Epidurals: Are labor epidurals really linked to autism?

[Regina] (0:00 - 0:12)
You know, Kristin, side note, every time you say leftover confounding, I am picturing leftover pizza, cold pizza the next day, leftovers, so yummy. So I just need to get that out of my head.

[Kristin] (0:19 - 0:29)
Welcome to Normal Curves. This is a podcast for anyone who wants to learn about scientific studies and the statistics behind them. I'm Kristin Sainani.

I'm a professor at Stanford University.

[Regina] (0:29 - 0:35)
And I'm Regina Nuzzo. I'm a professor at Gallaudet University and a part-time lecturer at Stanford.

[Kristin] (0:35 - 0:40)
We are not medical doctors. We are PhDs. So nothing in this podcast should be construed as medical advice.

[Regina] (0:40 - 0:45)
Also, this podcast is separate from our day jobs at Stanford and Gallaudet University.

[Kristin] (0:46 - 0:49)
Regina, today we are going to talk about autism.

[Regina] (0:49 - 0:55)
Autism. Oh, no. Is this going to be controversial or are we going to get some unhappy fan mail?

[Kristin] (0:55 - 1:13)
Let's hope not. This is a podcast about statistics. And we try to follow the evidence and stay clear of politics from any side.

And to be clear, we are not doing autism and vaccines today. We are not looking at autism and Tylenol either. We may tackle those eventually.

[Regina] (1:13 - 1:15)
All right. Eventually, but not now.

[Kristin] (1:15 - 1:29)
Today we are looking at a really interesting story that listeners might not have followed in the news already. This one is on autism and epidurals given for labor. It's a fascinating story.

I also happen to know some of the players involved.

[Regina] (1:30 - 1:40)
Oh, that makes it more fun, definitely. So epidurals, I just want to explain the terminology for people who don't know. That is pain relief when you're giving birth.

Exactly.

[Kristin] (1:40 - 2:09)
And, Regina, this story has larger lessons. So even if you don't care about epidurals per se, there are some bigger takeaways here for autism research. You know, autism is a really hard thing to study.

For one thing, it involves people's children, which understandably makes people very emotional. Also, autism isn't one single disease. It's a broad diagnostic category that lumps together a lot of different conditions.

[Regina] (2:10 - 2:24)
Which I think really surprises people. We talk about autism like it's one thing, right? But it's not.

I think of it like cancer. Cancer is actually many different diseases all under one umbrella label. Exactly.

[Kristin] (2:24 - 2:54)
And that heterogeneity makes it especially hard to study cleanly. Also, autism is statistically a thorny knot to untangle. There are a lot of confounders, and it's hard to statistically separate those confounders from causal effects.

And that brings us to the statistical topics that we're going to cover today. Cox regression, statistical adjustment, inverse probability of treatment weighting, unmeasured and residual confounding, and sibling analyses.

[Regina] (2:54 - 2:58)
Ooh, Kristin, I am so impressed. These are really juicy today. Nice job.

[Kristin] (2:59 - 3:19)
Oh, yeah. So the claim that we're going to examine today is that epidurals given during labor, during childbirth, cause autism.

This claim first showed up in a paper published in 2020, and so that's where our story begins. And we're also doing something a little different today on Normal Curves. We have a guest coming up later.

[Regina] (3:19 - 3:41)
Right. Our first guest, who is not at the moment getting drunk for science. And to understand what I mean by that, everyone should go and listen to our Holiday Guide Episode 1, where we experimented on some intrepid journalists at the Science Writers 2025 Conference in Chicago, where, yes, we tried to get them drunk.

[Kristin] (3:42 - 4:22)
And Regina, today's guest, the science journalist Laura Dattaro, was also at that conference, though we did not get her drunk. But today's episode comes out of her reporting for a publication called Spectrum, which focuses on autism research. And I should note that that publication is now part of a larger publication called The Transmitter that covers neuroscience in general.

I know Laura from when she was a senior reporter at Spectrum. I used to occasionally help Laura and other reporters there with statistical questions. Her piece on autism and epidurals was published in Spectrum in 2023, and it won a healthcare journalism award for excellence.

So we'll hear from her later about her experience reporting this story.

[Regina] (4:23 - 4:27)
Which is fabulous. And we will put a link to her piece in the show notes.

[Kristin] (4:27 - 4:42)
Absolutely. So this story starts with a paper published in JAMA Pediatrics in 2020. The authors reported that kids whose moms had an epidural during childbirth were more likely to later be diagnosed with autism.

[Regina] (4:42 - 4:49)
Hmm. Okay. So this 2020 paper was the first study to look at this link.

Had anyone looked at it before?

[Kristin] (4:50 - 5:09)
No, not really. At least nobody had looked at it carefully before. So the finding was a surprise.

And Regina, JAMA Pediatrics is not just some low-tier random journal. It is a top pediatrics journal. So this got noticed, and it was controversial from the start.

[Regina] (5:09 - 5:18)
Right. But Kristin, before we get into all of that controversy, why don't you just give us some background, tell us the details of the study itself.

[Kristin] (5:18 - 5:32)
Sure. This was what we call a retrospective cohort study. And this is a design that uses data that were already collected.

And they were usually collected for other purposes, like data in electronic medical records.

[Regina] (5:32 - 5:49)
Right. We talked about retrospective cohort studies in the shingles vaccine episode. And that was the one with the big database of medical records that was called the Optum Database.

And we described that basically as shopping for data at Costco.

[Kristin] (5:50 - 6:20)
Retrospective cohort studies are kind of an easy study design because the data already exists. So, for example, here, the autism diagnoses have already happened. That means researchers don't have to wait for children to develop autism.

And often the data are already de-identified and packaged for research, like with Optum, so researchers may not need ethics approval to do these studies. And Regina, sometimes I think people do them just to pad their CVs with extra publications.

[Regina] (6:20 - 6:45)
Wow. That is a hot take, Kristin. You've got your opinion.

I do see your point, though. Although a lot of studies are done this way, and they can be useful, the data were not collected for the purpose of research. They were collected for patient care and medical billing.

So the data are not always the cleanest data in the world.

[Kristin] (6:46 - 7:13)
Yeah, we have to keep that in mind, Regina, although I might just be grumpy because I end up reviewing a lot of these studies. Also, you should know that there are platforms coming out where an AI can do most of the work for a retrospective cohort study. So now it's going to be even easier for researchers to churn these things out.

Although to be honest, Regina, I'm not sure that the AI will do any worse than a lot of humans on this, at least from the papers I've reviewed.

[Regina] (7:15 - 7:28)
Another hot take. You're right, Kristin. We still need to be cautious because now with AI, people could churn these kinds of studies out with even less thought than before.

[Kristin] (7:28 - 7:51)
Yeah, that is definitely a risk. Absolutely. All right.

In this study, they did not use AI. It was back from 2020. And they did not use the Optum database.

They used data from Kaiser Permanente, Southern California. And this is a big database of people getting their health care through the Kaiser HMO. It makes it easy to link birth records to outcomes like autism.

[Regina] (7:52 - 7:58)
Wow. 143,000. That is a lot of babies.

That is a big sample size.

[Kristin] (7:58 - 8:25)
Yeah, we love big samples, though, Regina, it's important to keep in mind that big samples improve precision, but they don't magically get around statistical bias or confounding just because they're big. So in these records, they had data on which moms got an epidural during labor and which didn't. And they also had data on which babies went on to get a diagnosis of what we technically call autism spectrum disorder.

