Skip to main content
Search form
  • 00:09

    KATIE METZLER: It's really great to be here,chairing this panel on ethics and big data.I'm part of a new team at SAGE called "SAGE Ocean," whichwas set up with the mission of providing researcherswith the skills and tools they needto engage in computational social science.And one of the ways that we're doing thatis by providing books, videos, and online coursesto trained social scientists in stats

  • 00:31

    KATIE METZLER [continued]: and coding, as David called for in his keynote yesterday,as well as hosting events like this that bring togethercomputer scientists and social scientiststo debate key issues.And it feels like a really good timeto come together as a community to discuss ethics and big datafollowing a year of pretty high-profile media coverageof how our data has been used in,

  • 00:53

    KATIE METZLER [continued]: frankly, unsettling ways, and an increasing body of researchthat suggests that maybe all this technology isn't actuallymaking us happier.And in a way, it feels like maybe this is evena breakthrough year for computational social science.As David said to me following the Cambridge Analyticascandal, suddenly everyone knows whatcomputational social science is.Unfortunately, they think it broke the world.

  • 01:15

    KATIE METZLER [continued]: So although all researchers need to and do think about ethics,the questions computational social scientists faceseem, in some ways, thornier.Most of the data used in social research of this kindhasn't been generated by researchers specificallyfor research purposes.So a lot of the usual steps undertakenwhen collecting data from human participants,

  • 01:36

    KATIE METZLER [continued]: such as a formal ethics review, gaining informed consent,and the identification of data, either may not have taken placeor may be very challenging to do.In addition to continued debates about what constitutes privateversus public spaces online, questionsarise as to the role of informed consent.When possible future uses of the data

  • 01:57

    KATIE METZLER [continued]: we produce now through our networkdigital lives are harder to predict and control,the data we produce are often usedin ways that weren't anticipated when it was either generatedor collected, and often this databecome only valuable when linked with other data sources, whichcan lead to challenges, both technical and ethicalaround deidentification and anonymization of data.

  • 02:20

    KATIE METZLER [continued]: There's also ethical questions related to data access.And who holds the data we produce?If companies and social scientistsworking in these companies have access--they're the only ones with accessto the really large social and transactional datasets, where does that leave the broader social sciencecommunity, the broader scholarly community indeed?And how do we ensure that the algorithms companies are using

  • 02:42

    KATIE METZLER [continued]: to turn our data into profit are fair, accountable,and transparent?So this panel aims to address some of those issues,as well as others.Each of the panelists is going to talkfor about 5 to 10 minutes, and then we'llfollow with a group discussion, and Q&A.We will have Dr. Michelle Meyer, who's an assistant professorand associate director at Research Ethics in the Center

  • 03:04

    KATIE METZLER [continued]: for Translational Bioethics and Health Care Policyat Geisinger Health System.In addition to pursuing her own research,she chairs the IRB Leadership Committee,directs the Research Ethics Advice and Consulting Service,and plays a leading role in advancing Geisingeras a learning health system, as well as leading the Meyer Lab.Can you please join me in welcoming our panelists?[APPLAUSE]

  • 03:25

    MICHELLE MEYER: It's important to thinknot only about actions, but also about omissions.And so when we think about the ethicsof studying people, or policies, or practices,it's important to think about ethical issues,both in doing some of that work and also in choosingnot to do some of that work.So since my able copanelists have already

  • 03:48

    MICHELLE MEYER [continued]: addressed several important aspects of the former categoryof considerations, I thought I would spend my time todaythinking about the latter.And in particular, I'm going to give youjust a taste in the time I have of work under progress,a research program under progress,the first chunk of which has been submitted.

  • 04:10

    MICHELLE MEYER [continued]: And you can see my colleagues here.It's sort of a collaboration between myselfand others at Geisinger, along with Duncan Wattsand Will Kai at Microsoft Research.And in brief, this project, this first chunk of it,we've looked at--we've done about 14 experiments and replications,

  • 04:31

    MICHELLE MEYER [continued]: all but the first of which were preregistered at OSF.And I'll only be discussing two of those, don't worry.And they are vignette experiments.And I'll talk a little bit more about thatwhen we get to the data.But first, I want to take a little bit of a step back.So I started thinking about what I've called the "AB illusion."Like all good academics, I've coined a little phrase

  • 04:52

    MICHELLE MEYER [continued]: and tried to make it my thing.So I've written a little bit about this previouslyin normative work, which is my doctoral training, ratherthan empirical work.And this particular paper introduced this conceptand used it to analyze the Facebook mood contagionexperiment, which I hope never to speak about again,

