“We do a lot of process evaluations, Linda, as you know, reports which will ask questions like, did the money get stolen and did the people who got it appreciate getting it? And mostly the answer to those questions is no, it didn’t get stolen and people like getting the money. But that’s quite different from an impact evaluation where you’ve got a control group that didn’t get the program, randomly selected and so that before the program starts, treatment and control are the same.”
“The rise of populism has been substantial across the advanced world, indeed across developing countries as well. So those of us who believe in data need to be strong proponents of the publication of those data even when it produces results that make us uncomfortable.
It’s important to, as the old saying goes, you can have your own opinions but not your own facts. It’s important for the data nerds among us to point very strongly to the value of good data.”
“Trigonometry is intellectually interesting, but I feel like it’s a bit like the Latin of the maths world, something that we teach in order to stretch students’ minds and to filter who can analyze difficult problems, but which is used in relatively few professions.”
Transcript
Linda: Thank you for joining me for another episode of Make Me Data Literate. One of the real joys of this podcast for me is the excuse, it gives me to talk to an incredible range of really interesting people, but I did not expect to have a politician on the show. So it’s going to be really interesting to be talking to a self-confessed stats and data nerd and politician Dr. Andrew Leigh. Welcome, Andrew. Thanks for coming.
Andrew: Thanks Linda. Real treat to be talking to a fellow data nerd today.
Linda: Okay, so can you give us the summary? Who are you and what do you do?
Andrew: So I’m the Assistant Minister for Competition & Charities, Treasury and Employment, probably most relevantly for your podcast. I have responsibility for the Australian Bureau of Statistics, fabulous organisation with a storied record of being the custodian of Australia’s national statistics.
I’m also a former professor of economics and somebody who’s continued to do a bit of data crunching in my spare time. Indeed, I was playing with chatGPT’s data analysis tools over the weekend for looking at data from a survey. So very happy if you want to explore what AI can do in the data crunching world.
Linda: Oh, we could have a long conversation about that. I have big feelings about chat GPT, but let’s perhaps come to that a little bit later. So you’ve got a lot of books as well and a real focus on data and evidence. Did that kind of build up from your formal education or did that come from elsewhere? Did you learn enough in your studies that data became supreme or was it something else?
Andrew: Yeah, I mean, I was tuned into the notion of impact evaluation when I was studying my PhD at Harvard. My thesis advisor was someone called Christopher Jenks, who was passionate about poverty and disadvantage, but inherently sceptical about the ability of any individual program to transform lives. I later learned that’s what’s known as Rossi’s Law, the principle that the better you measure a social program, the more likely you are to find that its impact is zero or close to it. And that’s super important if you care about addressing poverty and disadvantage because you want to be doing it based on programs that have an impact.
And it also helps to explain the persistence of poverty in the face of a significant wave of social programs to address it. So I became an evangelist for randomized policy trials, which are the standard way in which we evaluate medical treatments, but are quite rare in social policy. I’ve seen examples of randomized policy trials showing rigorously that programs work, such as the New South Wales Drug Court randomized trial, or showing rigorously that programs don’t work, such as the randomized trial of “Scared straight”, a US program that put delinquent teens behind bars for a day, which turned out to actually increase their probability of reoffending.
Linda: Wow. I get so excited when people talk about evaluating programs, especially government programs, because they have such potentially wide reaching effects. But whenever I talk to teachers about evaluation, one of the things I always ask is, hands up if your school is very excited about implementing the latest, greatest program, but then doesn’t actually evaluate it before it moves on to implementing the next one. Every hand goes up and everybody kind of sags in their seats. And it’s a big theme of mine that “wouldn’t it be great if schools and businesses and governments all routinely evaluated their programs”, but it’s so, so rare to see someone actually advocate that from the government side.
