Triangulating the Truth: Strengthening Causal Inference in Epidemiology with Professor Deborah Lawlor

Dr. McGavock: [00:00:00] Welcome back to the Emerging Scholar Podcast. I'm your host, Dr. Jon McGavock from the University of Manitoba, and I'm thrilled to have you back for season two. Last season, we took you on a journey through your first year of research training, covering everything from crafting great research questions to understanding the foundations of both quantitative and qualitative methods.

Dr. McGavock: We met incredible scholars, explored fundamental concepts, and built a roadmap to help you navigate your early research career. This season, we're going deeper. If season one was about laying the foundation, season two is about mastering the craft. We're taking those methods we introduced and really unpacking them, diving into the nuances, the challenges, and exciting new possibilities that each method has to offer.

Dr. McGavock: Throughout this season, you'll meet a remarkable team of both established and emerging scholars with expertise across a diverse array of approaches, from natural experiments that reveal causal insights into real world [00:01:00] settings, to qualitative inquiry grounded in social justice, the complexities of implementation science, and yes, my personal favorite, causal inference.

Dr. McGavock: Each episode will introduce you to researchers who are pushing the boundaries, asking bold questions, and making novel and important scientific discoveries. First, we're celebrating Canadian research excellence, showcasing incredible work happening right here at home. Second, we're emphasizing the critical importance of patient engagement in shaping research that truly matters across a range of scientific methods.

Dr. McGavock: And finally, we're championing curiosity, the essential spark that drives all great science and keeps us asking what if and why. Whether you're designing your first research study, wrestling with methodological challenges, or simply passionate about rigorous and meaningful research, this season is for you.

Dr. McGavock: So let's dive in, stay curious, and continue building the skills that will shape your own scientific journey

Professor Lawlor: Welcome back everyone to the Emerging

Dr. McGavock: Scholar Podcast. I'm your host, Jon McGavock, and today I am thrilled to be joined by someone who I have admired for decades, [00:02:00] Professor Deborah Lawlor. Deb is a professor of epidemiology and a British Heart Foundation chair at the University of Bristol, where she's based in the MRC Integrative Epi Unit.

Dr. McGavock: She trained as a medical doctor before moving into epi, and over the past two decades plus, she has built an extraordinary body of work spanning life course epidemiology, maternal and perinatal health, cardiometabolic disease, and what we're going to talk about today, strong methods for causal inference.

Dr. McGavock: Her early career research examined how socioeconomic position across the life course shapes insulin resistance and type 2 diabetes risk. This work was published in the BMJ and The Lancet, and that set the stage for her career focused on an understanding for how and why exposures truly cause disease, particularly in relation to maternal and perinatal health.

Dr. McGavock: Today's conversation centers on a paper that I think every epidemiologist and every emerging scholar should read and read more than once. The title of the paper is Triangulation in Aetiological Epidemiology. It was published in IJE in twenty sixteen and co-authored with Kate Tilling and George Davey Smith.

Dr. McGavock: In it, Deb and her colleagues lay out the framework for strengthening causal inference by integrating [00:03:00] results from multiple study designs that each carry different and ideally unrelated sources of bias. The core idea here is elegant. If approaches with very different biases all point in the same direction, our confidence in a causal interpretation increases substantially.

Dr. McGavock: If it doesn't, our causal interpretation is weakened. We call this Sherlock Holmesing, and I hope you enjoy it. But we dig into what triangulation actually means in practice. We walk through some different brilliant real-world examples, from breastfeeding and child health outcomes using cross-world comparisons, to the effects of maternal BMI on pregnancy outcomes using Mendelian randomization, alongside paternal negative controls to study the long-running blood pressure and coronary heart disease literature.

Dr. McGavock: We also get into the work that has changed long-standing assumptions about life course epidemiology and the developmental origins of health and disease, including touching on something that I think is ultra important for anyone working in this space, a paper that she co-authored called It's the Mother, It's the Mother, and what happens [00:04:00] when we interrogate the assumption of maternal causes for offspring outcomes.

Dr. McGavock: Along the way, Deb shares how mentors like George Davey Smith and the late Shah Ebrahim shaped her thinking and how she came to the field of Mendelian randomization in its early days. How her research on assisted reproductive technology and fertility has opened up yet another window into causal questions into perinatal epidemiology and the life course determinants of offspring obesity and health trajectories.

Dr. McGavock: This is a rich,

Professor Lawlor: wide-ranging conversation on a very important but complicated topic. Let's dive into it.

Dr. McGavock: Welcome everyone to the Emerging Scholar Podcast. It is my great pleasure to have Dr. Deborah Lawlor here with us today from the University of Bristol. Deborah, welcome to the Emerging Scholar. Can you tell us a little bit about yourself?

Professor Lawlor: Thank you. Thank you very much. So I'm an epidemiologist, as you said, at the University of Bristol.

Professor Lawlor: My background was initially clinical, and then to be honest, I didn't [00:05:00] imagine being an academic. I just got to a stage where I felt I wanted to test my brain, and I wanted to do a PhD, and that was it. But as a clinician, the way to do that is a kind of clinical fellowship or a clinical lecturer post. And I knew I wanted to do an epidemiology.

Professor Lawlor: That was the other thing. Um-

Dr. McGavock: Why? Why epi?

Professor Lawlor: It, it was a process of elimination. I wanted to be working on people rather than laboratory analysis of

Dr. McGavock: whole people.

Professor Lawlor: What was common for clinicians was randomized control trials, and they are absolutely essential and important. But what I wanted, something that I'd done maths at school and was interested in maths, and randomized control trials are so

Professor Lawlor: important, but intellectually, this is what- ... potentially offend a lot of people. You can say

Dr. McGavock: it.

Professor Lawlor: You

Dr. McGavock: can say

Professor Lawlor: it. They didn't feel that stimulated. Right. But epidemiology is like, uh, uh, be an investigator. Y- [00:06:00] you have to pull the evidence together- Yeah ... make some sense about it. So it was that kind of challenge that I was interested in.

Dr. McGavock: The Sherlock Holmes approach, which I often say, like when I talk about triangulation, I'm like, "You're Sherlock Holmes. This is applying Sherlock Holmes thinking." What was your first project when you entered the PhD program?

Professor Lawlor: So the name of my PhD was Life Course Determinants of Cardiovascular Disease: The Role of Insulin Resistance.

Professor Lawlor: Oh, wow. And that was a county thing. It was a sort of advertised clinical lecturer post, which was the way you could do a PhD, and it was with George Davey Smith and Shah Ebrahim as my supervisors, and Shah was setting up this women's heart and health study. And so the idea of the PhD would be that I would do some of the field work and organize some of the field work, [00:07:00] which was fantastic training.

Professor Lawlor: It worries me nowadays that epidemiologists don't do their own field work, and I understand why- Yeah ... because of all the data that's there and because of the pr- professionalized field workers, but I learned so much from doing that. So to begin with, I didn't have the data. I was going out with trained field workers once a month, and we would go to different towns around the UK.

Professor Lawlor: We'd pack the van on a Sunday, spend a week there in church halls and GP surgeries, and weighing, measuring, realizing the importance of calibrating weighting scales, recognizing-

Dr. McGavock: Yeah

Professor Lawlor: Realized at one stage it was important to look at between field worker differences, and it was really notable. Two of the field workers were very different.