  • 05:13

    MICHELLE MEYER [continued]: and the OKCupid experiment, and Ishould say that's not a comment on the experimentor the company.I'm just really tired of talking about that particular example,and I hope maybe some of the rest of you are, too.So what is the AB illusion?What do I mean by that?I mean, it's the tendency to judge an experiment designedto determine the comparative effects of an existing

  • 05:34

    MICHELLE MEYER [continued]: or proposed policy or practice, often called an A/B test,as more morally suspicious than simply universally implementingeither A or B for everybody untested.And why do we call this an illusion,or perhaps better an anomaly of moral judgment?Because if it would be morally permissible to give everyone

  • 05:56

    MICHELLE MEYER [continued]: in a population A and it would be morally permissible to giveeveryone in a population B, and there'sno reason ex ante to prefer A or B than conducting an A/Btest to determine the comparative effects of A and B,should also be morally permissible.In fact, I would say that, if anything,it should be more morally permissible, i.e. laudable,because we learn something from it.

  • 06:17

    MICHELLE MEYER [continued]: And, in some cases, conducting an experiment is actuallymorally obligatory, although I won't bedefending that assertion here.So here's a recent example.Pearson, you may have remembered this--a little bit of a brouhaha.So Pearson, the big educational company, has software for--I think it's called MyLab Programming.It's for college students who are learning how to code,

  • 06:40

    MICHELLE MEYER [continued]: how to do computer programming.And it's part of the choice architectureof this educational software.When students try a problem and theydon't get it right, the software,you have a few different options.You could provide some sort of encouraging statement.So you could say, for instance, look, some students triedthis 26 times.Don't worry if it takes you a few times to get it right.

  • 07:02

    MICHELLE MEYER [continued]: Or you could say something like, no one'sborn a great programmer.Success takes hours and hours of practice.Or you could just be silent and ask themif they want to try again and not provide any encouragement.So the status quo, until recently,had been C. So the status quo for Pearson software,

  • 07:25

    MICHELLE MEYER [continued]: as far as I know, is that there wasno particular encouragement, but they had-- they hypothesizedthat maybe one or both of the A and B interventions,these forms of encouragement, might help studentseither attempt more problems or get more eventually correct,et cetera.And I think that's an intuitively plausiblehypothesis.As far as I know, no one ever complained about the fact

  • 07:48

    MICHELLE MEYER [continued]: that everyone using Pearson softwaregot the software that had no encouragement.No one said, it's an outrage that we aren't--that the software isn't prompting studentsto try again, as far as I know.This is another universal implementationin the world that was never run, so I can't prove this.But I suspect that had Pearson said--

  • 08:08

    MICHELLE MEYER [continued]: had the CEO of Pearson said, this sounds pretty good to me.It sounds like a really good idea; make it so;and told his team to change the software for everybodyshowing either A, or B, or both, Idoubt that many people would have been especially outraged,either.But what they did instead, as you might have guessed,if you aren't already familiar with this,

  • 08:29

    MICHELLE MEYER [continued]: is they ran an A/B test.So it's important to know this was a cluster randomizeddesign.So these were not people designed within the same classor even within the same university.People were randomized at the level of universitiesor colleges to either get one version, one of these threeversions of the software.And I won't belabor the results, except to saythat the researchers were surprised,

  • 08:50

    MICHELLE MEYER [continued]: and they were quite mixed, at best, for the interventions.And so the reaction, though--this is just one example of many I could offer--was a fair amount of moral outragethat Pearson was conducting experiments.And of course there was no particular consent or evennotice.And so there was this sort of outrage.

  • 09:13

    MICHELLE MEYER [continued]: So what are the animating ethical principlesof this project?Well, from Mill, we take the broad ideathat untested policies and practicescan set back the welfare interests of those they affect.And here, this could be consumersof a product or service, users, employees, patients,any sort of end user.

  • 09:36

    MICHELLE MEYER [continued]: If you're more of a Kant person than a Mill person,I've got you covered, too.So the failure to study the effects of one's policiesand practices on others so long as one's own interests, sayprofit, are met, might constitute treating themas mere means to your end, and thereforebe inconsistent with those people's dignitary interests.So I think there are both welfare interests and dignitary

  • 09:57

    MICHELLE MEYER [continued]: interests at stake in the alternative to an A/B test,which is often simply based on intuitionor the powerful person in charge intuitingthat certain policy will probably work, and let'sjust implement it, and then walk away to the next thing.All right, so here's the first study I want to talk about.So catheters, these large tubes--