Andrew: Yeah, that’s right, we do a lot of process evaluations, Linda, as you know, reports which will ask questions like, did the money get stolen and did the people who got it appreciate getting it? And mostly the answer to those questions is no, it didn’t get stolen and people like getting the money. But that’s quite different from an impact evaluation where you’ve got a control group that didn’t get the program, randomly selected and so that before the program starts, treatment and control are the same.
If we use the observational approach to evaluate which COVID vaccine worked, we would have ended up rolling out a range of quite problematic or ineffective vaccines. But we didn’t. We had a control group, a group of people who didn’t get the vaccine or got saline solution injected into them, and a group of people who did. And that really allowed us to do our very best job in tackling COVID.
I’d love to see the same approach rigorously used to evaluating new programs, whether that’s programs to help struggling kids learn to read or whether it’s something like an afterschool program. I think about the example of the 21st century learning communities which were put in place in California under the former Governor Arnold Schwarzenegger. They sounded a great idea, these sort of large scale homework and basketball clubs, but it turned out that they had no positive impact on academic achievement and a negative impact on behaviors, potentially because a group of teens who had been acting up learned from one another how to act up a little more. Just because the program feels good, doesn’t mean it actually works.
Linda: So you said that you learned about the importance of evaluation in your PhD. Did you have anything in your undergrad that kind of pointed that way? I asked because I felt it was sorely missing from mine and I have a science degree. You would think it should be in there.
Andrew: Yeah, I studied economics for one year at the University of Sydney, but then went off to focus on law and political science. I enjoyed those courses, but in retrospect I would have been better to stick with more economics, because by the time I’d finished six years of university, I wasn’t able to think sufficiently rigorously about the problems that I was most passionate about. The legal framework of rights is very useful in the courtroom, but I’m not sure it’s so tractable in a policy sense.
The economic framework of incentives and the ability to sift correlation from causation really is the way I think about problems now. So my 2018 book, Randomistas, made the case for more randomized policy trials, drawing the examples from how the randomized policy trial went from being rare to standard in medicine, but also from some of the innovative randomized policy trials that have been done in developing countries and in various parts of social policy. One of the things I’m most excited by now is that in the last year’s budget, we funded the Australian Centre for Evaluation, set up within Treasury, to do high-quality evaluations, including randomized trials.
Linda: It’s super exciting to see that as part of government. I hope that continues and expands as we go. Is there one thing that you wish everyone knew about data?
Andrew: I think understanding the difference between correlation and causation is probably the most important thing. My view on the school curriculum is we spend too much time teaching trigonometry and too little time teaching probability and statistics. Trigonometry is intellectually interesting, but I feel like it’s a bit like the Latin of the maths world, something that we teach in order to stretch students’ minds and to filter who can analyze difficult problems, but which is used in relatively few professions. By contrast, probability and statistics are used all the time. If you’re thinking about whether to reduce the premium on your car insurance, for example, a little bit of probability goes a long way. If you’re understanding claims that are being made about the impact of something, it really helps to be able to distinguish whether that claim is being made based on a before-after study or on a double-blind randomized trial. The quality of that evidence is very different between those two evaluations.
Linda: Yes. Even simple things like being able to evaluate graphs would be a fine thing to see. We saw in Covid, We actually teach – everybody learns, or rather Everybody meets exponential growth in their schooling, but very few people actually understand it as we saw when we started seeing exponential growth in COVID cases. People were like, “Ah, those cases aren’t scary.” Well, if you put them on a curve, they’re very scary indeed because of what might happen in just 10 days’ time. It was astounding to me how few people had internalized that information and made sense of it. It’s not just what we teach. I think it’s the way we teach it that’s the problem. We need to put it in context more.
Andrew: Absolutely.
Linda: What are some of the worst data mistakes that you’ve seen?
Andrew: I certainly notice challenges in distinguishing signs and magnitudes. One of the things you see often in politics is people are interested in whether something will go up or down more than whether it’s big or small. To take the example of the deadweight cost of taxation, that is, how much economic activity gets destroyed when you raise a tax, it is true that both income taxes and insurance taxes have a deadweight cost. That is, both of them have a negative sign when you look at the impact on economic growth. But what matters there is the magnitude.