Professor Lawlor: One who had sort [00:08:00] of on average higher blood pressure, and there's no reason why this was the case. Um,

Dr. McGavock: but- So they were measuring higher blood pressure? Their own blood pressure wasn't

Professor Lawlor: higher ... well, I mean, I mean basically the thing was I could have worked out without having the numbers who it was. So one of them was fantastic at getting the, uh, forced expiratory, um, measures because you, you could hear her down the corridor going blah, blah, blah, blah, blah, blah, blah.

Professor Lawlor: But I think that same kind of- Yeah ... attitude was what shot the blood pressure up as well. And it's things like that, if you don't go out and do it, you have no idea

Dr. McGavock: that's what's happening. Yeah. Oh my God, we did the exact same thing for, it's now called The CHILD Study, but it was a precursor to that here.

Dr. McGavock: One of the early randomized controlled trials on the hygiene hypothesis was done, and we were out in the field. Like same thing. The whole day you're hearing people screaming at these kids, blah, blah, blah, blah, blah, blah. Oh, that's so fantastic. How did you get hooked up with George Davey Smith? That must have been an amazing experience, right?

Dr. McGavock: To your first foray is in this great group.

Professor Lawlor: Yeah, that [00:09:00] was w- what was advertised. In fact, I had applied for a similar post in London at St George's because I actually wanted to go and live in London, and- I was offered it, and then the professor who was gonna go there, who I would work with, decided not to take the contract, and so it was put on hold.

Professor Lawlor: And I was just like, you know, I've made my mind up, I'm ready for the

Dr. McGavock: move. Yeah.

Professor Lawlor: And that was, in a sense, how I ended up applying for the Bristol post. And yeah, it was an amazing, still, both Shah and George, I learnt so much. I think more than anything, they gave me a confidence in myself.

Dr. McGavock: Awesome.

Professor Lawlor: Which I kind of recognize as an important thing to do as a supervisor.

Professor Lawlor: You know, day one, the background to your PhD is gonna be this book chapter, which Diana Kuh is writing this book on life course epidemiology for women's health, so mine was always gonna be around women's health. So there you go.

Dr. McGavock: That's great [00:10:00] mentoring. And this would've been the height of, like, the Barker hypothesis, like the life course- Yeah

Dr. McGavock: the big boom in sort of DOHaD research or developmental origins of health and disease, AKA DOHaD research. Yeah. So you're right at, right

Professor Lawlor: at the fore- Very much, very much so. And you know, again, that sort of fitted with my interests. But in particular, that sort of notion of a life course approach as opposed to assuming that intrauterine exposures determine everything.

Dr. McGavock: Right. Looking through your CV and, you know, someone who's reading your work a lot, and your work informs a lot of the work we do, I also find that no matter what question you're tackling, you're at the forefront of emerging epi methods. You know, early work with ALSPAC and looking at blood pressure risk, like just always thinking about what are better ways for me to answer this question using the data that we have.

Dr. McGavock: And I think that's so important for trainees to understand and to think about. And for you, what is that push? Like, at every level I feel like you're [00:11:00] always looking forward to say, "Can we do this better?"

Professor Lawlor: It's what interested me in a sense around this detective type of work and, and it's partly that I was in a department with people like Shah Ebrahim, George Davey Smith, Stephen Frankel, Jenny Donovan, and others who were pushing at the boundaries of qualitative research in amazing ways.

Professor Lawlor: So there was always that, how can we do this better? With that emphasis on better, not perfect. Mm-hmm. That it's, perfect study probably doesn't exist. So it's about part of how do we do this better is being upfront about where the limitations are, where the biases are, what are the trade-offs.

Dr. McGavock: And so maybe some of the fundamental questions when you're asking, either when you're reading a paper, like for the first-- as for someone who coming in, you're reading more papers than you're producing knowledge, right?

Dr. McGavock: Yeah. So where is the bias [00:12:00] coming from? What was the framework? What else could contribute to this? I think that's one. I guess when you're designing a study nowadays, what are some of the things you're thinking about when you're designing a study to answer a research question and thinking about these things?

Professor Lawlor: First of all, getting the clear research question, aim, or objective. Doesn't have to be question. Hypothesis, aim, objective. I feel like at times I'm becoming a grumpy old woman. But this kind of thing that I see all the time now in epidemiology papers, including my own group will write the same thing, and I feel like I'm banging my head against a big wall.

Professor Lawlor: It is the end of the introduction, instead of our aim was, our hypothesis is, or... So it's here we use Mendelian randomization or whatever. What I want

Dr. McGavock: is...

Professor Lawlor: And sometimes it has to be a broad aim. That's also fine, but just being clear, [00:13:00] this is what I set out to do. And the reason I think that has to be clear is that for me, you know, when you said how you lucid in there, when you sort of identify the gaps and why it's important, and that informs your aim.

Professor Lawlor: You, you have to have that sort of rationale for the aim. And then because there are all of these data available and data sharing, so epidemiologists do less field

Dr. McGavock: work,

Professor Lawlor: which I think has disadvantages, but the advantages of having all of these data available is it's an efficient approach to research.

Professor Lawlor: And in that open science framework, you've also got the opportunity in certain circumstances to replicate things, to validate things, to, to do triangulation.

Dr. McGavock: Right.

Professor Lawlor: So, you know, my start would be if I think of [00:14:00] things I'm interested in currently, it's o- one example would be around assisted reproductive technologies.

Professor Lawlor: And with colleagues, I did some work starting in the early 2000s around issues. There was a big question at the time, should we have- is mandated single embryo transfers or should we allow double embryo transfers. At the time in the UK it was possible to have up to three transfers. We had the opportunity to talk with couples who were going through this, and it, it was very clear that it was important.

Professor Lawlor: There'd been two big randomized controlled trials. You need massive randomized controlled trials for this. Set up one in the UK, one in Australia, and they both didn't recruit. They had to be stopped. They were meant to be recruiting thousands. And it was interesting because in the UK couples didn't want to be randomized [00:15:00] because they wanted to be allowed to have two or three transfers.

Professor Lawlor: In Australia, it was the other way around. And-

Dr. McGavock: Hmm ...

Professor Lawlor: that was coming from a strong push in Australia that it was unhealthily potentially dangerous to have more than two. So there wasn't the randomized control evidence, and we did a very classical observational analysis and then illustrated that sort of emphasize you'd need to be able to do the stratification by age to say, you know, well, in the older women, should they be allowed more?

Professor Lawlor: And our research suggested they should. But following on from that then, the sort of things that are coming from couples, coming from clinicians are

Dr. McGavock: increasingly about, well, what happens to the child

Professor Lawlor: if they've been conceived by ART? And so, you know, that was, you know, next set of work and as you say, you apply for the grant and everything.

Professor Lawlor: And one [00:16:00] of the things that was fascinating in terms of what might happen to the children was when I reviewed the literature, you had a shift. Uh, so most of it was opinion pieces or commentaries with very little data, but it shifted from these babies are all gonna be skinny and weedy and weak, and they're gonna grow up and be skinny and weedy and weak, and they won't be able to do the things that needs greater physical strength, to as the obesity epidemic emerged, and that was literally it These babies are all gonna be fat

Dr. McGavock: Yeah.

Professor Lawlor: And unhealthy.

Dr. McGavock: Yeah. At

Professor Lawlor: that point you think, "Okay, we really do need some good evidence." So that's just an example of, I suppose, how I ended up doing work. One of the projects I'm doing

Professor Lawlor: at the moment, which is around the long-term offspring outcomes in relation to assisted reproductive technology.

Dr. McGavock: Yeah, and that's awesome. I think the best questions [00:17:00] come from those observations or from those contradictions or someone who's aware of the literature and they're like, "You guys just flipped your viewpoint in the last two years without...

