in doing this article and looking across disciplines as well. And then finally, just one last port when I gave my very first talk on statistics. And that was now I think, in hindsight how bold I was not being an expert, and still not an expert in this field. Somebody from industry stood up and said, How do we bring this to industry? And she meant it. But I don't think the industry always feels that way about that. But how do we bring these ethical dimensions of using data, which is part of the premise of the GDPR. Behind that are the teeth of that?
John Bailer
Well, I'm afraid that's all the time we have for this episode of Stats and Stories. Stephanie and Donna, thank you so much for joining us today.
Stephanie Shipp
Thank you.
Donna LaLonde Yep, thank you for having us.
John Bailer
Stats and Stories is a partnership between Miami University’s Departments of Statistics, and Media, Journalism and Film, and the American Statistical Association. You can follow us on Twitter, Apple podcasts, or other places you can find podcasts. If you’d like to share your thoughts on the program send your email to statsandstories@miamioh.edu or check us out at statsandstories.net, and be sure to listen for future editions of Stats and Stories, where we discuss the statistics behind the stories and the stories behind the statistics.