[Regina] (8:25 - 8:33)
Ah, so that is the technical name for autism to recognize the fact that, as we mentioned, it covers a lot of things.

[Kristin] (8:33 - 8:38)
Exactly. But for simplicity today on this podcast, we're just going to call it autism.

[Regina] (8:38 - 8:46)
OK, so the researchers were able to find this label, this diagnosis of autism from medical billing codes, correct?

[Kristin] (8:47 - 9:28)
Exactly. Which we need to keep in mind means that there could be mistakes, right? These codes may not be perfect.

So we could be misclassifying some babies. And that's one of the big problems with retrospective cohort studies. Now, for each baby, they followed them in the medical records from their birth until the baby was either diagnosed with autism or they died or they left Kaiser or until the study end date, which was December 31st, 2018.

So for each baby, the outcome is basically for what amount of time was that baby followed? And in the end, did they get an autism diagnosis or not?

[Regina] (9:29 - 9:44)
Right. We call this time-to-event data because it has this binary outcome part and the time part. And we talked about this in some depth in the episode on exercise and cancer.

[Kristin] (9:44 - 9:54)
Right. And we talked about in that episode how the methods that we use for analyzing time to event data, those are called survival analysis methods.

[Regina] (9:55 - 10:04)
And that episode was the one in which we voted people off the island as if it were a game show to illustrate survival analysis.

[Kristin] (10:04 - 10:13)
That's exactly right. Fun episode. Definitely check it out.

Before we dive into the results, Regina, can I complain to you about something in the paper that was driving me nuts?

[Regina] (10:13 - 10:16)
Well, even if I say no, you will anyway.

[Kristin] (10:16 - 10:50)
Yeah, that was kind of a rhetorical question, Regina. You got me. First, though, I'm going to give the authors credit because the paper was nice and concise and I really enjoy concise writing.

In fact, I teach a whole course on Coursera called Writing in the Sciences, which is largely about writing concisely. But they did commit one of my biggest pet peeves, Regina, whenever they referred to the word exposure in their paper, instead of just saying epidural, they instead replaced the word epidural with an acronym, LEA, L-E-A.

[Regina] (10:50 - 10:53)
And what does LEA stand for?

[Kristin] (10:54 - 11:23)
Labor Epidural Analgesia. And that is the technical term for an epidural given during labor. But they could have just given the technical term at the beginning of the paper and then referred to it thereafter as an epidural.

By using the acronym LEA, every time I saw that, I was thinking of a girl's name, not a medicine. And this is one of my biggest pet peeves, how much we overuse abbreviations and acronyms in research papers. It makes papers really hard to read.

[Regina] (11:24 - 11:37)
Amen. Epidural here, especially, is a very straightforward term. No need to disguise it with an acronym.

But are you done ranting, Kristin? Are you ready to tell us what they actually found?

[Kristin] (11:38 - 11:58)
Yes. All right. Let's start with some simple numbers.

First of all, about three quarters of the moms had an epidural, which to me means that about three quarters of the moms are smart. Because honestly, why would you give birth without one? I mean, I am crazy enough to run hard marathons, but I am not crazy enough to forego the epidural.

[Regina] (11:59 - 12:02)
Even you have your limit, huh?

[Kristin] (12:03 - 12:17)
Absolutely. Yes. All right. They divided the moms into two groups, epidural, no epidural. In the epidural group, 1.9% of the babies went on to an autism diagnosis versus 1.3% in the no epidural group.

[Regina] (12:18 - 12:43)
OK, let's look at those numbers again. 1.9 minus 1.3, that's a difference of 0.6%, which taken out is six per 1,000 babies. So we could say an additional six babies per 1,000 went on to be diagnosed with autism in the epidural group compared with the no epidural group.

[Kristin] (12:43 - 13:01)
Right. But, Regina, these proportions are a little crude because they don't take into account time. And remember, we have time to event data.

So if we want to be more precise, we need to use survival analysis methods, and here they used Cox regression.

[Regina] (13:01 - 13:19)
Kristin, I know you're pausing because you're waiting for me to make some Cox jokes, but I feel like I got all my Cox jokes out of my system already in that previous episode. So I'm just going to restrain myself here and be quiet.

[Kristin] (13:20 - 13:27)
Thank you for that restraint, Regina. And I should point out that Cox is a person's name. It's spelled C-O-X.

[Regina] (13:28 - 14:05)
About Cox regression, though, the not funny but very relevant part, see, I can be serious, is that it's the regression technique for time-to-event data, right? Cox regression models, they give us something called hazard ratios, and I don't know of a way to make that dirty or funny. Hazard ratios is a ratio of two rates of some events that we care about.

And so here, Kristin, I assume the model is comparing the rate of autism diagnoses in the epidural group versus the rate of autism diagnoses in the no epidural group.

[Kristin] (14:06 - 14:16)
That's it exactly, Regina. And when they put the data into Cox regression with no other variables in the model, they got a hazard ratio of 1.48.

[Regina] (14:17 - 15:06)
1.48, which we would interpret as a 48% increase in the rate at which babies in the epidural group get autism, which actually, Kristin, lines up pretty closely with the numbers that you gave us before, because before that was 1.9% had an autism diagnosis in the epidural group versus 1.3% in the no epidural group. So that's kind of similar. But we need to remember, of course, that that hazard ratio is more precise because it accounts for the time of diagnosis.

That's important. And just those simple proportions ignore time. That's why we like hazard ratios.

But overall, Kristin, it looks like more autism in the epidural group, right?

[Kristin] (15:06 - 15:18)
Yeah. But this, of course, does not mean that epidurals cause autism, because this is classic. Correlation is not causation.

There are really important confounders here.

[Regina] (15:18 - 15:24)
Right. Correlation, not causation. We know that.

Tell us about the confounders. Do you have any examples?

[Kristin] (15:24 - 15:40)
Sure. For example, more educated, wealthier mothers are more likely to get an epidural. And also, women who have longer labors or more complicated births or bigger babies, they are also more likely to get an epidural.

[Regina] (15:41 - 15:49)
OK. All of that does make sense. But then when you think about it, all of those factors could also affect the risk of autism.

[Kristin] (15:49 - 15:58)
Exactly. So there is strong potential confounding here, which is why we have to do something statistically to try to account for that confounding.

[Regina] (15:58 - 16:13)
This is like a big theme of our podcast and of statistics and of these papers, right? Strong confounding. We have to do something to account for it.

So are we going to get into some fun things? How did they account for confounding here?

[Kristin] (16:13 - 16:21)
We are going to get into a lot of fun things, Regina, because they actually addressed confounding with two different methods.

[Regina] (16:21 - 16:28)
Oh, hey, this is a plot twist. All of a sudden, I'm sitting up straight. Now I want to hear more about it.

OK. Two methods.

[Kristin] (16:29 - 16:49)
Yeah, we find this really cool that they did it two different ways. All right. So method one, they used traditional multivariable adjustment.

That basically means they just threw all the confounding variables into the Cox regression model. And of course, these were things like age, education, health of the mother, and the size of the baby.

[Regina] (16:50 - 17:00)
So this is the typical way of approaching confounders. We've talked about this multivariable adjustment in a number of episodes. Now I'm very curious about method two, though.

[Kristin] (17:00 - 17:08)
So method two, they did inverse probability of treatment weighting, also called IPTW.