Most public finance economists would say that the deadweight cost of insurance taxes is maybe five times as big as the deadweight cost of income taxes. Insurance taxes, I’ve seen on some estimates, cost a dollar of economic activity for every dollar of revenue that you raise.
Linda: Wow!
Andrew: The income tax is maybe 20 cents of economic activity for every dollar you raise. Yet sometimes, the fact that both of them have a cost is what dominates. The need to focus on size is really important. Likewise, being able to reduce very bad things from small probabilities to tiny probabilities can have a huge payoff. For example, if you were to accept the view that nuclear holocaust is likely to wipe out a significant fraction of the world’s population, sorry, has a 1% chance of wiping out a significant fraction of the world population over the next century. Now, 1% is really small, but if you could reduce that to 0.1%, then the expected value of that is massive. A policy intervention that reduces the chance of nuclear holocaust from 1% over the century to 0.1% has an extraordinarily large payoff. You’re talking in expected value terms about millions of lives.
So, thinking about probability can help you calibrate the value of insurance policy type interventions.
Linda: Yeah, it’s that kind of everyday context, the ability to put it into some sort of meaningful framework rather than the way I was taught probability, which was black and white balls in an urn and those classic probability problems that have no interest at all to most people, I think.
Andrew: Yeah, I mean, the urns can be useful. It can be handy to be able to think about probabilities of events occurring together. But just the basics of probability, learning about expected value, which is to take the cost or the benefit of something happening and multiply it by the probability. That’s one idea in economics, which is very powerful.
So, how much should I be willing to pay for something that will reduce the cost by $1,000? Well, if there’s a 10% chance of that happening, then I should be willing to pay $100, which is just 10% of $1,000. That expected value calculation is really handy when you’re thinking about probability challenges.
Linda: I might have to apply that. My car insurance has just come up and it’s awfully high. You start to wonder, is it worth it? And we forget, even I forget as a data nerd, that I could actually apply some basic maths to that and work it through.
Andrew: Yeah, that’s right. So, some economists would say that you should only insure things that would be catastrophic for your household budget. So, don’t insure your mobile phone, but do insure your house. And that changes as your wealth goes up. So, governments, for example, don’t insure their cars. They fix them themselves and just internalize the risk because if you’re a government, then having a car written off is a small thing. Whereas for many households, they would choose to ensure their car because that’s catastrophic. But on the other hand, if you’re a household that’s bringing in $100,000 a year, you might think that ensuring a $10,000 car really isn’t a good use of money because the loss of that car, while really, really annoying, might not be considered catastrophic.
Linda: Have you ever seen data deliberately misused?
Andrew: Absolutely. So, we saw a lot of this during the COVID pandemic. The touting of unproven cures such as hydroxychloroquine for COVID, which was based on early observational studies, essentially just looking at people who took hydroxychloroquine and saying, “Oh, look, they seemed to get better.” But then the randomized evidence came in. It was very clear that when you had a control group, hydroxychloroquine wasn’t an effective treatment for COVID, and yet people continued to spruik it as a cure despite the fact that low-quality evidence had been replaced by high-quality evidence. And you see that a lot of the time in the area of health.
So, for example, we’ve had for a while the suggestion that people who had one drink a day were healthier than people who didn’t drink. And that led some doctors to say, “Oh, a glass of red wine is good for your heart.” But when people looked at the data more closely, they started saying, “Well, hang on, there’s some odd things about this.” Like, for example, the fact that some of the people who are non-drinkers are former alcoholics, and so we’d expect their health to be a little worse.