Dr. McGavock: And I don't know of any evidence that forced you to flip your, your viewpoint on this." And in life course epidemiology too, very hard to do randomized control trials throughout the life course, so you're gonna rely on either prospectively collected data or administrative data. I'm interested too, I know that you're a pioneer in sort of the life course epi and the application of Mendelian randomization, and I still feel like people are not doing it right.

Dr. McGavock: There's... It exploded, people became really interested in it, and now we're seeing people doing MR for particulate matter, MR for noodle intake, MR for things that should not be used i- in the context of Mendelian randomization. But you made this leap into Mendelian randomization within observational data early.

Dr. McGavock: You're an early adopter. What was the push to say, "Let's become expertise in this, [00:18:00] and let's think about the methodology there"? Was, what was the rationale?

Professor Lawlor: So I think again, being in, you know... So George and Shah's first paper describing this was in 2003, so it was in very early days- Yeah ... of me being in an academic environment.

Professor Lawlor: I'm still doing my PhD. And it- As well as the sort of, the environment in Bristol was, you know, I look back and I think in some ways that the London job didn't work out was great because there has always been this island environment of collaboration, openness, supporting each other, and things like George and Shah and others being able to see the future.

Professor Lawlor: And you should get George on, he may say, "Didn't you?" Yeah. But my memory of the time was he-- the, the ability to do genome-wide association studies was emerging, and he's from a, a social [00:19:00] epidemiology

Dr. McGavock: background. Yeah.

Professor Lawlor: So he's saying, "We need to know what's going on here. We need to know how this can help us or if it can help us, and how it might be important or not."

Professor Lawlor: And I remember we often have book clubs, and we had a book club which was reading The Rough Guide to Genetics, 'cause that was kind of where we were. That's amazing. And you know, and then George starts talking about this idea of Mendelian randomization and-- but it's like reading it through a lace curtain.

Professor Lawlor: It's not all there. And then you kind of get to the point where you, "Yeah, got it." And then something comes up and you think, "No, I'm still a bit behind that lace curtain." And it, it kind of eventually embeds. So I guess the answer to, you know, suggesting I'm an early adopter, it was those serendipitous challenge things that, you know, I was in the right place at the right time.

Dr. McGavock: So that's a huge lesson for trainees. Like, [00:20:00] don't go to the environment where it's, you know, the biggest center or whatever. You're living in a big city. Go to the place where it's gonna be the most collaborative, the most open science. People are talking about this, like 2003, we mapped the human genome in what?

Dr. McGavock: 2000. GWAS studies are slowly coming out, and then your group is saying, "Oh yeah, wait, watch out. Mendelian randomization's gonna come out." And now when you look at the arc and the number of people using that method, wow, that's amazing. And the reason I ask that is it is a core examples you give here in this r- this pioneering paper.

Dr. McGavock: So we're 10 years out of this paper that you wrote in 2016 called-- that was published in the International Journal of Epidemiology called Triangulation in Aetiolog- Aetiological Epidemiology. It's something that when I first read it, I was like, "We have to do this in our lab. I love this." When I explain it to people, I'm like, "It's applying Sherlock Holmes' methods to understanding causal inference or causal effects."

Dr. McGavock: And so maybe you could just talk to us a little bit about what is triangulation in the world of [00:21:00] epidemiology?

Professor Lawlor: So quite rightly, there are some different views on this, and with any new method you get people critiquing, questioning, and, and that's really important. Just an aside to early career researchers and clinicians, you can often see you get debate in public, in papers, and then I hear people saying, "Oh, so-and-so hates so-and-so," and you go, "No, no, no."

Professor Lawlor: Yeah,

Dr. McGavock: yeah.

Professor Lawlor: They don't. They absolutely don't. So when you see Miguel disputing something that Neil Pearce has said or that George has said, it's healthy and it's good.

Dr. McGavock: Right.

Professor Lawlor: So, um-

Dr. McGavock: They're still going for a pint at the end of the day.

Professor Lawlor: Absolutely.

Dr. McGavock: Good.

Professor Lawlor: Or-

Dr. McGavock: Good to know ...

Professor Lawlor: coffee.

Dr. McGavock: Or coffee, yeah, yeah. Or- Depending on where you're thinking.

Dr. McGavock: Yeah, yeah. Right.

Professor Lawlor: So in that paper, and I think so for me, triangulation is primarily about trying to identify two or more studies [00:22:00] where you're selecting them because they have

Professor Lawlor: differing key sources of bias. So one of the examples in the paper, we talk about that Mendelian randomization, the thing that we're often most worried about is that the genetic variants we're using to instrument the, the exposure we're interested in might include genetic variants that relate to your outcome.

Dr. McGavock: Right.

Professor Lawlor: Then instead of estimating the causal effect or your best go at the causal effect, you're estimating the joint effects of the path through the exposure and any other path that goes around. Um-

Dr. McGavock: So an example there might be, let's say you have 35 SNPs associated with fasting [00:23:00] glucose, 10 of those SNPs are related to BMI And you're thinking about childhood birth weight, right?

Dr. McGavock: Now, you can't disentangle the SNPs that are serving as an instrument for blood glucose from the SNPs that are also potentially driving body weight. Is that right?

Professor Lawlor: Yeah. Let's say what we're interested in is the effect of body mass index on birth weights or some pregnancy outcome, something like that.

Dr. McGavock: Mm-hmm.

Professor Lawlor: Or cardiovascular disease, and you would expect that the genetic variants that reach genome-wide significance for body mass index, some of those will relate to glucose, lipids. So there, that's not really a problem because you're interested in body mass index, and if body mass index [00:24:00] acts through glucose or lipids or blood pressure, then that's all part of the body mass index.

Professor Lawlor: It would be a little bit like, you know, we, we do now Mendelian randomization for drug targets, so, and you can compare that to a randomized control trial. So if you're looking at the effect of statins on cardiovascular disease-

Dr. McGavock: Mm-hmm ...

Professor Lawlor: you wouldn't want to adjust away the lipid effects. So that kind of what gets called vertical pleiotropy, where the genetic variants that do relate to a risk factor for the outcome, a downstream of your main

Dr. McGavock: exposure.

Professor Lawlor: So it's more when you're sort of finding, uh, if you're looking at body mass index and you're looking at cardiovascular disease, inflammatory markers that are not downstream of the body mass index, and also that cause cardiovascular disease. So you've got to have those two conditions. Right. You want them [00:25:00] to be independent of the exposure, not downstream of it, and to be a causal factor for the outcome that you're interested in.

Dr. McGavock: Right. And so MR allows you to say, because we have this instrument, and if we can, in theory, rule out horizontal pleiotropy or rule out some type of pleiotropy, the instrument itself is giving us an unbiased estimate of the exposure on the outcome. And so we're dealing with there's no unmeasured confounding or there's no residual confounding using that method, but it does have limitations.

Dr. McGavock: So there are risks.

Professor Lawlor: Yeah. So definitely just another caveat. We're looking at what's called unbalanced horizontal pleiotropy. So again, if you've got lots of different paths, and just by chance those that increase, that are pleiotropic paths, those that increase the risk of the outcome are of a similar magnitude [00:26:00] to those that decrease, so they cancel each other out.

Professor Lawlor: Mm. Um, it probably doesn't happen commonly, but the sensitivity analyses we do are testing that idea- Right ... that there's no unbalanced one. So for the big problem that we worry about in Mendelian randomization is this horizontal

Dr. McGavock: pleiotropy.

Professor Lawlor: It's biologically not plausible that common confounders like socioeconomic position, behaviors and so on, can causally relate to your SNPs, your genetic variation.