[Regina] (17:08 - 17:34)
Oh, IPTW, our old friend. So we saw this back in the shingles episode again, except I remember we thought IPTW was a boring, jargony name. And so I accepted your challenge to come up with a better acronym to rename it.

And remind me what we renamed it, Kristin. I'm a little scared.

[Kristin] (17:35 - 17:41)
Yeah, we renamed it SEX for Statistical Exception Expansion.

[Regina] (17:44 - 18:09)
Of course we named it that. What else would we name it? OK.

But that is actually brilliant, though. I had completely forgotten this. OK, Statistical Exception Expansion.

You've got to be a little creative with those letters there, but it actually works in this case because we are expanding the exceptions, which I think we're about to talk about.

[Kristin] (18:10 - 18:41)
Exactly. And speaking of brilliance, Regina, that does remind me, I want to give a shout out to Phil Kearney, who wrote a lovely blog post promoting another mnemonic that is a Normal Curves original. And that was from our Holiday Guide 1 episode.

We made up the SMART framework, S-M-A-R-T, to help walk Uncle Joe through the results section of a research paper. And thank you, Phil, for helping us make SMART go viral. And Regina, I'm hoping it will become as famous as PICOT.

[Regina] (18:41 - 18:44)
Well, it's certainly better and easier to remember.

[Kristin] (18:45 - 18:45)
Exactly.

[Regina] (18:45 - 18:54)
SMART. Yes, that was your brainstorm. I loved it.

And that was the episode where we got the journalist drunk. So that was just a great episode all around.

[Kristin] (18:54 - 19:13)
We'll put a link to Phil's post in the show notes as well. All right, back to sex, IPTW. The idea of this method is that we take two groups that are really different and we try to balance them by upweighting the oddballs in each group.

In other words, the statistical exceptions.

[Regina] (19:13 - 19:36)
The statistical exceptions, exactly, the oddballs. And we explained this in the shingles episode, I remember, with a great analogy using the Barbie and Oppenheimer movies, the Barbenheimer effect, because those movies are way more fun than most epidemiology examples. But the analogy, surprisingly, it works.

[Kristin] (19:37 - 20:20)
Yeah, that was a brilliant analogy on your part, Regina. A little refresher for everyone. The idea here is that the crowds for those two movies are really different.

So if we just had each group rate the movie they saw, that would be an unfair comparison. But we can rebalance the groups by effectively counting some people more. So toy example, let's say there's one teenage girl and 10 middle-aged men in the Oppenheimer movie and the reverse in the Barbie movie.

You would then weight the teenage girl in Oppenheimer 10 times and the middle-aged man in Barbie 10 times. So on paper, both crowds would end up with the same mix of ages and sexes.

[Regina] (20:20 - 20:25)
Exactly. That's what we mean when we say we are upweighting the oddballs in each group.

[Kristin] (20:25 - 20:49)
Yeah. We artificially make the groups balanced in the confounders and then we put this adjusted sample into the Cox regression. So both of the methods they used, multivariable adjustment and IPTW, these are both different ways to remove confounding.

But in both cases, they are running a Cox regression and what we get out is an adjusted hazard ratio.

[Regina] (20:50 - 21:08)
Adjusted hazard ratio. So just a moment ago, Kristin, you told us the hazard ratio was 1.48, but that was before accounting for confounding. So when they use these two methods to try to adjust for confounding, how did that number change?

[Kristin] (21:09 - 21:17)
Method one gave an adjusted hazard ratio of 1.37 and method two gave 1.38. Nice.

[Regina] (21:17 - 21:40)
They basically gave the same answer, which we like. We're happy to see that because both methods, they're really aiming to do the same thing. So it's giving us more trust in the results to see that whatever results we get out of that Cox regression, they're not completely dependent just on exactly which of those two methods you chose to use.

[Kristin] (21:40 - 22:01)
Yeah, we like it when things line up mathematically like this. And it's also a little bit of a lesson, Regina, because sometimes we get really excited about newer, fancier methods like IPTW, but these don't always gain us anything over more old-fashioned, tried-and-true methods, you know, other than sounding fancy.

[Regina] (22:02 - 22:41)
Well, that's not nothing. Sometimes the peer reviewer likes the fancy-sounding method.

[Kristin]
Yeah, you've got to impress the reviewer two, exactly.

[Regina]
You do. You do. Uh-huh.

Okay. So let's look at those numbers again. Okay.

You said 1.37, 1.38, that is a 37 or 38 percent increase in the rate of autism after controlling for all of those characteristics. Now, before controlling for all those characteristics, confounding, we had a 48 percent increase. So the fact that it went down from 48 percent to, you know, 37, 38 percent, that is telling us that confounding, that stuff that we talked about, was indeed an issue.

[Kristin] (22:42 - 23:06)
That's right. But these hazard ratios were still statistically significant, or as you like to say, Regina, statistically discernible. So the authors concluded that the association persists even after adjusting for confounding.

Now, of course, Regina, they didn't come out and say epidurals cause autism. They put the obligatory cautionary line in the discussion.

[Regina] (23:06 - 23:32)
But it sounds like they were kind of implying that there is a causal link. So I can imagine, Kristin, someone reading this, looking at this, they could say, well, look, there's a statistically significant result after adjusting for confounding and epidurals clearly happened before the autism diagnosis. So why can't we just say this is a demonstration or evidence of causality?

What gives?

[Kristin] (23:32 - 23:57)
Right. I can see a lot of people walking away thinking that, Regina. But as we've talked about on this podcast before, statistical adjustment is not magic.

And just because we have, quote, adjusted for confounding, it doesn't mean that we have actually stripped away all of the confounding. We always have to worry about unmeasured and residual confounding.

[Regina] (23:58 - 24:22)
Like Botox. That's how we described it in previous episodes. I still think that analogy works.

So statistical adjustment is effective, just like Botox is effective. But Botox cannot remove all of your wrinkles, sadly. And statistical adjustment cannot remove all of your confounding.

You're still going to have that residual and unmeasured, wrinkly confounding.

[Kristin] (24:24 - 24:50)
Exactly. So, Regina, my first reaction in reading this paper is, oh, this is probably just residual or unmeasured confounding. And it turns out that was a lot of people's first reaction when reading the paper.

The paper got a lot of attention, and it also got a lot of criticism from the scientific community. And that was actually a lot of what Laura wrote about in her article for Spectrum.

[Regina] (24:50 - 25:35)
Ooh, OK. I want to hear about all of the criticism and the blowback, but after the break.

Welcome back to Normal Curves.

We're looking today at a 2020 paper that claimed to find a link between epidurals for labor and autism. And Kristin, you were about to tell us about the blowback that this paper got.

[Kristin] (25:35 - 26:03)
So Regina, this is where it gets interesting, because a lot of researchers, anesthesiologists, OBGYNs, they saw this 2020 paper and felt it was pretty weak evidence. And they were really worried that it might mislead the public and scare women away from this very effective and safe pain relief during labor. So there were a flood of letters to the editor and online comments for this paper.

[Regina] (26:03 - 26:12)
Ooh, I imagine they did not want a rehash of that whole vaccines and autism scare fiasco.

[Kristin] (26:13 - 26:29)
Yeah. One of the letter writers wrote, similar to persistent skepticism related to the safety of vaccines, we are concerned that it may be difficult to reverse false notions, even with contradictory scientific evidence.

[Regina] (26:29 - 26:40)
I get that now, but I'm getting the impression that criticisms were not just about this possible scare, right, in the general public. What scientific concerns did they raise?