  • 10:19

    MICHELLE MEYER [continued]: very, very important, lifesaving.However, they are subject to infections.Those infections can lead to morbidity, even mortality,and a lot of health care costs that are really devastating.There are CDC-approved precautions.And these are pretty straightforward.These are like, wash your hands before you insert or remove

  • 10:40

    MICHELLE MEYER [continued]: a catheter; don't put the catheter inlonger than it needs to be--pretty self-explanatory, standard precautions.So in this experiment, which was MTurk,we randomized participants to one of three conditions.In the A condition, we described--it was a vignette-- we described a hospital director whowants to reduce these types of catheter infections

  • 11:02

    MICHELLE MEYER [continued]: in his hospital, in his ICUs.And he thinks of an idea, which is to print the CDC-approved--a checklist of the CDC-approved stepson the back of doctors' ID badgesso that they can refer to them and kind of crossthem off mentally.We randomized-- this is between subjects design,

  • 11:23

    MICHELLE MEYER [continued]: so another group of participants,we gave the same vignette, except,in this case, the hospital directorthought of a slightly different idea, which was to simply printthose precautions on a poster thatwould go on the wall of the procedure rooms.In the third condition, the A/B condition,as you might have caught on at this point,the same hospital director thinks of two ideas,doesn't know which is best, and so

  • 11:44

    MICHELLE MEYER [continued]: decides to randomize patients to one of those two things,either a badge treatment or a poster treatment.And after a year, we tell participantshe will look at the results, see whichis better for his patients, i.e. which prevents the mostcatheter infections, and he will make that policythroughout the hospital.So our main dependent variable was participants' rating

  • 12:07

    MICHELLE MEYER [continued]: of the appropriateness of the agent's decisionon a one-to-five Likert scale.And so in the A condition, this was badge.What I'm showing you now is the percentageof participants who, on that one-to-five Likert scale,said that the agent's decision was either somewhator very inappropriate.So this is percentage of participants saying this

  • 12:27

    MICHELLE MEYER [continued]: is an inappropriate decision.So you've got 15% or so in the badge.We had free response, things that participantsgave their reasons, and we've got a hold code book for that.It's a whole separate thing.And so some of these--plausibly thought, well, if you're touching your thing,and there could be infection field.OK, fair enough.People liked the poster better.

  • 12:48

    MICHELLE MEYER [continued]: So very few people objected to the poster.Now in the A-B condition, was it appropriate to run an A/Btest to figure out these are overwhelmingly unobjectionablepolicies?Was it appropriate or inappropriate to comparethe comparative effects of two otherwise unobjectionablepolicies?Big difference, right?

  • 13:09

    MICHELLE MEYER [continued]: So about half of people thought it was somewhat or veryinappropriate to do that.We replicated this experiment three times,two more times on MTurk, and a third timeon Pollfish, which is a mobile platform.All right.Second of two studies, in this case, in this vignette,we described a walk-in clinic.

  • 13:31

    MICHELLE MEYER [continued]: So in this walk-in clinic, some doctors prescribe drug A,and some prescribe drug B for blood pressure.And that's all we tell them.We literally call them drug A and drug B.And we tell them that they're both FDA approved,which means they're minimally safe and effective.We tell them that they are affordable, both affordable,and that they both have side effects that patients tolerate.

  • 13:54

    MICHELLE MEYER [continued]: And, again, some doctors like drug A, some doctors like drugB. Patients get to see whatever doctoris available when they walk into the clinic.As a result, if you're a patient in this walk-in clinic,the drug that you're prescribed is random.It's random contingent on who happens-- whichdoctor happens to be there.

  • 14:14

    MICHELLE MEYER [continued]: So it's random in all three conditions.And so in this vignette, we describe one Doctor Jones,who either decides to prescribe drug A for all of his patientsor, in the B condition, he decidesto prescribe drug B for all his conditions,or in the A-B condition, again, he randomizes his patients

  • 14:35

    MICHELLE MEYER [continued]: to either of them.And after a year, he'll figure out which patients do best,and he will offer that, not force a change,but offer that drug and that informationto all of his patients.We have very low inappropriateness ratingsfor the drugs.And happily, they're basically equivalent, A and B. Well, howcould you possibly distinguish?We have made them exactly the same.

  • 14:56

    MICHELLE MEYER [continued]: There's no reason to have any preference here.And remember, it's already randomwhich drug you get in those conditions.But when we say, let's scientificallyrandomize it so that we can learn from this, again,not quite as big of an effect as in the checklist study,but a sizable effect nonetheless.So in conclusion, as I said, we did 14 studies, experiments,

  • 15:21

    MICHELLE MEYER [continued]: and replications.So I happened to show yourself in health care.I work in a health care system.It's important to me.But we also looked at things like interventionsto address global poverty, autonomous vehicle design,opting into retirement, employment nudges,all sorts of things.And we observed the A/B illusion in many different domains.