Professor Lawlor: Um, so in general, that's why we think confounding is not a problem. There are... If you've got a population that has, for example, different ethnic groups, then you've got what gets called population stratification. That can cause confounding, which is why most studies have been done on white European ancestries and adjust for ancestral principal [00:27:00] components.

Professor Lawlor: Again, that has downsides because we don't just want to know about white people.

Dr. McGavock: Right.

Professor Lawlor: Um, but if we park that for a second, that's the main thing we worry about in terms of this idea of key source of bias. Whereas in the observational multivariable regression data, w- we ca- more concerned about residual confounders because not all of the confounders have been measured, or they've been measured but not with a lot of detail.

Professor Lawlor: So have you ever smoked? Yes, no. Rather than maybe pack years of smoking, or they're not modeled, you know, if there's nonlinear relationships with the exposure or the outcome. So the idea then with triangulation is if you are able to apply Mendelian randomization and conventional best adjusted multivariable regression to the same research aim or objective, [00:28:00] and you get a consistent answer, by that I mean the direction is the same in both and the magnitude is similar in both, that increases your confidence in saying that's likely to be the causal conclusion that this exposure increases this outcome or decreases it, or has no association with or no effect on it.

Professor Lawlor: Because it just seems unlikely that you'd get the same answer if you've got these two different sources of bias.

Dr. McGavock: Right.

Professor Lawlor: And obviously the more different studies you've got there to triangulate, the stronger you are under the assumption that they've all got different key sources of bias.

Dr. McGavock: So if... We'll go into an example in a minute, but first off, you're answering the same question, you have the same exposure, you have the same outcome, you have the [00:29:00] same population, I'm assuming Like it's not in the same population- Or-

Dr. McGavock: but it's, you're targeting the same group of individuals?

Professor Lawlor: Yeah. So that's broadly I would say yes. There's a really lovely paper which Julian Higgins, Kate Tilling, and Qing Yang Chapman were key people in putting that together, and they very much focus on this, you've got to have the same question, and the key question being very focused.

Dr. McGavock: Right.

Professor Lawlor: In terms of amount of exposure and timing of exposure and things like that. Yeah. And, uh, I think there are objectives in epidemiology where it can be a broader aim, and we can give this, what I've just described as a qualitative answer. So qualitatively they're in the same direction and of similar magnitude.

Professor Lawlor: Whereas the paper I've just mentioned takes a, a very focused approach to the question [00:30:00] and provides a framework. It's in a pre-print form at the moment. It's not been published after peer review, but it provides a framework for being able to pool the estimate On the basis that they are measuring the same things, and also i- put it within that adjustment for the bias.

Dr. McGavock: Mm.

Professor Lawlor: So that's really neat. Again, no perfect study comes with its challenges. So the interesting thing in doing it, and it was amazingly hard work by, I think particularly those three people, but just being involved in that, you get some sense of it. In the end, it was really difficult to work out for the majority of papers whether they were asking the same question when they were asking the question that we decided we were interested in.

Professor Lawlor: So you do the best you can.

Dr. McGavock: Mm-hmm. But ideally you're honing in on a really defined exposure, a really [00:31:00] defined outcome.

Professor Lawlor: Yeah.

Dr. McGavock: And you're finding design studies where the sources of bias are different. So in your paper you give an example. We have a randomized control trial. Are there sources of bias? Of course.

Dr. McGavock: Concealment, retention, intention to treat. Mm-hmm. That there's a strength there. There's also other sources of bias in terms of sample size, selection bias maybe, or a distinct group. You have an observational study with regression. In a randomized controlled trial, you're assuming there might be no residual confounding if the randomization and the methods are okay.

Dr. McGavock: In a multivariate regression, you have residual confounding. Doesn't matter how you match or adjust for, or propensity score, inverse probability weight propensity score match, there's still gonna be unobserved confounding or residual confounding. And then with Mendelian randomization, we already talked about it, there are sources of bias, but they're different in all three studies.

Dr. McGavock: Yeah. And so the direction of the result [00:32:00] in each of those analyses is gonna give you, and this is where the Sherlock Holmes end comes in, is seeing those results together as a package. You know, how do I interpret these different findings? Can you give some examples here about how you would do that? So one is a cross-context comparison.

Dr. McGavock: Can you give us a definition or like what is a cross-context comparison?

Professor Lawlor: Yeah. So, the idea here and the example we talk about with Marie-Jo Brion's work is where you're comparing results, straightforward multivariable regression with the best confounder adjustment that you can do in two different populations where the underlying assumption is that the confounding structures differ between those two populations.

Professor Lawlor: And then if that multivariable regression is compounded in one or both, if [00:33:00] it's driven by confounding where you've got those two compounding differences, y- you'd expect them to be different. Right. Whereas if they're similar, it makes you think, okay, this, you know, in terms of Sherlock Holmes, this is, okay, this is one piece of evidence that is now taking us towards what this similar result is.

Professor Lawlor: And I think the work that Marie-Jo Brion had just done at the time, we were, around the time we were writing the triangulation paper, was a really good example of that. And it started looking at a low- middle income country, Pelotas cohort, and comparing it to ALSPAC, a UK birth cohort. And it was looking at breastfeeding and looking at the long-term effect of that on a range of things, adiposity, cognitive function.

Professor Lawlor: And, uh, again, with that idea that [00:34:00] in Brazil- Breastfeeding was more common than in... A priori, this was a-

Dr. McGavock: Mm-hmm ...

Professor Lawlor: you then have to show it in your data. So I think a priori, the argument was we would expect there to be more breastfeeding in Brazil and less socioeconomic association. Whereas in the UK, we expected there to be less breastfeeding and for socioeconomic position.

Professor Lawlor: So, so basically for those with higher socioeconomic position to have very socially patterned and, and be more likely to be breastfeeding, both initiating breastfeeding and breastfeeding for at least three months, I think were the exposures. Marie-Jo's an example of a PhD student kind of next generation to me, who benefited from that, you know, we're all open, we collaborate- Yeah

Professor Lawlor: we talk, we think about these things, and, and she just did this fantastic. She looked at Pelotas compared to ALSPAC. It sort of looked like [00:35:00] breastfeeding didn't affect later adiposity, but potentially there's cognitive function in a bene- in a beneficial way.

Dr. McGavock: Mm-hmm.

Professor Lawlor: Panel wasn't great, so she finds some more lower middle-income country studies and merges those, and then took the analysis from this amazing Belarus study, uh, which is a randomized controlled trial that was done because at the time Belarus had the lowest rates of breastfeeding, and then it got followed up by Ch- oh, I can't remember his first name, but his second name is Kramer.

Professor Lawlor: Um-

Dr. McGavock: Michael Kramer. Canadian.

Professor Lawlor: Yeah.

Dr. McGavock: And in season one- Yeah ... we did an interview with his co-investigator, Robert Platt. Yeah.

Professor Lawlor: Fantastic. So he set up this randomized control trial, and then the long-term follow-up, and you could see there again that when you pull all of that together, really no strong evidence that breastfeeding reduces your risk of obesity or-

Dr. McGavock: Or

Professor Lawlor: high blood pressure

Professor Lawlor: but it [00:36:00] does seem to increase cognitive function.

Dr. McGavock: Right. So let me try to summarize this, 'cause one of my grad students, Nicole, and I presented this at a journal club, and the confusion was huge. So first off, if we look at breastfeeding in the UK or in Canada or in the US, there's a huge socioeconomic gradient.