[Kristin] (26:41 - 26:57)
Yeah, I mean, the bottom line was that people found the evidence quite unconvincing. One of the letters actually said that they had, quote, strenuous concerns about the conclusions of this study implying a causal link between epidurals and autism.

[Regina] (26:58 - 27:24)
Hmm, strenuous concerns. That is very beefy, meaty language, especially for academia. And maybe I've been spending too much time on Reddit, but I can't help but to try to translate what this would look like, subreddit.

And I'm picturing something like, no, hard pass or the side-eye gif, you know, the one I'm talking about. I love that one.

[Kristin] (27:24 - 27:44)
No, I actually don't. I'm not schooled enough in social media, clearly, Regina. But yes, strenuous concerns, that is really harsh language for academia.

And one of the strenuous concerns that critics raised was that this association does not have a lot of biological plausibility.

[Regina] (27:45 - 27:49)
Oh, I've been wondering about that. OK, why not? Like, what other argument?

[Kristin] (27:50 - 28:21)
Yeah, so scientists believe that the key brain changes associated with autism actually happen during fetal development, which is, of course, before birth. Epidurals are given during birth, so the timing does not make a lot of sense. Also, the medication is delivered locally to the mom around the spine and hardly any of it reaches the baby.

And even if a little bit did reach the baby, the amount is likely to be really tiny and the exposure window is short. So it's hard to imagine how this could lead to autism.

[Regina] (28:21 - 28:28)
OK, I feel like that is, that's very plausible, what they're saying, that it's biologically implausible. Yeah.

[Kristin] (28:29 - 29:14)
The authors of this 2020 paper did have one biological hypothesis, but they ended up disproving it themselves. So epidurals are known to cause fever. And in their study, about 12 percent of the epidural group got a fever versus only about one percent of the no epidural group.

And fever could plausibly harm the baby. But in their study, even though getting the epidural was strongly linked to fever, getting a fever was not at all linked to autism. So that cannot be the mechanism.

Besides biological plausibility, Regina, the other major problem that critics pointed out is something we already alluded to, the worry about unmeasured and residual confounding.

[Regina] (29:15 - 29:25)
Right. OK, unmeasured, residual confounding. Let's talk about that for a moment.

Unmeasured confounding means we did not measure it. So we can't even try to account for it.

[Kristin] (29:25 - 29:59)
That's right. This study, for example, did not account for some really important, what we call perinatal factors. And those are things that happen around labor and delivery, like birth complications, fetal distress or oxygen problems.

And these may be important confounders because they may lead a woman to get an epidural, but they also may be related to autism because a baby who is already having problems, they may have more complicated births, too. And Regina, they did not account for these perinatal factors at all.

[Regina] (30:00 - 30:02)
Oh, that seems like a big weakness in the study.

[Kristin] (30:02 - 30:30)
It is. Yeah. And then, of course, Regina, there's always residual confounding.

So as an example here, we talked about how maternal demographic factors like education and income are important confounders. But of course, we cannot measure those things perfectly. We don't have precise data on someone's exact economic situation.

So if we don't measure these variables perfectly, we cannot adjust for them perfectly. And there's always going to be some leftover confounding.

[Regina] (30:31 - 30:43)
You know, Kristin, side note, every time you say leftover confounding, I am picturing leftover pizza, cold pizza the next day, leftovers so yummy. So I just need to get that out of my head.

[Kristin] (30:45 - 31:29)
Yeah. Leftover pizza, really fun. Leftover confounding, not fun.

Regina, you know, something unusual also happened with this paper that Laura wrote about in her piece. On the same day that the paper was released, this was October of 2020. On that same day, a whole bunch of major medical bodies issued a joint statement saying that, hey, this JAMA Pediatrics paper does not provide credible scientific evidence that epidurals during labor cause autism.

And they tried to reassure everyone that epidurals are safe for both mothers and infants. These medical bodies were groups like the Society for Pediatric Anesthesia and the American College of Obstetricians and Gynecologists.

[Regina] (31:30 - 31:56)
Wow. OK, so no slouches when we're talking about these medical groups. And the fact that their statement, their joint statement, came out the same day as the paper.

First of all, how did that happen? They got an advance copy of the paper and there were enough doctors in these groups that were concerned about the paper. And the conclusion that they managed to get together and all agree on a statement to send out immediately.

And I think that really said something.

[Kristin] (31:57 - 31:58)
That is no small feat. Yeah.

[Regina] (31:59 - 32:20)
And Kristin, just thinking back on everything that we've been talking about today with this paper, it's interesting because the concerns that we are talking about here, they're a little different than the typical concerns that we encounter in this podcast when we're going through papers. It's not that there are errors. So can you talk a little bit more about that?

[Kristin] (32:21 - 32:58)
Yeah. A lot of times what we're pointing out in papers, Regina, are what we call statistical cockroaches. This paper did not have any statistical cockroaches that we found.

We think they did their modeling correctly and soundly. They did two different methods. So the methodology is not wrong.

It's just inherently weak. The misstep here is not in the statistics per se, but it's in how we interpret the results. So a lot of scientists looking at this would simply say, yeah, it makes sense that these two things go together, but there's no good reason to think that the relationship is causal.

[Regina] (32:58 - 33:14)
So with something like this, the next natural thing to do would be to try to replicate the whole thing, replicate the results, maybe adjust for those extra confounders like the perinatal factors you talked about earlier. So what about replications? Has anyone done that?

[Kristin] (33:14 - 33:50)
Oh, yes. There actually have been numerous replication attempts now. This got a lot of attention and people wanted to see, did they find this in their own datasets?

And this is good. This is how science is supposed to work. Regina, the first replication came out actually very quickly in 2021, and it was published in the same journal, JAMA Pediatrics.

Fun fact, the senior author on that paper is Dr. Alexander Butwick, and he is a professor of anesthesiology at Stanford University. And he also happens to be one of my best former students.

[Regina] (33:51 - 33:54)
Well, then the statistics ought to be amazing.

[Kristin] (33:54 - 34:24)
Oh, yeah. I mean, I'm going to take full credit, of course, for the robustness of the statistics in that paper. Actually, Regina, Alex was one of my best students ever.

He is super sharp and was just an all-around star in my classes. And anecdotally, I have to say, the anesthesiologists in general have been some of my best statistics students in my several decades of teaching at Stanford. Regina, I think it must just be very competitive to get into anesthesiology.

[Regina] (34:25 - 34:28)
They also have to do a lot of math, actually, I believe.

[Kristin] (34:28 - 35:12)
Oh, yeah, that's a good point, Regina. So maybe a lot of them are just naturally quantitatively inclined. They also have the good drugs and they are your favorite people during labor.

So Alex partnered with a researcher from Canada and they used data from the province of Manitoba. They had about 123,000 births and they used a very similar design to the 2020 study because they were trying to replicate it. So singleton births, vaginal deliveries and births from 2005 to 2016.

One interesting difference between the populations, though, Regina, that I noticed, only 38 percent of the women in this data set got epidurals. And that's about half the number from the American paper.

[Regina] (35:12 - 35:16)
Oh, that's interesting. I guess different health care systems, different practices.

[Kristin] (35:16 - 35:48)
Yeah, I found that kind of fascinating. Just like the earlier paper, they compared the epidural and no epidural groups. But big difference here, they had data on those perinatal factors.

And so we can actually see in their descriptive table that there are big differences in these perinatal factors between the two groups. For example, about 10 percent of the births without an epidural had what we call fetal distress compared to 20 percent of the births with an epidural.