Now, they also noticed that the one drink a day people tended to engage in better health behaviours, do more exercise, than the people who were complete abstainers. And then a team of researchers said, “Well, here’s something that looks a bit like a random experiment. There’s a small portion of the population that have a genetic mutation that makes them unable to drink alcohol. What if we use that as a kind of random experiment to look at the impact of alcohol on health?” Once you do that, it seems that not drinking is actually a little healthier than even having one drink a day. So while a drink a day isn’t going to kill you, it’s also not extending your life according to the very best evidence.
Linda: Yeah, medicine is an endless fund of stories like that. As a 52-year-old woman, I think, particularly of the studies on menopause and hormonal treatment of menopause and the whole saga of negative impacts of hormone replacement therapy, which turned out to be wildly overstated at best, non-existent at worst.
Andrew: Yeah, and there’s been a number of waves of those studies as well. So I haven’t fully kept up with the literature, but certainly aware one of the recent studies had to be stopped early because the adverse effects appeared so large.
Linda: Yeah, it’s actually one of the things I teach when I teach kids to use real data and teach teachers to teach with it because you have to stop and go, “What is the context and what are the problems with this data?” And that’s something that’s really difficult to get across in a headline or a press release. And so we often see these things reported in ways that even the scientists themselves don’t endorse, but they get through without that important lens of skepticism. What are the issues with the data that was collected? As you say, what are the differences between the populations and that kind of thing? It’s a whole mindset that I’m trying to get everybody to adopt, but it might take me a minute.
Andrew: Yeah, it’s really important also to consider how big some of these effects are. So if you look at, for example, the effect of eating bacon on cancer risk, it may well increase your cancer risk, but it’s probably worth looking at how large that effect is. Some of the studies seem to suggest that it is tiny fractions of a single percent. If you enjoy bacon and you have it every now and then, then you might be willing to say, “Look, I buy that there is an increased risk, but I’ve looked at that increased risk. It seems so tiny that it’s offset by the enjoyment that I have from eating bacon.” And very clearly, the impact of bacon on cancer is orders of magnitude smaller than the impact of cigarettes on cancer.
Linda: Yeah, and it raises the question of absolute risk versus change in risk as well. You can double a risk that’s 0.00001 in a million and you still haven’t made much difference.
Andrew: Yes.
Linda: What do you look for when you’re presented with data? How do you spot the deliberate or accidental misuses? What are some of the tells?
Andrew: Sometimes it’s just a matter of making sure that what’s being presented accords with what you can characterize as a reasonable story. If it doesn’t pass the laugh test, then it’s worth looking into a little bit more. If somebody is telling you that the way to solve Australia’s obesity problem is to eat a particular fruit or vegetable, you have to ask yourself whether the magnitude of that could really be so large.
But so much of the challenge in use of data is plagued by selection effects. So I’ll come back again to nutritional epidemiology. I think we’re now realizing that just about every study which says the effect of this food on that health outcome is X, is plagued by the fact that our food choices are endlessly complicated. Peter Attia and Bill Giffords’ book [The Science and Art of Longevity] is very good on this, essentially making the case that when you do randomised trials, you put volunteers in a dormant tree for a couple of weeks and you randomly change their diet, you see almost entirely different results than when you simply ask people what they ate and how their health is. The former study tells you about causal effects. The latter study tells you about correlations which are often eating choices, you’re learning about eating choices rather than about food impacts.
Linda: You’re also relying on people’s memory and willingness to admit how little broccoli they actually eat and things like that.
Andrew: Yes, exactly.
Linda: And how much chocolate, how many glasses of wine.
Andrew: Precisely.
Linda: It’s so interesting to explore the different results you get when you explore something slightly in a slightly different way or you ask slightly different questions. Just the idea that survey questions can push people in one direction or another, something that people are often astounded by. “How wonderful was my course?” as opposed to “what did you think of the course on a scale of terrible to wonderful?”, it gets you very different results.