Dr. McGavock: Mm-hmm. And in those studies, you know, like we have here, we have the child study, you have this wonderful study called ALSPAC. You measure, let's say, household income or your position, your job class, white collar, blue collar, and you can say, absolutely there's this association. Individuals in higher income groups are more likely to breastfeed longer and more exclusively.

Dr. McGavock: Mm-hmm. And those who are in the lower income class are less likely to initiate breastfeeding, less likely to have it exclusive b- breastfeeding, and do it less long. So in your association study, you're like, oh my goodness, we see this clear, powerful association or s- very robust association between breastfeeding and body weight at five.

Dr. McGavock: But because they're so tightly correlated, is that [00:37:00] just a function of all of the factors that come with affluence? Yeah. How do we isolate the association to breastfeeding? Now, in the Pelotas cohort, which is in Brazil, everyone's breastfeeding more often, initiation and for longer periods of time, so there's no clear gradient.

Dr. McGavock: So now if you see a positive association between breastfeeding, the range would be different, I assume, like the variation in breastfeeding exposure is different because it's a different cross-context. Across the gradient of breastfeeding, if you don't see an association, then you would assume that breastfeeding is not associated with that, and there isn't an association with income, right?

Professor Lawlor: It, y- yeah. So-

Dr. McGavock: It's a way to disentangle- Yeah, so- ... income from breastfeeding ...

Professor Lawlor: yeah. So the idea is, you know, we a priori, or Marie-Jo was leading this, separated [00:38:00] these two cohorts. And it's, again, because we have access to these data, 'cause we've been working with Victora and the Pelotas cohort for some time. So we have the access, and we think this is very likely, and we think it's likely because, you know, we know what the situation is like in the UK, and Alicia and Cesar and the group in Pelotas know what the situation is like in Brazil.

Professor Lawlor: But we tested that it was true that socioeconomic position, by a number of different measures, strongly related to breastfeeding in the UK and didn't

Dr. McGavock: in Brazil. And nor did Pelotas.

Professor Lawlor: Yeah. So it, that's not really an outcome. We're kind of setting it up so that if we had found that wasn't the case, we would've still wanted to publish it, but said, "This is not how our sort of idea behind this doesn't work."

Dr. McGavock: Yeah. Go find another context. Yeah. But, so if there is an association, now you have an example where there is no confounding by socioeconomic [00:39:00] status. So if that association holds up, then you've isolated breastfeeding from socioeconomic status. If it doesn't hold up, then this is the Sherlock Holmes piece, it's likely that the association you saw it in the UK, in the ALSPAC cohort, is probably driven by unmeasured confounding related to probably socioeconomic status.

Dr. McGavock: That's only- Yeah, yeah ... two studies. Your third study is, well, let's find a randomized control trial and see what they find. And then lining up those three studies, again, now you're interpreting whether or not there's an effect. And in that paper, which we'll link to, there is an effect in all three studies on IQ, so there's something very powerful there.

Dr. McGavock: But again, all the other biological factors, there's no association. And the beauty of that, for me, is it just took some Sherlock Holmes-ing, some relationships, some fearlessness to reach out to other cohorts and say, "We want to answer this question really, really well. We don't want to just rely on the ALSPAC data," right?

Dr. McGavock: "So can we come together on this study?" And so I thought it was a beautiful example of doing that. [00:40:00]

Professor Lawlor: Credit to Marie-Jo, I think

Dr. McGavock: Oh, credit to Marie-Jo. Absolutely. So your more recent paper, which was published in 2024, looks at multiple lines of evidence for maternal BMI on pregnancy related outcomes and perinatal outcomes, and I think the advantage here that we've adopted in our lab that I really, really like, that Nicole did a wonderful job on, one of the grad students in my lab, is laying out your framework in a figure to show exactly where, in a pictorial format, where you see that those sources of bias and how you've- Yeah

Dr. McGavock: specifically chosen these designs to say there's a different source of bias. And then also a key aspect of your paper is to say, "We predict this," right? So you make an assumption. Based on the bias structure in this cohort, if our question, very specific question, is true, we would expect the results to go this way.

Dr. McGavock: If the answer to the question is no, then we would expect our results to go this way. So do we want to walk through that paper a little bit as an [00:41:00] example and make sure we acknowledge the PI, M- Maria Carolina Borges, who led this. So do you want to kind of walk us through the different ways of that you tried to-- You had three s- specific studies, a multivariate regression, Mendelian randomization, and paternal negative control.

Dr. McGavock: So what was the question? Do you remember?

Professor Lawlor: Yeah. So again, it's the obesity thing. Some of the obesity causes everything, and there's a lot of literature, w- with or without varied e- evidence. But the collaboration that Carolina and I set up and drove me wrong was to try and enable Mendelian randomization and other methods to be applied

Dr. McGavock: to a number of pregnancy and perinatal outcomes,

Professor Lawlor: so that you have the bigger picture.

Professor Lawlor: You know, taking higher BMI as an example here, being able to say to women, couples, midwives, physicians, "Well, you know, [00:42:00] this suggests-" reasonably robustly that higher body mass index increases these outcomes, decreases these outcomes, has no effect on these, and is too uncertain was that we had a fourth category in that paper-

Professor Lawlor: we really can't say here. So that gives the picture to, I think, women couples. They'll sort of want to know that. So once we've got that collaboration up and running, this is one of the first papers, and it brings together a lot of different birth cohorts. And the three... And I, like you, I love the pictorial abstracts and I can't do them, but people like-

Professor Lawlor: yeah, really good. It's that it

Dr. McGavock: might make a whole- Me neither. Me neither. Trainees are really good at

Professor Lawlor: it. Yeah. Yeah. But they're great, and I think, you know, again, I have a passion about research that is both written and presented in a language, in a way that is understandable to a wide variety of people, the general public, [00:43:00] academics in other disciplines, and so on and so forth.

Professor Lawlor: So I really like those. And here we looked at Mendelian randomization, so looking at the body mass index In pregnant women using genetic variants related to body mass index with an instrumental variable for that, and then looking at this range of outcomes, pregnancy perinatal outcomes.

Dr. McGavock: And so those included like- Mm-hmm

Dr. McGavock: stillbirth, miscarriage, perinatal depression, gestational diabetes, large for gestational age- Yeah ... APGAR scores, breastfeeding duration. So a, a whole whack of outcomes.

Professor Lawlor: Yeah. And then comparing those results directly with the multiva- standard multivariable regression analyses of the measures and the pregnancy and perinatal outcomes.

Professor Lawlor: And then lastly, doing the multivariable regression with the partners on the basis [00:44:00] that for many of these outcomes, we wouldn't expect the partner to have a direct effect.

Dr. McGavock: Mm-hmm.

Professor Lawlor: We would expect the partner's body mass index to be correlated with the same confounders as in the mother's examples, but we would expect any associations would be null or, or much, much weaker than what we would see in comparison to the mother's own body mass index during pregnancy.

Dr. McGavock: Can I just highlight that for just one second- Yeah ... just to explain, 'cause this is so critical, that if you have this negative exposure, so in this case, we're trying to isolate the exposure to the in utero exposure. So if you bring in- Yeah ... any information about dad, he is bringing in with him the same confounding structure genetically.

Dr. McGavock: Not exactly the same, but close. He's also bringing in environmental confounding, [00:45:00] but he's a negative exposure in that he doesn't have a uterus. So Yeah ... any association with dad suggests residual confounding is present in any of the associations that you find. So the confounding structure is exactly- Yeah

Dr. McGavock: the same, but the exposure is different, which is a wonderful example, and I think not enough people are using these negative exposure controls, especially in the life course, because a lot of times it's possible to use where people are just not using it enough to isolate those two factors.