[Regina] (35:48 - 35:50)
Oh, that is a huge difference.

[Kristin] (35:50 - 36:15)
It is. And we don't think this is because the epidural causes fetal distress, but rather that you are more likely to end up with an epidural when the birth is complicated and you are having problems like fetal distress. But of course, fetal distress may also be related to autism because maybe sicker babies are more likely to have fetal distress and also more likely to be diagnosed with autism later.

So this is exactly the kind of confounder we need to adjust for.

[Regina] (36:16 - 36:19)
So it's good that they controlled for this. What did they find?

[Kristin] (36:19 - 36:30)
In the raw, unadjusted data, 2.1 percent of kids in the epidural group were later diagnosed with autism versus 1.7 percent in the no epidural group.

And that corresponded to an unadjusted hazard ratio of 1.25.

[Regina] (36:30 - 36:42)
1.25. OK, so that is a 25 percent higher hazard in the epidural group.

[Kristin] (36:42 - 36:58)
Exactly. So just like in the 2020 paper, in the unadjusted analysis, they are finding an elevated hazard in the epidural group, although the magnitude of the hazard ratio is a little smaller here. Now, interesting statistical detail, Regina, for those nerds out there.

[Regina] (36:58 - 36:59)
Out there? In here.

[Kristin] (36:59 - 37:30)
Yeah. OK.

OK. Juicy detail for us, too, Regina. When they adjusted for confounders, they only used IPTW, not multivariable adjustment, but they did it iteratively.

So they did the whole IPTW analysis several times, meaning they rebalanced the groups several times, progressively accounting for more and more confounders. And this is nice because we can see what happens as they adjust for more things.

[Regina] (37:30 - 37:46)
Oh, I love this. It's like stop action animation. You see every every stage in the development.

And it also is showing their work. So you see how these things reveal themselves. So what they find?

[Kristin] (37:46 - 38:12)
So, Regina, when they adjusted for the basic socioeconomic factors like income, the hazard ratio didn't really change. It was 1.28. But when they added pregnancy and mom related factors like maternal health and smoking, the hazard ratio dropped to 1.15.

[Regina] (38:13 - 38:24)
Oh, 1.15. So now we are down to a 15 percent relative increase in the rate of autism.
So already we're seeing that a lot of that original association, that increased hazard rate, it's getting explained away by those confounders.

[Kristin] (38:24 - 38:38)
Exactly. And in the final model, when they additionally adjusted for perinatal factors, things like the fetal distress, the hazard ratio dropped to 1.08 and was no longer statistically significant.

[Regina] (38:38 - 38:56)
Oh, wow. 1.08 is just an 8 percent relative increase in the rate. And again, you can actually see the confounding being stripped away in action in real time.

And it left us with what is basically then just a null result.

[Kristin] (38:57 - 39:32)
Right. A hazard ratio of 1.00 would mean a zero percent increase in the rate of autism because when the groups have equal rates, that means you are dividing two equal things and you get a ratio of exactly one. 1.08 is very close to 1.00. So it's almost completely null. Regina, after this IPTW analysis, they did another analysis that's even stronger, what's called a sibling analysis.

And this is where they compare siblings within the same family where the mom had an epidural for one child, but not for the other.

[Regina] (39:32 - 39:54)
Oh, I love it. So it's like they are their own control in a way. So that can control for a ton of the really important shared confounding that we couldn't easily measure.

Like genetics, right? Or family environment and all these things that you can imagine really do strongly affect autism diagnoses.

[Kristin] (39:54 - 40:30)
It is a stronger design. But, of course, you need a huge data set to do this because you only get information from families where the epidural status differed across births. Right.

You need enough moms who switched it up, had an epidural with one child, but not the other. And that's not common. But with 123,000, there's enough of those pairs in the data set.

And in that sibling analysis, the hazard ratio was 0.97. So below one and not statistically significant, meaning the baby who got the epidural compared to their sibling who didn't, they did not have a higher risk of autism.

[Regina] (40:30 - 40:33)
So basically no effect at all.

[Kristin] (40:33 - 40:43)
Right. 0.97 is pretty much 1.0. No hint of an association. So the conclusion from Alex's paper was they did not find an association between epidurals and autism.

[Regina] (40:43 - 40:51)
Wow. So a total opposite conclusion from the previous paper in the same journal in just the last year.

[Kristin] (40:51 - 41:04)
Yeah, exactly. And this made this a really interesting story for Laura to report on. Interesting side note, Regina, when Alex's paper was published in JAMA Pediatrics, it garnered an editor's note.

[Regina] (41:05 - 41:15)
Ooh, fancy. Getting an editor's note in a journal. Oh, that is, I don't know, kind of like your book, getting a book review in the New York Times.

It's a big deal.

[Kristin] (41:16 - 41:52)
It is. And that's a great analogy, Regina. The editor pointed out that this was the first time they had published two conflicting papers within 12 months.

And I really liked his note, so I'm going to read part of it. He says, science is an imperfect and iterative process. And our responsibility as journal editors is to manage the process as best we can.

Publishing two conflicting studies in such a short time frame serves as testament that we recognize the process for what it is. For now, my personal assessment is that the association is yet to be definitively established. If a more definitive study is done, JAMA Pediatrics will publish it.

[Regina] (41:52 - 42:05)
Oh, I love this. So he's a good writer, this editor, but I love also how he's recognizing that this is how science is supposed to work. There's supposed to be conflicting studies and that's OK.

That's a good thing.

[Kristin] (42:06 - 42:20)
Yeah, follow up studies are a good thing. And Regina, this story does not end here because after Alex's paper, there were a slew of additional replication attempts and even some meta-analyses.

[Regina] (42:20 - 42:56)
OK, so follow up studies right after the break.

Welcome back to Normal Curves. Today, we're looking at a 2020 paper on epidurals and autism.

And now we are getting to the slew of follow up studies.

[Kristin] (42:57 - 43:10)
After Alex's paper, a bunch of other groups published similar studies. I'm going to highlight a few of them because the pattern is pretty striking. Two more papers came out in 2021, and they were both published in the same issue of JAMA.

[Regina] (43:10 - 43:18)
Oh, JAMA, that is even one tier above JAMA Pediatrics in the prestige hierarchy.

[Kristin] (43:18 - 43:50)
It is, yes, because we've now dropped the pediatrics part. So this is now all of medicine. Quick summary of the two studies.

One study was from Denmark, about half a million births. And in that study, the unadjusted hazard ratio was 1.29. After full adjustment, it dropped to 1.05. The other study was from British Columbia, almost 400,000 births. Unadjusted hazard ratio was 1.32. Adjusted dropped to 1.09.

[Regina] (43:51 - 44:49)
Wow, that is amazing consistency. So I've been I've been taking notes here. Right. I got all your numbers because I know you like your numbers.

So if we include Alex's paper, we can look at the unadjusted hazard ratios, the raw ones. And we have 1.25, 1.29, 1.32, which is just beautiful. All right.

And then after statistical adjustment, we are at 1.08, 1.05, and 1.09, which is again, it's just gorgeous. It's beautiful to a statistician. It might not be obvious to everyone, but we can look at those numbers and we feel reassured because we know that they are similar enough.

We're going to have a little bit of variability, sample to sample, you know, study to study, but they they're so close enough that it gives us more trust in the conclusions and the pattern.