Andrew: Yeah, that’s right. And even the context in which surveys are asked. And so if you’ve asked a whole series of questions about a particular topic, you may be warming up someone to think about that topic. So if we ask you a whole set of questions about your feelings on neckties, what colour of neckties you like, whether you prefer bowties or neckties. And then we move on to saying, is a tie an important part of a wardrobe? You may well say, well, yeah, actually, it does feel like an important part of the wardrobe, but only because the previous part of the survey has warmed you up to think too much about ties.
Linda: Oh, I tend to hang out with engineers and computer scientists. So the tie is not a feature of my friends’ wardrobe. The answer would always be no, regardless of how you primed them. But yeah, you make a really good point. What’s the first question you ask when you look at graphs in the media?
Andrew: So one of the things I noticed is actually there’s not enough graphs in the media. I really enjoy the work of somebody like Greg Jericho who’s been doing data journalism for The Guardian. There’s Ed Tadros, the Australian Financial Review has been doing something similar. But there’s not that many data journalists out there who really enjoy putting together a good graph.
So for me, the biggest problem isn’t that there’s too many problematic graphs out there. There’s too few graphs in general. So I’d like to see more presentation of data within articles. And then you’re able to look at the source. One of the challenges is always response rates. We know that in the market research field, response rates are now down below 10%.
As Nick Terrell and I documented in our book, Reconnected a couple of years ago, Australian Bureau of Statistics response rates for the Labor Force survey have been steadily falling. It’s been reported that other statistical agencies are struggling with their response rates. So I suppose that I have in the back of my head how representative is this? Newspaper opinion polls never published their response rates. My guess is that that’s because as many as four out of five people are saying no when asked if they want to do the survey.
Linda: Yeah, that’s an interesting point that I think I haven’t stressed enough in my work is that I always stress the number of respondents, but response rate is a really important measure. How many people are you missing out on? What percentage of the people you asked actually got engaged with the survey and why. I did a survey recently on the after effects of COVID and people’s level of COVID caution, how many times they’ve been infected and what after effects they’ve suffered. And that is, I’m very aware, an incredibly biased sample because although I got over 500, I think nearly 600 responses, that was people who were motivated to answer. So there’s going to be, I think, a massive overrepresentation of COVID cautious people answering that survey.
Andrew: That’s such a good example. Pew recently had an exercise where they were talking about the problems of opt-in polls where you just ask people if they would like to do your survey. They showed that with an opt-in poll, the share of Americans who say that they’re licensed to power a nuclear-armed submarine is something around 5%. In reality, of course, the true answer is less than 0.1%. But the fact that about one in 20 people are willing to lie and claim that they can drive a submarine, pilot a submarine does point to broader challenges in opt-in polls.
Linda: Yeah. Yeah. It’s so tricky to get a really good representative dataset without some very complex and expensive processes, I think. As a combination data nerd and politician, how does it feel when you see governments implementing policy that flies in the face of evidence and data?
Andrew: So I think it’s frustrating. We saw a bit of that during COVID. We saw, for example, Jair Bolsonaro talking about the benefits of hydroxychloroquine. In that case, you really were playing with lives. But ultimately, I’m an enlightenment person. I believe that good evidence crowds out bad evidence. And, that that process isn’t immediate, but that over time, better evidence will deliver better policy. You look at the early childhood space where randomized trials such as the Perry Preschool, Abecedarian, and Early Training Project trials of the early 1960s really reshaped the conversation in the coming decades, moving people’s thoughts away from the idea that early childhood education was really just babysitting towards the notion that it was another form of education which could have significant payoffs later in life. So that’s perhaps one of the best examples of where high quality evidence makes change, but it takes decades for that process to fully take effect.
Linda: I’m interested in your comment that you believe that good evidence crowds out the bad because it feels like the pessimist in me feels like that has shifted a lot in the current age of widespread and viral misinformation. Do you still really believe that?