Professor Lawlor: Yeah. That's interesting 'cause it is, uh, very feasible to do.

Professor Lawlor: As always, there are some caveats, there are some things to do. The father's exposure is somewhat unrelated to the mother's, and so you have to mutually adjust the mothers and fathers when you do these analysis, which breaks all the rules of DAGs, but completely sure that's what we have to do. And, you know, it might...

Professor Lawlor: That sort of idea of assortative mating and eating the same thing, so that's why we have to have that caveat and do the mutual adjustment. [00:46:00] But yeah, fundamentally, that's the case. And what we also do is some kinds of sensitivity analysis. We have less information on the fathers. Yeah. Fewer numbers of the partners-

Dr. McGavock: Yeah

Professor Lawlor: fathers than the mothers. So you kind of also want to compare the mothers when you restrict it to the same population of the father, the same data for the partner.

Dr. McGavock: Right. And so then you lay out this wonderful sort of forest plot where you're comparing the results from a multivariate regression to the results from a Mendelian randomization study, the results from a multivariate regression to a paternal control, and this is where the Sherlock Holmes then comes in a bit, right?

Dr. McGavock: This is unconventional. Most readers, you're looking at one point estimate with a confidence interval. Here you have multiple ones. They're going in different directions. Now you go to Sherlock Holmes, the conclusion, right? And, and I think that's the beauty in this, is it tells a much more [00:47:00] robust story. You don't have to go pull 20 papers in to answer the question.

Dr. McGavock: Everything's laid out there for you. So do you remember some of the key findings that came out here?

Professor Lawlor: Yeah. So, you know, some of them, I think people would generally go, "Go figure."

Dr. McGavock: Yeah.

Professor Lawlor: As in we saw that higher body mass index related to higher risk of hypertensive disorders of pregnancy and the components gestational hypertension and pre-eclampsia.

Professor Lawlor: Similarly, higher risk of gestational diabetes, um, and, and also higher risk of large for gestational age or macrosomia, large birth

Dr. McGavock: weight. So the result looks like both studies have... are demonstrating a positive association. So the MR study, no residual confounding, positive association. Multivariate logistic, positive association, lots of unmeasured confounding.

Dr. McGavock: But even though we adjusted for unmeasured confounding in the other study, it still looks exactly the same. Very conclusive finding. [00:48:00]

Professor Lawlor: Yeah. And then when we look at the comparison of the moms with their partners, we see that positive effect in the mother to those outcomes, and w- we don't see it in the fathers or the partners.

Professor Lawlor: And the reason I was sort of smiling and saying, oh, it's like that goes figure, but- That was really important to see that.

Dr. McGavock: Hmm.

Professor Lawlor: So we-

Dr. McGavock: It's like a calibration ...

Professor Lawlor: it would be... Yeah. Yeah. So it kind of gives us confidence then in, in some of the other things we saw. So we saw higher body mass index was associated with lower breastfeeding, and there have been studies showing that before, but I think this gave it a more robust flavor.

Professor Lawlor: We didn't see any evidence for a relationship with perinatal depression, which is interesting. Mm-hmm. But because we have this kind of what we expected to see quite strongly, gave us a bit more confidence to think about that and, you know, future work to try and maybe understand why that's the [00:49:00] case. And then the one that was, well, we're not sure, was preterm birth, where, and I may get it the wrong way around, but there were different direction.

Professor Lawlor: It wasn't just different magnitude, different direction from the Mendelian randomization and-

Dr. McGavock: Yes ...

Professor Lawlor: the multivariable regression. I think firstly, this is a collaboration with a large number of pregnancies, but for some things we lack some power. So, you know, we didn't have much power for the preterm birth.

Professor Lawlor: But also, I think that certainly what we have seen in observational linked record studies across different populations is there's a definite difference between spontaneous and pregnancy-induced, medically-induced preterm birth. Mm-hmm. But in the, when we did this paper, we really didn't have the power to separate those two out, and so I think [00:50:00] the outcome there was combining two very different outcomes where you've had a reason to induce the pregnancy or have a planned cesarean section compared to where none of the midwives or the couple or the obstetricians are expecting this baby to come out early.

Dr. McGavock: So yeah, that, those dichotomies are areas to go digging, right? Yeah. Where you see that, oh, wow, like this went in the opposite direction. We need to go understand what's happening there. I was really surprised the maternal BMI was associated with a lower risk of breastfeeding initiation, and that was seen both with MR and with traditional logistic regression, and it was a null finding for the paternal control.

Dr. McGavock: So there was something going on about body weight. Mechanistically, we don't know why, but an interesting finding there.

Professor Lawlor: Yeah, that's been seen observationally that women who are larger are less likely to breastfeed, and whether that [00:51:00] has, for some, a kind of almost mechanical reason or whether it's the kind of maybe embarrassment to breastfeeding.

Professor Lawlor: I don't know. Mm-hmm. But yeah, something for further investigation. It's, you know, again, a, a- colleague who worked in Bristol and who's now back in Italy where she's from originally, Luisa Zuccolo, who appears very closely with this.

Dr. McGavock: Yes. Yeah. Met her in Winnipeg as part of a-

Professor Lawlor: Oh, nice ...

Dr. McGavock: with Meghan Azad. Yeah.

Professor Lawlor: Of course, yes.

Dr. McGavock: Yeah. Yeah. So, we'll hopefully we'll have both of them on at some point to talk about their interesting life course work. So these are great examples. Just to follow up on the paternal control, what you're looking for there is a positive association. So for example, in Nicole's thesis, which hopefully will be published and we'll link to this, we found that exposure to paternal diabetes was associated with the offspring's risk for hypertension.

Dr. McGavock: A lot of people look at that association and say, "Oh yeah, of course, like if dad has diabetes, there's probably some underlying [00:52:00] genetic thing that he has that's transferred over to his child, and his child's got a higher risk for hypertension." And we're like, "Yeah. Exactly. It's not in utero," right? It's, it's so funny that you're used to seeing just one association.

Dr. McGavock: You're not used to seeing the different associations and comparing them. Mm-hmm. So it's really, like on the right-hand side of that figure one, in blue you see all the paternal controls, and they should all be null. Everything should be null. Any positive thing is saying to you in the triangulation approach, this is residual confounding.

Dr. McGavock: This is some unmeasured confounding that is clustering with the family in some way, and so that you should account for that when you're interpreting the other point estimate that is just above it, right?

Professor Lawlor: It's why in that paper, the two figures had to be separated.

Dr. McGavock: Mm.

Professor Lawlor: And the first figure, where you think there's an effect, whether it's increasing or decreasing, your expectation is the two are in the similar direction and similar point estimates.

Professor Lawlor: Where yeah, you then go to the negative control [00:53:00] and you're expecting it to be null.

Dr. McGavock: Right. It's

Professor Lawlor: negative in the sense that we don't expect something to be going on here.

Dr. McGavock: And most of the time you're not excited about a null finding.

Professor Lawlor: Exactly.

Dr. McGavock: But in this case, the null finding is helping you answer your question, which I think is great.

Professor Lawlor: I was just gonna say, if we've got a well-formulated aim question that's justified, I think null findings can be just as important as positives.

Dr. McGavock: Totally agree. I shouldn't have said that. I do not want anyone to be excited about positive. But yes, a, a null finding should be as exciting if the question is well-designed, and that I think in a grant, like I say this to students all the time, if you write a grant where whether you prove your hypothesis is true or untrue, the answer is important either way No matter what the result is of your experiment, people are gonna be fascinated by it and say it's important

Professor Lawlor: And it's often beneficial public health or clinically.