[Kristin] (44:49 - 45:39)
Yeah, from our point of view, this consistency is actually really amazing that it's so tight. And remember, we have different countries, different health care systems, different rates of epidural use, different autism rates. But this same basic pattern repeats and they are so close.

It's amazing. Regina, both of those additional 2021 papers also did sibling analyses, just like Alex did. The results from the sibling analyses, they got 1.05 and 1.07 as their hazard ratios. Remember, Alex's paper was 0.97. So they're all, you know, hovering right around 1.0, no association. But we're not done yet, Regina. Another paper came out in 2022, and this was a massive study from Scandinavia, Norway, Finland and Sweden, about 4.5 million births.

[Regina] (45:39 - 45:45)
I think the technical term for that is ginormous. That is a ginormous sample size.

[Kristin] (45:46 - 46:43)
Ginormous, yes. Huge. I feel like we should make a penis joke here.

[Regina]
Look at you. I didn't even go there.

[Kristin]
OK, not going there.

All right. But in that huge study, the adjusted hazard ratio, I'm starting with the adjusted hazard ratio here, that was 1.12. Mm, again, consistent. Yeah, we've got like 1.09, 1.05, 1.08, 1.12. So very consistent. In this case, though, Regina, this hazard ratio was statistically significant, meaning statistically different than a null hazard ratio of 1.00. And that's just because there's such a large sample size here that we have extreme precision. For the nerds out there and in here, the confidence interval was 1.10 to 1.12. 1.04.

[Regina] (46:43 - 47:02)
Oh, yes. Really narrow, right? The nerds can see that. So, Kristin, it looks like we are getting here some statistically discernible results, statistically significant results.

And that is probably due to the huge samples just giving us really high precision, the ability to really discern small effects.

[Kristin] (47:02 - 47:16)
Yeah, exactly. And, you know, Regina, they also did a sibling analysis with this ginormous data set. And the hazard ratio there was 0.98, very close to one and not statistically significant.

[Regina] (47:17 - 47:32)
So no association in the sibling analysis, I feel like that's kind of tying it together. No association in the sibling analysis and a significant but tiny one in the non-sibling analysis that could just be residual confounding.

[Kristin] (47:33 - 48:32)
Yeah, exactly. And, you know, this 2022 study lines up very nicely with another study from 2024. So I'm just going to share this last study.

So once a bunch of individual studies had been published on this topic, then people started doing meta-analyses where they pool the results from multiple studies. The most complete meta-analysis that I saw was from 2024 in the Journal of Pain Research. Regina, they pooled data from seven studies, including the original 2020 paper and the three papers from 2021 that we talked about.

This did not include the data from the Scandinavian study. The pooled hazard ratio came out to be 1.12, statistically significant. So exactly the same as the Scandinavian study.

Regina, they also tried taking out the 2020 study because that one had a much bigger hazard ratio than the rest of them. So it was a bit of an outlier. And when they pooled the remaining six studies, the pooled hazard ratio was then 1.09, but still statistically significant.

[Regina] (48:33 - 49:07)
OK, so looking over all of the evidence so far, we now have two very large studies that suggest that after you adjust for confounding, we have a tiny increase in risk that's statistically discernible, statistically significant. And the magnitude is about a 9 to 12 percent relative increase in hazard. And it's very consistent.

We've seen that. So now the question that we ask ourselves is, could this be a small but real causal effect?

[Kristin] (49:07 - 50:16)
Yeah, you know, Regina, that is a possibility that we have to consider. We are getting a tremendous amount of consistency across studies. Think back even to Alex's study.

He had an adjusted hazard ratio of 1.08. That is very close to 1.09 and 1.12. That one did not come out statistically significant, but he did not have a ginormous sample. He had a large but not ginormous sample, so it wasn't quite as precise. So, Regina, there could be a real but very small effect here.

We can't rule that out. I personally think, though, that the more likely explanation why we're getting hazard ratios close to but not quite one is that what we are seeing is probably just residual confounding. Remember, we can never perfectly remove all confounding.

We are always left with a little bit of leftover confounding. And to me, the sibling analyses are more convincing because that design precludes a lot of confounding. And remember, in the sibling analyses, none of those results were statistically significant, and all of the estimates were even closer to one.

So that suggests to me that the 1.09, 1.12 that we're seeing, those might just be residual confounding.

[Regina] (50:17 - 50:54)
You know, Kristin, all of this reminds me of the vitamin D episode where we talked about how hard it is to study vitamin D. And it seems like it's also really hard to study autism in observational designs because autism, from what we know, seems to be really intertwined with other variables. And with vitamin D, didn't we liken it to a big ball of string, right, where it's really hard to untangle just one thread from that entire ball?

And it seems like autism is so complicated, it might be the same thing.

[Kristin] (50:54 - 51:21)
Yeah, I think it's very similar, Regina, except for vitamin D, we had a way out. We had a way to cut all the strings because we could do randomized trials for vitamin D. We can't do randomized trials here because women don't want to be randomized to get an epidural or not, right?

It's a pretty personal decision. And even if a woman was willing to be randomized, if she got put in the control group, there's a good chance that she might change her mind halfway through.

[Regina] (51:21 - 51:26)
I can imagine there might be a lot of noncompliant in the control group once the pain starts hitting.

[Kristin] (51:26 - 51:44)
Yeah, definitely. So no one is going to be doing a randomized trial of that anytime soon. And this is true with a lot of autism research.

We're kind of just stuck with observational data. So we do the best we can statistically, but we have to interpret the results, keeping confounding in mind.

[Regina] (51:45 - 51:57)
So, Kristin, I think now is a good time to bring in our pre-recorded discussion with Laura and share her take on this entire saga.

[Kristin] (51:57 - 52:05)
Yes, I'm going to play that now.

So, Laura, welcome and congratulations on being our first podcast guest who is not getting drunk for science.

[Laura] (52:06 - 52:11)
Thank you. I'm really excited to be here. This is actually my first podcast interview.

[Kristin] (52:11 - 52:14)
First of many more, I'm sure. So, Laura, tell us a little bit about yourself.

[Laura] (52:14 - 52:31)
Yeah, so I'm a freelance science journalist based in New Jersey. So I cover lots of things, but my main focuses are in neuroscience and I do science and health policy. But I've also written about appendicitis a couple of times and baby sharks.

And that's kind of the fun part about being a freelancer.

[Kristin] (52:31 - 52:32)
Now I want to hear about the baby sharks.

[Laura] (52:33 - 52:36)
We don't know where they're born.

That’s the story.

[Regina] (52:37 - 52:42)
That is fascinating. I'm not sure we can cover that on Normal Curves, but I want to learn more about it anyway.

[Kristin] (52:42 - 52:44)
Regina, there might be statistics behind that somewhere.

[Regina] (52:44 - 53:14)
There might be, and we can play the baby shark song.

[Laura]
Oh, the baby shark song was actually what led me to the baby shark story because my grown up brain was looking for something to think about while listening to that song over and over and over and over again. And so then it just made me want to write a story about baby sharks.

And then I found this story about how we don't know where the baby sharks are. And some guy invented a thing that you put in the mommy shark so that when they give birth, it sends a signal to a satellite.

[Kristin]
Wow.

Yeah, I love that story.

[Regina]
We'll link to this.

[Kristin] (53:15 - 53:30)
Absolutely. So, Laura, tell us a little bit about how you got wind of this story, because the original paper came out in 2020 in JAMA Pediatrics that we've been talking about. And then your story was in 2023 after a lot of the follow up was done.