Andrew: Yeah, look I think populism is testing our institutions, our international engagement, and also those of us who believe in expertise and that there are definitive truths to be identified. The rise of populism has been substantial across the advanced world, indeed across developing countries as well. So those of us who believe in data need to be strong proponents of the publication of those data even when it produces results that make us uncomfortable.
It’s important to, as the old saying goes, you can have your own opinions but not your own facts. It’s important for the data nerds among us to point very strongly to the value of good data. When I first got the job as the Assistant Minister responsible for the Australian Bureau of Statistics, I went to ABS House in Belconnen in the ACT and gave a talk to staff where I talked about the fact that when dictators come to office, often one of the very first things they seek to do is to shut down the national statistical agencies because the one thing a dictator can’t abide is an independent source of truth. Those of us who believe in data need to be fierce defenders of independent, well-funded national statistical agencies.
Linda: It is that discomfort that’s often the issue that even when it makes you uncomfortable, you said that that’s one of the problems with things like COVID, and climate change, and inequality. They’re forcing us to face some ideas that we’re uncomfortable with. We might have to change the way we do things. We might have to fly less, drive less, change our manufacturing. These things that are difficult and make us uncomfortable and I think that’s where facts struggle against the comforting narrative of, “Ah, it’s going to be fine.”
Andrew: Yeah, that’s a good climate change is of course a great example of this and the way in which climate change deniers have sought to undermine experts and institutions such as the Bureau of Meteorology has been really problematic over the years. We do need to make sure that that debate is conducted on a bedrock of hard facts.
Linda: Yeah, that’s the challenge. That’s a good segue to come back to what you were talking about at the very start about chatGPT and data analysis. I’m interested in your perspective on that. I have some strong feelings but I don’t want to sway the conversation. So you go first.
Andrew: Yeah, I’m keen to hear your views, Linda, but mine are certainly that the chatGPT is a better data analyst than I expected it to be. So for example, with what I was doing on the weekend was analysing a survey. I uploaded both the survey results in Excel format and the survey instrument in PDF format. I told chatGPT that it was the same survey and then set about in natural language analysing the data, asking for means of a range of variables, cross tabulations and graphs. It produced them all very efficiently.
Even now and then it would make mistakes. There were eight categories for one variable and when I asked it to tabulate it initially, it gave me six categories. So I wrote back and said there’s eight categories of this. What’s gone wrong? And it immediately apologised and produced the right graph.
It’s interesting sometimes to see how it can miss categories or make those sorts of mistakes. But by and large, it was producing results which were, you know, I had the statistical analysis program, Stata, up and running at the same time. The results from Stata matched the results from chatGPT. But the chatGPT material was quicker to produce and more, it produced in a much more user friendly format where you could very easily just drop it into a presentation.
I did a speech for the Australian Agriculture on Resource Economics Conference recently in which I used these tools to analyse an agricultural forecasting data set just in order to show the audience of research economists how the tools can be useful and what their limitations are.
Linda: Yeah, it worries me. I’m very happy with it in the hands of someone like yourself who’s checking the answers. But it worries me when, as you said, you know, it made some mistakes in places and some of them might be easy to spot, like when there’s eight categories and it only produces six of them.
But what about the mistakes that aren’t easy to spot? If you’re going to, you know, double check every single response you get out of it, then it kind of negates the benefits of using it for, you know, speed and efficiency and ease of use. If you’ve got to go and check it in Stata or in Python, which would be my preferred system, but we can get into that debate another time. I saw that in that ABS talk that you wanted to learn R but not Python and I had feelings about that but we’ll move on and not get into the tech disputes.
But those mistakes, you know, there is an element of the way large language models like chatGP2 work that is not designed to produce truth or accuracy. It’s designed to produce plausibility. And that just feels like such a dangerous place to go in this era of plausible half truth and outright lies. Building in AI to do more of that seems like a really risky path to go down.