Professor Lawlor: So the sort of evidence suggesting vitamin D doesn't cause cardiovascular disease, then people don't [00:54:00] have to buy vitamin D tablets for the prevention of cardiovascular disease. Or showing that having conceived your child via IVF, ART, doesn't mean they're going to be obese later is quite reassuring and useful.

Professor Lawlor: Or, or other work we've done, which is really important, is evidence where, again, a great PhD student, Kurt Taylor, the sort of triangulation suggested that things like smoking and alcohol didn't increase the risk of your child having congenital heart disease. And for the parents and the patients we work with, that's so important.

Professor Lawlor: I mean, anybody with children knows that you spend a lot of your life worrying that you've harmed

Dr. McGavock: them in some way. Yes, exactly. We wrote that in the grant. It's so important for this work on gestational diabetes and long-term cardiovascular risk to be able to say to parents, "You know, there's no association here, and if your child develops high blood pressure, you don't have to worry [00:55:00] about having diabetes in pregnancy."

Dr. McGavock: There is a great paper, I think, from one of your colleagues about reducing the shame. Stop blaming the parents, right?

Professor Lawlor: Yeah.

Dr. McGavock: And I- That's

Professor Lawlor: Gemma, Gemma Sharp.

Dr. McGavock: Gemma Sharp. Yeah.

Professor Lawlor: Yeah. Right. Exactly.

Dr. McGavock: Yeah. Now, the last thing about triangulation that I really like from this paper, and you graphically represent this beautifully, is the timing of the exposure.

Dr. McGavock: And again, the value of an MR or an observational study is you can get the lifetime exposure without doing a randomized control trial. So looking at-- One example here was looking at blood pressure and coronary heart disease risk, where you can see the effect of a 50-year exposure, and then it, you line it up with the effect of a shorter randomized control trial or a shorter prospective study.

Dr. McGavock: So where does the timing c- fall in, in this triangulation approach?

Professor Lawlor: So, the paper you're referring to was in our triangulation paper and- It was in your paper, yeah.

Dr. McGavock: Yeah ...

Professor Lawlor: but it's a paper by Brian Ference and, and I, he ... just thought it was an amazing paper- Elegant ... and for many reasons. He didn't use the [00:56:00] word triangulation.

Professor Lawlor: It fits the criteria, so it was a nice example to put in, in the original paper, and he was comparing the effect estimates of blood pressure with cardiovascular disease. And he used a big prospective collaboration cohort, multi-variable regression. He used Mendelian randomization, and he used randomized control trials as an instrumental variable.

Professor Lawlor: Mm-hmm. So it was a published systematic review and meta-analysis, and they had not said, "What's the effect of antihypertensives on cardiovascular disease?" Coronary heart disease more specifically. They had said, "What's the effect of variation in blood pressure- Mm-hmm ... on coronary heart disease?" And that's a nice comparison.

Professor Lawlor: There are other examples that we and other people have done where you take the [00:57:00] randomization arm, and if you've got the target of the intervention, so here blood pressure as well as the longer term outcome, you're doing exactly the same as most Mendelian randomization analysis does. You're doing this instrumental variable analysis.

Dr. McGavock: Mm-hmm.

Professor Lawlor: And it made me ... because for this triangulation paper, I wanted to get clear about the ... and because Brian hadn't called it a triangulation paper and he hadn't talked about the different sources of bias, so I put together a table that brought that out. And it was so interesting reading that systematic review that was doing this instrumental variable because they used wordings in the discussion like the fact that we get this consistent effect, and again, they have a forest plot and

Professor Lawlor: but we get the consistent effect of you reduce blood pressure by this amount, and this is how much you reduce the relative risk of cardiovascular, coronary heart disease. So the fact that we see that across [00:58:00] all the different types of antihypertensive tells us two things. It's not about the mechanism of a specific hy- antihypertensive, and there's no pleiotropy in these studies.

Professor Lawlor: And I've never heard that word used other than in genetics- In that context, yeah ... and it's really inter- yeah- Yeah ... really interesting. So what the paper shows first of all is all three methods show that lower blood pressure, lower risk of coronary heart disease. But the actual estimates are quite different with the randomized control trial results having the weakest association effect, the multivariable regression being intermediate, and the Mendelian randomization being the strongest.

Dr. McGavock: Right.

Professor Lawlor: So he then plots them on [00:59:00] his estimation, and these things have some assumptions under them, of the amount of exposure to the lower blood pressure. And that's shortest in the randomized control trial because the randomized control trials are short in the follow-up- They're short, yeah ... measured. Five years, six years.

Professor Lawlor: And then it's intermediate in the observational study, which was based on the length of follow-up, the average length of follow-up across these different studies.

Dr. McGavock: 13 years.

Professor Lawlor: And then for the Mendelian randomization, he assumes that your genetic liability is lifetime. Yeah. And therefore, that's measuring your lifetime lower-

Dr. McGavock: Yeah

Professor Lawlor: blood pressure. 55 years. And it sits on this nice diagonal.

Dr. McGavock: Oh, beautiful.

Professor Lawlor: I think, yeah,

Dr. McGavock: yeah. It was, yeah. It was beautiful. It was so great that you put that in there as an example, and then you did the same thing with gestational diabetes further down- Yeah ... cumulative exposure to gestational weeks. So just to close out here, what triangulation is not.

Dr. McGavock: So it doesn't have to be three. You don't have to [01:00:00] actually make a triangle. But it has to be at least two.

Professor Lawlor: Yeah.

Dr. McGavock: And it's not multiple studies. So a meta-analysis doesn't answer this question. You could have 50 studies that are all the same design, but they all have the same bias structure. So a point estimate in a meta-analysis is just a replication of the same bias if it's there.

Professor Lawlor: Y- yes. And this work I mentioned earlier does sort of pool as you might in a meta-analysis, but it's pooling different types of studies. Mm-hmm. So it's pooling, again, randomized control trial as an instrumental variable, multivariable regression. That sort of idea of let's pool all the results from all randomized control trials that have looked at- Right

Professor Lawlor: the effect of this intervention on this outcome isn't triangulation. It's really useful and important, um- It is

Dr. McGavock: important ...

Professor Lawlor: but it, it's not the same as triangulation.

Dr. McGavock: And pooling or meta analyzing data from observational [01:01:00] studies all pointing in the same direction could still be replicating the same residual confounding, so- Yeah

Dr. McGavock: it could be one of multiple studies in your triangulation approach. And then it's the interpretation of the final data that's the key element of the triangulation, and explaining in detail how to interpret the results from multiple studies with different sources of bias or different confounding structures.

Professor Lawlor: Yeah. Setting out at the start what you think the key sources of bias are-

Dr. McGavock: Mm-hmm ...

Professor Lawlor: and what direction you think they will go in, in this particular example that the data that you're using. And sometimes you're seeing that original triangulation paper, you're seeing others where w- you just have to say, "We don't know."

Professor Lawlor: Mm-hmm. You know, we're concerned about selection bias or, you know, certain other things, but we don't know whether that would [01:02:00] bias away from the null, towards the null, towards the expected hypothesized direction or not. And you just have to be upfront about that.

Dr. McGavock: Yes. And are there any exposures where triangulation is particularly useful for studying certain exposures?

Dr. McGavock: So thinking about target trial emulation, sometimes you think like, we need to be able to manipulate this exposure. With a triangulation approach, you don't necessarily need that manipulation assumption or caveat.

Professor Lawlor: I think of target trial emulation as another tool in our toolbox that can contribute to triangulation.