How did you get tipped off onto this story?

[Laura] (53:31 - 54:45)
Yeah, so when the first study came out, my colleague at the time, Peter Hess, who was another reporter at Spectrum, covered the study and some of the initial, you know, skepticism and research response. And then at the time, I was writing a newsletter called Null and Noteworthy for Spectrum, which covered null results and replications.

[Kristin]
Oh we love this.

[Laura]
Yeah, because it was just, they don't often get news coverage. And I wrote this monthly and I had all these Google alerts for finding null results and stuff to include in the newsletter. And it was like every few months I was like, and yet again, we have a study that says there's no association between autism and epidurals.

And so when you're a journalist and you keep seeing something over and over again, eventually your instinct is like, what is going on? And so I just wanted to dig in a little bit. And actually my initial story idea, the story I thought I was going to find was this study came out about autism and epidurals and a bunch of people heard about it and got nervous and now they don't want to do epidurals anymore.

I just wanted to see if that had happened because that was the initial fear around the study. I also, in the interim, went on maternity leave. And so I came back from maternity leave and I was like, I want to write this story now.

And I had had an epidural, which was fantastic.

[Kristin] (54:46 - 54:51)
I had them too with childbirth. And yes, they're absolutely fantastic and highly recommended.

[Regina] (54:51 - 56:44)
So Laura, what was your strategy on finding people to be able to comment on this paper?

[Laura]
So the first thing I did was a literature search of what papers had come out. And then I just started contacting like the authors of pretty much all of the response papers.

And you can see in all of these papers, they explicitly say in the introduction, this study came out in Java Pediatrics in 2020. And so as a result, we are doing this study. It's a really direct, like explicit.

They were only doing it because this 2020 came out. So I just started reaching out to people. So that was my initial strategy for the research side.

I also did a whole, because I was trying to see if I could answer this question of had this information gotten out there? And if so, what effect did it have? I did a bunch of searching on social.

So I downloaded TikTok and tried to figure out how to use it. I, at the time, was more of an Instagram person. But I just searched like every combination of like autism, epidurals, things like that, that I could search for.

I did that on Instagram, Twitter. I reached out to some doulas and midwives who see patients and just kind of asking, like, has anyone asked you this question? And what was interesting was not a lot came up.

So it was a lot of social media sleuthing. But overall, there was very little there. And then a lot of the researchers that I talked to also are clinicians.

And they similarly said either no one had ever said anything to them or maybe one or two people. So I thought my story was going to be this study caused misinformation. And when that didn't happen, I didn't know what my narrative was or what the point of this story was.

I thought so many times that I had no story. I remember I filed a story with my editor and I was like, I think I just gave you 2,500 words of nothing. He's the main reason the story actually got published is because he kept telling me, he's like, no, there is a story here.

We just have to, like, write it the right way. At the end of the day, it became like a narrative of the scientific process. And how do you find, how do you get close to the truth?

[Kristin] (56:44 - 56:54)
So, Laura, this paper doesn't seem to have spread that widely. Spectrum covered the original 2020 study, but do you think it was not that widely covered elsewhere in the media?

[Laura] (56:55 - 57:57)
Yeah, there wasn't a ton of coverage about it in the wider press, like publications like the New York Times did not cover it, which actually was surprising to me because autism, unfortunately, is something that tends to get out into the larger press and epidurals and for a journalist who's not a science expert, it's like there's a clear sort of takeaway there to like at least explore. But one thing as we put in the story was this paper came out in October 2020. It was in the middle of a pandemic.

It was a month before one of the most contentious elections in US history. There was a lot going on. I can't definitively say that's why they didn't cover it or that's why it didn't get out there.

But I would bet that had like at least something to do with it. People were just really occupied with other things. And I mean, the research response was pretty quickly negative.

There was comments on the original paper. And so unless you had pre-reported it ahead of time, then you would probably quickly see like, oh, researchers don't really like this. So I'm not going to cover it.

But again, this is all speculation. I didn't ask the New York Times why they didn't cover it.

[Kristin] (57:57 - 58:01)
Well, that's interesting to think. It's good if journalists are using that kind of judgment.

[Laura] (58:01 - 59:26)
Yeah, I mean, we try.

[Regina]
So, Laura, when you spoke to researchers and clinicians, what sort of takeaway were they getting from this paper? Did they think, OK, this is causal, that they were going in and believing the result to be causal?

[Laura]
The short answer is no. No one thought it was causal that I talked to. To be fully transparent, most of the people I talked to were people who went out and did a study, like in response, which meant they were people who pretty strongly felt there needed to be more evidence.

But that means those were also the people who looked at the study really closely and looked at their own data. And there wasn't anyone I talked to who at all was concerned or thought there might be any real... I mean, no one can say there's no association, but if there was a real association, it's very, very, very small.

And even there was a couple of papers that came out that found small associations, and those papers say they think it's confounding. I also wanted to point out the original 2020 paper never says epidurals cause autism. And they're pretty careful to not say that.

And that was something that my sources pointed out. But they were like, anyone who reads this is going to take that away. You don't put out a study that says we found a correlation between autism and epidurals if you don't think there is something there.

If you think your association is completely due to nothing, you wouldn't put it out there. But I think it is important to note that these researchers weren't going around saying that epidurals cause autism.

[Kristin] (59:27 - 59:31)
Yeah, I believe we said that they put the obligatory cautionary line in there. Right.

[Laura] (59:32 - 59:51)
And I will note that the researchers of the study did not talk to me. They answered my email. I emailed them a whole bunch of detailed questions, and they basically responded with the same language that was in the conclusion of the paper.

So they didn't really tell me anything new, and they didn't answer any of my specific questions, and they did not want to talk.

[Kristin] (59:51 - 1:00:00)
I wonder if they got a lot of flack when it was initially published and they decided to not interact with the media because they were getting a lot of pushback from the scientific community.

[Laura] (1:00:01 - 1:00:07)
But I do think they published other similar studies on other topics after this. So if the blowback bothered them, it doesn't seem to have like.

[Kristin] (1:00:08 - 1:00:26)
I think they even published on autism and epidurals again. There was a study that did find some association. So I think they were.

So, Laura, the researchers you spoke with were skeptical of this 2020 paper from the start. Even though it did find a statistical association, what were some of the reasons for their skepticism?

[Laura] (1:00:26 - 1:02:45)
So one thing to point out is that statistics are not the whole story. They're super important and they give you so much information, but like they can't tell you everything. So one of the questions we're always taught to ask as science journalists is what is the mechanism?

Like how how would an epidural potentially cause or increase the likelihood of autism? And this was an issue that came up in the reporting a lot because most of the scientific community, we mostly believe that autism is largely genetic, perhaps gene environment interactions. But brain development happens really early on in pregnancy.

Right. Like in your first trimester, you're building something crazy like a hundred thousand neurons a minute. That's when brain development is happening.

And, you know, if you look at like biological studies of what genes linked to autism do in the brain, it's stuff super early about how neurons form and how the layers in the cortex form. And so one question that came up is like if those processes are related to autism, how does something that happens over the course of like a few hours at the end of a pregnancy contribute? Which, again, is not to say that it doesn't, but you have to ask how could that be happening?

And it gets to this question of why do you do a study? One of my sources said just because you have data is not a good reason to do a study. So before this 2020 study came out, scientists had never wondered if epidurals cause autism.