Andrew: Yeah, I think it’s a really important point. You do need to check your work at the moment and the tools have some strengths and some weaknesses. So they’re quite good at data cleaning. They’re very good at data matching. They’re terrific at creating graphs, but they’re not always, you know, precise to the last decimal point. And so at the moment you need to do a fair bit of checking.
Now, these models are the worst large language models that we will use in the remainder of our lifetimes. The models are steadily getting better. And I think the programmers are realising that scaling down the tendency of these models to hallucinate/extrapolate is important in terms of people’s confidence in using their data analysis side.
On other things, they’re terrific fun. So if you want an image which illustrates a point in a talk, then you can very easily produce anything in the style of digital art or impressionism or anime. And given that humans tell stories and enjoy pictures, having a quirky picture tailored to the point that you want to make, it can be kind of fun. So for the agricultural forecasting talk, I discovered that poultry forecasts were more accurate than wool forecasts. So I asked it to create an image which would cement that result in reader’s minds, an image of a chicken winning a race against a sheep. It did so beautifully. And that graphic, I think, is the kind of thing you could easily imagine throwing into a generalist presentation where you’re trying to cement an important point in the audience’s mind.
Linda: Are you worried about the ethics of that given these image models were largely trained non-consensually on the work of human artists whose work was stolen and is now being, in a sense, regurgitated or repurposed for AI art?
Andrew: Yeah, I think there’s an emerging discussion about how that compensation should take place. It is certainly true, though, that I was also trained on looking at examples of art by other artists and that in general, I didn’t pay them for the fact that I got to view their images online or in an art gallery or in a book. So we need to make sure that the remuneration schemes are appropriate but also that we’re not building in a set of norms which is different from what already exists in the community. I realised then what I said I suggested that I’m an artist, this is in no way true.
Linda: I was about to follow that as an exciting new dimension to the Andrew Leigh that we knew.
Andrew: Yeah, definitely not. No, no, as my stick fingers can attest, I have not an artistic bone in my body which is probably one of the reasons why I like AI art generators so much.
Linda: What excites you about data?
Andrew: The ability to find fresh things. I remember one of the early case studies that I did when I was studying econometrics at Harvard was to look at the digital divide, the relationship between internet access, computer ownership and income across US households. Seeing those results pop out for the first time at an era where not many people had good empirical results on the digital divide was super exciting and that sense of excitement is still with me when I do empirical research more broadly. It does feel like a way of exploring the world. There’s a sort of spelunking aspect to it. Just as a caver would put on their headlight and go out into a part of the world that hasn’t been explored before, so too. You can do that as a data analyst.
Linda: It’s that thrill of discovery. I feel it too. How do you find the time though between your books and your marathons and your multiple portfolios?
Andrew: A wonderfully tolerant family and an electorate which I think is happy to have a member of parliament who’s talking about big ideas. If we’re envisaging how we ought to regulate artificial intelligence, I actually think there is value in having members of parliament who don’t just talk about it but have actually used these tools and seen how they work and where their limitations are and where their strengths are.
There’s also a sense in which as a professor turned politician, I want to make sure that I’m keeping a foot in both camps, the world of ideas and the world of power, and that I can act as effectively a conduit, a transmission belt from academia through to politics. I can best do that if I stay connected with academia and at least have a sense as to how the tools are developing, what are the techniques that are being used if I’m part of those conversations with colleagues. Part of book writing and article writing and data analysing is just to keep enough of a hand in that I’m able to straddle those two worlds and to serve them both as effectively as I can.
Linda: I guess that’s one of the other exciting parts about data as well. The ability to say when someone makes some assertion, the ability to go and actually check it yourself is…
Andrew: Absolutely.
Linda: I love that. I’m building projects around kids fact-checking politicians and I just love the idea. Thank you so much. This has been such an interesting conversation.
Andrew: Such a pleasure, Linda. Thank you for the terrific questions and all the best in your important work in talking to people about how to use data and how to be more confident data analysts in their everyday lives. It’s such crucial work.
Linda: Thank you. Thank you very much.