Dr. McGavock: Yeah.

Professor Lawlor: So there's lots of different ways in which triangulation can be done. The fact that we have studies like UK Biobank, all of us in the States- Mm-hmm ... Our Future Health now being, you know, ongoing in UK, where you have observational data and you have within [01:03:00] that the ability to, again, I don't mean a majority of them, but to do at least two comparisons, if you've got to do Mendelian randomization and multivariable regression within UK Biobank for some questions.

Professor Lawlor: I know Chechen has said that that's the best form of, of triangulation because you know it's in the same sample and you're, you're not bringing in noise where actually you think you've got the same sort of population, but you may not. It, it has disadvantages for different reasons. But yeah, I think it's Trying to identify these differences that you think may alter the, may, if the bias is driving it, then will drive the estimate.

Dr. McGavock: And so the paper that we were just talking about that was published in BMC Medicine had 13 cohorts in it. There's [01:04:00] so many more, especially in the developmental origins world, so many more cohorts where you do have a measured exposure and a genetic risk, and you're creating resources for trainees or anyone to be able to use these large data sets together.

Dr. McGavock: So there's so many opportunities out there for trainees to dive into this. They should really think about it.

Professor Lawlor: Yeah. I think so. I would say where you can do more than one thing in a study, then that's useful. But just going back to the target trial emulation, you need quite big studies for that, and so far it's predominantly done in electronic health record linkage studies.

Professor Lawlor: I have in my head that what would be nice would be to have almost these open triangulation studies that people could put things in, and you know, that may be something a reality going forward, but you also have to be pragmatic. And I'm at the stage in my career now where the work is predominantly being done by PhD students and postdocs, and they have to get publications that often it's a more readable [01:05:00] paper if you don't try and put everything in.

Professor Lawlor: So we've just talked about the paper we, this MR-PREG collaboration that we have, and there are things we can do in that, including around treatment during pregnancy. And then, and we're doing that work, and then we've also got a lot of electronic health record linkage studies, and we're using that with negative controls within sibling analysis, target trial emulation, conventional.

Professor Lawlor: But it's an unwieldy paper in practice if we try and put all of that. We've tried it. We just

Dr. McGavock: can't People don't understand it. So- It's so hard.

Professor Lawlor: Yeah. You lose

Dr. McGavock: the story. Yeah.

Professor Lawlor: But when your PhDs are more generally with grants, it's kind of you, you have to have two or three papers maybe with the different sorts of data, but then a paper at the end that brings it all together kind of.

Professor Lawlor: Yeah, I think we've got to have space for reviews of triangulation or,

Dr. McGavock: or something like that. Oh, that's great. Yeah, that would be beautiful. [01:06:00] You read one paper, chapter one, chapter two, chapter three, summary. I don't know how editors of journals appreciate four papers coming in at once. Yeah. But I like that.

Dr. McGavock: It would be great. I really like that.

Professor Lawlor: Yeah.

Dr. McGavock: Yeah. And for a PhD thesis, 100%, so. Yeah. Well, that was amazing, Deb. I think we'll wrap it up here. Is there anything else we missed? Anything else you want us to point in certain directions that trainees should be looking at in terms of triangulation?

Professor Lawlor: No, I think it's been nice talking in general about openness, who you work with. There is no perfect study, but we do the best we can.

Dr. McGavock: Yes. And you are an absolute legend in this field, and to hear you say that sometimes when you're reading a new paper or a new field, that it's like reading it through lace and you have to read it three or four times, I can't tell you how satisfying that is for me to hear, that I'm not the only one.

Dr. McGavock: That even the giants in the field are still trying to digest. Yeah. Like when I was reading MR, it took me so long for it to kick in. So I- Yeah ... can't thank you enough. Hopefully we get to see you soon in Canada, and keep doing the amazing work you're doing. It's, it's just so great. I look forward to [01:07:00] reading anything that comes out of your lab, so thanks again for doing the amazing work that you do and the contributions to science.

Dr. McGavock: Great.

Professor Lawlor: Well, thank you very much. Thank you.

Dr. McGavock: I can't thank Professor Lawlor enough for making time for this fascinating conversation. I hope this conversation gave you a deeper appreciation for what triangulation really looks like when it's done really well, not as a checklist, not as a buzzword, but as a genuine intellectual practice of how you design studies with complementary biases and honestly weighing the results from each of those studies, seeing where they converge, seeing where they don't, and making causal inference based on those findings.

Dr. McGavock: We covered a lot of ground today. I'll have links in the show notes to all the key papers that we discussed, including the seminal 2016 triangulation review in IJE that we focused on, as well as some of the studies that were referenced in that paper. My particular favorite is the breastfeeding cross-comparison study, as well as the work that she led [01:08:00] on maternal BMI and pregnancy outcomes using Mendelian randomization and paternal negative controls.

Dr. McGavock: Importantly, you should take time to read that thought-provoking article, It's the Mother, It's the Mother piece, but I think that's foundational in how we understand life course epidemiology and epidemiology, uh, around perinatal health. We'll also link to Deb's, uh, faculty page and some of the work done by George Davey Smith and Shah Ebrahim that launched her career, as well as their foundational 2003 Mendelian randomization paper that argued for this tool being a cornerstone for causal epidemiology.

Dr. McGavock: If you enjoyed today's episode, please share it with a colleague. You can also go back to season one and listen to the two episodes that we did on causal epi with Dr. Robert Platt. Or you can go back to our previous episode on natural experiments with Dr. Claire Boone from McGill University. Like we say in the past, if causal inference is your thing, we have more episodes coming up, and, uh, we look forward for you to join us for those conversations.

Dr. McGavock: If you have any questions, comments, suggestions for future [01:09:00] guests, please reach out to us and we'd love to hear from you. Until next week, where we're going to dive into the topic of implementation science,

Professor Lawlor: what it is and what it isn't. See you next week.

Dr. McGavock: The Emerging Scholar Podcast was recorded in the UMFM studios located at the University of Manitoba in Winnipeg, Manitoba, Canada. This is Treaty 1 territory, the original homeland of the Anishinaabe, the Ininew, and on the national homeland of the Red River Métis. Manitoba is also homeland to the Anishininew, the Dakota Oyate, and the Dene peoples.

Dr. McGavock: We acknowledge that our water is sourced from Shoal Lake 40 First Nation. As a settler scientist working in health research, I acknowledge TRC Call to Action number 18, that health inequities between Indigenous and non-Indigenous people are the direct result of colonization. Winnipeg is the center for reconciliation in Canada, and where we are speaking from today is at the [01:10:00] crossroads of Canada's history and Canada's future.

Dr. McGavock: We are three kilometers away from one of the most sacred meeting grounds for Indigenous people on the plains and within Turtle Island. We are also several kilometers away from one of the last residential schools in the province. We are working collectively on reconcili-action with youth, elders, and community leaders to heal from the trauma of colonization, past and present, honor the original treaties, and create a Canada that addresses and eliminates the inequities and injustices of the past and present to create a future for Indigenous children to live and thrive.

Dr. McGavock: On behalf of the team at the Emerging Scholar Podcast, we want to thank Osani Balkaran for the great production for our opening and closing themes. We also want to thank UMFM 101.5, especially our executive producer, Jared McKetiak, for technical assistance and production on this project. We also want to thank all of our guests for the discussion and the insight that they've provided.

Dr. McGavock: [01:11:00] Once again, thank you for listening. Be sure to subscribe to the Emerging Scholar Podcast on Apple Podcasts, Spotify, and umfm.com.