There had never been like clinical reports of this being a concern. There weren't data of like babies whose mom got epidurals, showed developmental delays a few months later. None of that was there.

And that would be a reason why you'd go, oh, there might be something going on here. Let's go look at our data and find out. I think that's a really important part when you ask this question of like, oh, they found this little association.

Is it real? It's important to bring in all this other non-statistical evidence and context of, again, why were you even looking for this in the first place? I point this out in the story, but they don't like cite real biological evidence in the paper either, which is something that sources pointed out to me you would normally expect in a study like this.

The only real study they cite is a 1998 study on 11 baby rhesus monkeys. And people questioned if the dose those monkeys got was even relevant for the kind of dose a human baby would be exposed to through their parent. So I just think it's really important to think about this isn't happening in a vacuum.

There is real biology involved and you can't just pull associations out of nowhere.

[Kristin] (1:02:46 - 1:03:04)
Regina, I think Laura just backed up my hot take from earlier that sometimes people do retrospective cohort studies just because the data is there and it's an easy publication. Let's throw some data on the wall, find some association and publish it. I'm not saying that's what the researchers here did, but that's always one of my concerns with retrospective cohort studies.

[Regina] (1:03:04 - 1:04:48)
So, Laura, what you're talking about here, these are sophisticated statistical ideas. And you're writing for an audience, yes, of researchers, but not all of them are statisticians or love statistics. So what was your thought process in how you're translating these ideas to be broadly accessible and understandable?

[Laura]
This is one of those times where my own lack of knowledge helps because I am not a statistician. I do not have a Ph.D. I am an English major. I know about science from doing science journalism and talking to a lot of scientists all the time.

And so I think writing at a level that I understood it using metaphors that I understood or using ways of explaining things that I understood was sufficient for our readers. And there's a lot of stuff I was like, oh, man, I wish I had listened to your podcast episode about my story before I wrote the story. But I wasn't doing a deep dive on the statistics.

I was telling the story of how the research community responded to this really interesting paper in an interesting way. Like this doesn't happen very often that someone publishes a paper and then a bunch of researchers are like, I want to publish a paper, too. And then they rush to put it.

That just doesn't happen very often at all. So the statistics were necessary to that story, but not at a level where I really needed to like deep dive into, for example, what a Cox regression is or a hazard ratio or that sort of thing.

[Regina]
So how would you summarize this in one sentence?

What would be your your nut graph here?

[Laura]
Well, a nut graph is usually like three sentences. One sentence is a challenge, let me think.

My summary was that scientists saw a result they weren't sure about, and they did a lot of work to try to make sure the correct information was put on the record. Yeah. And so we said a couple of times, this is a story of the scientific process working as it should.

[Kristin] (1:04:49 - 1:04:59)
This has been a really great conversation, Laura. And we always wrap up on Normal Curves by doing our smooch rating scale and our methodologic morals. And we're hoping that you'll join in on that today.

[Laura] (1:04:59 - 1:05:17)
I will do my best.

[Regina]
So our smooch rating scale is highly unscientific, but that's OK. We do it anyway.

One to five smooches, where one smooch means little to no evidence supporting the claim, five means a lot of evidence. So, Kristin, do you want to start off for us?

[Kristin] (1:05:17 - 1:05:46)
Yeah, Regina, this one's pretty easy for me. I'm going one smooch, our lowest rating. I'm not throwing a martini in the face, but I don't see any actual evidence here that epidurals cause autism.

There's that little statistically significant nine to 12 percent increase in relative risk that I think is probably just due to confounding. And as I've said, the sibling analyses are really what convinced me that that little effect that's left over is just confounding. So I'm going one smooch.

[Regina] (1:05:47 - 1:06:05)
Me, I'm going one smooch, too, exactly for the reasons that you said. Sibling analysis does not support this and it's a very small effect. And there's a confounding that we just cannot get rid of.

One smooch for me. Laura?

[Laura]
I think I agree.

It's funny. Part of me wants to say like 1.2 smooches. That's just my journalistic...

[Kristin] (1:06:05 - 1:06:08)
You can. That's allowed on this podcast.

Go for it.

[Laura] (1:06:09 - 1:06:41)
Well, I think if I was being honest, I would say one. But there's always that journalist skepticism that wants to leave a little bit of flexibility for like, maybe there's something I miss or maybe it's one of those examples where nobody believed them, but they were right. And I don't think that there's no evidence that that's true.

But, you know, so maybe we'll give them like a little bit of flexibility and say 1.2, but also because no one that I talk to thinks anything is there. And for all of the reasons I gave about the biology also just to me really says there's not really anything. There's no there there.

[Kristin] (1:06:41 - 1:06:43)
What does 0.2 of a smooch look like, Regina?

[Regina] (1:06:43 - 1:06:54)
I don't know. I think that would be maybe an air kiss.

[Kristin]
The smooch emoji maybe?

[Regina]
Yeah. The smooch emoji. Oh, even better.

[Laura]
I think it's an air kiss that doesn't land, goes in the wrong direction.

[Kristin] (1:06:54 - 1:06:56)
An air kiss in the wrong direction. Perfect. I love it.

[Regina] (1:06:57 - 1:07:33)
And we also do methodological morals where we try to pull out something from the episode that's not just about this particular claim or the subject matter, but what are the bigger lessons we can learn about scientific evidence or the methodological process or statistics? And a little like Aesop's fables, maybe, or kind of fortune cookie. And maybe Kristin and I will start with ours and that might inspire something in you.

How about that? Kristin, go ahead. What do you have?

[Kristin] (1:07:33 - 1:07:52)
I'm going with, Regina, every time you adjust the model and the effect gets smaller, that's the universe whispering, maybe don't build a causal story out of this. So, I'm thinking of the 1.25 to 1.15 to 1.08, you can see it shrinking. That's a good clue that we've got a lot of confounding going on.

What about you, Regina?

[Regina] (1:07:53 - 1:08:12)
I think I'm going to go with consistency across studies is gold. And that was getting to the idea that we had so many different studies with different populations and different study designs, but we were seeing consistent effect sizes in there. And that is really telling us something that's giving us confidence in the results.

[Kristin] (1:08:12 - 1:08:13)
I love that, Regina.

[Regina] (1:08:13 - 1:08:32)
Laura, do you find any inspiration in these? Anything you want to add?

[Laura]
Okay.

I think my moral is that there's more to the story than the statistics.

[Kristin]
I love it.

[Laura]
Yeah, the story always matters to the science and you need to think about the context in which your statistics appear.

[Regina]
As a statistician, I still love that.

[Laura]
I'm so glad.

[Kristin] (1:08:32 - 1:08:40)
We're not offended that everything is not statistics because that is what we try to do on Normal Curves is to put things in context and it's not just about the numbers. I love it.

[Laura] (1:08:40 - 1:08:45)
Yeah. If you don't have numbers, you're a bit lost, but they're also not enough on their own. Yes.

[Kristin] (1:08:45 - 1:08:45)
Love it.

[Regina] (1:08:46 - 1:08:55)
This has been so much fun, Laura.

[Laura]
Yeah. Thanks.

I never get to just spend an hour talking about a story that I worked on for months. So thank you.

[Kristin] (1:08:55 - 1:09:02)
Thanks so much, Laura, for being our first non-drunk guest on the podcast. Hopefully the first of many. And thanks, Regina.

This was a lot of fun.

[Regina] (1:09:03 - 1:09:05)
Thanks, everyone, for listening.

[Laura]
Thank you so much for having me.