  • 15:41

    MICHELLE MEYER [continued]: So it doesn't depend on the patient-physician relationshipor something special about that or something specialabout health care medicine.And I want to suggest that the AB illusion createstwo possible dangers.So the first-- and this is some of my earlier workon the Facebook mood contagion experiment--is that, to avoid backlash, the kind of backlash

  • 16:04

    MICHELLE MEYER [continued]: certainly that Facebook and OKCupid experienced, but alsothat Pearson experienced, and there are many other animatingcases from health care that I won't bore you with,to avoid that kind of backlash, it's certainly possiblethat organizations might stop being as transparent about someof the experimentation that they're doing.So they'll continue to do it or at least some of it,but they might not be transparent about it.

  • 16:25

    MICHELLE MEYER [continued]: They might not report the results.They might not publish the results.And that is unfortunate for at least two subreasons.So first of all, the rest of us don't get to learn from that.And second of all, if they aren't being transparentabout what they're doing, when thereare actual problems with the design of experimentsor other conduct of research, which of coursethere often are, it makes it a lot more difficult

  • 16:46

    MICHELLE MEYER [continued]: to detect if all of this has sort of beendriven underground.The other possible danger of the A/B illusion--and this one might even be worse--is that some organizations or individuals might simplydecide not to conduct an experiment at all.Because the sober results of our studiessuggest that a decision-maker whodecides to just unilaterally implement something

  • 17:10

    MICHELLE MEYER [continued]: for everybody, even if it's purely based on intuition,it's an idea they just came up with,there's no data to support it, but theyare a hippo-- they are the highest paid person's opinionright, that person might simply implement thingsrather than actually collecting datato see what the effects of their decisionsare on other people, which would certainly be unfortunate.

  • 17:34

    MICHELLE MEYER [continued]: And so the AB illusion, it's possiblethat it may reflect a sort of heuristic ethic, somethinglike, experiments treat people unequally per se,or per se disrespectfully, or that they're inherently risky.And then if we mechanically apply that heuristicto all experiments, ironically, wecan end up with some of those same results, actually.

  • 17:55

    MICHELLE MEYER [continued]: So sometimes implementation of an untested practice is,what I would say, treats people disrespectfully.If you don't care about the effects of your practice,your product, or your service on other people,I mean, companies like Facebook-- yeah,maybe some people quit Facebook, but theycan afford to do basically whatever they want in termsof a news feed algorithm.They don't have to study the effects of it, frankly.

  • 18:18

    MICHELLE MEYER [continued]: And, to me, that is not responsible innovation.So to introduce something new into the world that necessarilywill have some effect on other people,to not really care to study what those effects are,is not respectful of those peoplewho are affected by your services or your products.And the other lesson, I think, is there reallyis no one-size-fits all rule or framework

  • 18:41

    MICHELLE MEYER [continued]: for research or for practice.Sometimes consent is really important.Sometimes it's got to be study-specificconsent. sometimes notice suffices or broad consent.sometimes none of that is ethically required.And both of those things can be true,both for research or for learning activities,and also for practice that isn't tested.And instead of relying on heuristics,

  • 19:03

    MICHELLE MEYER [continued]: like human experimentation alwaysrequires specific informed consent,or it's always risky, or whatever,or unilateral corporate decisions to create policiesare always the domain of-- the right of a CEOto do, we should think through, specifically,each specific decision, either to conduct research

  • 19:23

    MICHELLE MEYER [continued]: or an experiment or not to do so.


As part of the SAGE Ocean team's Ethics in Computational Social Science panel discussion, Dr. Michelle Meyer, PhD, Assistant Professor and Associate Director at Research Ethics in the Center for Translational Bioethics and Health Care Policy at Geisinger Health System, reviews ethical issues resulting from omissions, including several examples of using the A/B test approach (Pearson's MyLab, a catheter study, and a drug prescription study), pitfalls of the A/B test, and lessons learned.

Looks like you do not have access to this content.

Ethics in Computational Social Science– Michelle Meyer

As part of the SAGE Ocean team's Ethics in Computational Social Science panel discussion, Dr. Michelle Meyer, PhD, Assistant Professor and Associate Director at Research Ethics in the Center for Translational Bioethics and Health Care Policy at Geisinger Health System, reviews ethical issues resulting from omissions, including several examples of using the A/B test approach (Pearson's MyLab, a catheter study, and a drug prescription study), pitfalls of the A/B test, and lessons learned.

Copy and paste the following HTML into your website