The Fastest Retraction Ever

On December 26, 2023, someone posted a message on Reddit saying that one in five husbands leave their wives if their wife is diagnosed with a serious illness. I'm guessing they were referencing a 2009 study exploring whether MS and cancer diagnoses affected divorce rates, but they may have also been thinking about a larger 2015 study, which asked a similar question. Or, perhaps they weren't thinking about that 2015 study, which, as I'll explain later on, could be equally problematic.

The finding that significantly more men leave sick wives than the reverse can't be found in the 2015 paper. That's because it came from the first version of the paper. That version was published in March 2015 and made headlines because the difference in divorce rates among couples when one of them gets ill was striking. According to that version, a wife getting seriously ill leads to a 6% higher probability of a divorce, while a husband getting ill doesn't change the probability of a divorce (the study only looked at heterosexual married couples).

This seemed to substantiate the equally headline-making 2009 study and the idea that men, at least on average, are less loyal partners than women. The finding "made sense" because males have an advantage in the marriage market (at old age, there are more single women than men) and because of cultural factors (the relative value of youth for women in the marriage market, compared to men).

Then something unusual happened. Another lab drawing on the same kind of data couldn't replicate the 2015 study's numbers. Members of this lab contacted the authors of the study and the authors discovered there had been a typo that miscategorized a large group of data, leading to completely incorrect results. The authors quickly issued a retraction notice, corrected the mistake, re-did their data analysis, and released an updated version of their paper. In the updated version, the difference in divorce rates is only statistically significant for one category of illness (heart problems), but negligible overall.

Retraction Watch, a website that monitors retractions in the peer-reviewed literature, held up this retraction as one where almost everything went right: the mistake was caught quickly, the authors responded transparently, and, at the time of the retraction, no papers had cited the original paper. Great, right?

Well, kind of.

SagePub, who published the article, followed a "retract-and-republish" strategy in this case. The updated article has exactly the same name as the old one. And it doesn't mention that there was ever a retracted version of itself. Though the retraction notice appears as related content in a sidebar.

The PubMed version of the updated article, however, does inform readers at the end of the abstract that it was a republished form of a retracted article. And if you somehow come across the original article on PubMed, that does say it is a retracted version and sends you to the updated article. But the retraction is just an abstract. You can't read the retracted material. The only direct link that I know of to the retracted material comes from Retraction Watch.

This is fine for future researchers – maybe even an ideal state of affairs. If you're reading about this topic for the first time, you will come across the corrected version of the article and be unaware that there ever was a correction.

But for public discussion of the article, the retract-and-replace strategy seems incorrect. During the month that the original, incorrect article was up, many outlets wrote about the article's findings. Some of these outlets, revised their articles in some way. Huffpost added an update. Washington Post published a follow-up article. Others did not. But the brief window of time between publication and retraction was enough to shape the conversational landscape on this issue. And without some notice that the updated article is an updated version of a retraction, it's easy to get confused. Especially when the replacement has the exact same title, the same authors, and the same year of publication as the retracted article (the only difference is the month and issue).

A related problem comes from how people talk about the finding. Notice the difference between the titles of the research articles and the Reddit post: In Sickness and in Health? Physical Illness as a Risk Factor for Marital Dissolution in Later Life (2015) and Gender disparity in the rate of partner abandonment in patients with serious medical illness (2009) vs. More than one in five men leave their wives if they are diagnosed with a serious illness. In sickness and health?. The Reddit version may be a mash-up of the 2009 and 2015 studies or maybe just a reference to the 2009 study. It also fronts the conclusion. The Reddit headline also contains an added assumption that the data cannot support: that men were leaving their wives.

Data on who initiated divorce proceedings wasn't available in the 2015 study. The 2009 study uses the term "partner abandonment," but it's unclear to me if they knew who initiated divorce proceedings or why the marriage dissolved (just that it did). You could find some plausible reasons that would support either conclusion: it's likely easier for men to re-marry for the reasons stated earlier (a reason to support the idea that husbands leave their wives), and men, in general, are perceived as worse caretakers than women (a reason to support the idea that wives leave their husbands). Both of these possibilities (and others) are discussed in the 2015 study.

As the recent Reddit post suggests, however, online articles continue to act as though the 2015 article supports the general principle that the disparity between divorce rates is large, usually including both the 2009 and 2015 findings as if they are mutually supportive. The 2009 article has not been retracted, but it had a much smaller, more homogeneous sample (~500 compared to ~2500; a single hospital compared to nationally representative panel study), examined a smaller number of serious illnesses, examined a shorter time period, and used a different methodology.

But the 2009 paper and the 2015 article are in conflict. The 2015 paper didn't try to replicate the 2009 paper, but it asked the same basic question in a different way. A wife's onset of illness was associated with a 1% higher probability of divorce (than if the wife had not gotten ill). This was higher than the probability for men. Using standard statistical hypothesis testing, however, the authors couldn't reject a null hypothesis that the divorce rates were the same. This is wildly different than finding that ~21% of marriages where the wife contracts cancer end in divorce (vs. ~3% for husbands). Even the "6% increased probability of divorce" of the retracted version seems at odds with the 2009 paper, though it can be tricky to compare because they are evaluating slightly different things.

You can't just pick the study you like and choose to believe it. Presuming both findings are accurate, we should be in conclusion-limbo. Some evidence points one way, some evidence points another. Certainly, there's no evidence that the husband's illness drives divorce rates, but how much difference there is in divorce rates in cases where the wife becomes seriously ill compared to cases where the husband becomes seriously ill... it's hard to say. Based on the research in this area as a whole, it doesn't even seem like we can or should say, because the results depend on the patient cohort and the diseases that the researchers choose to examine. Not to mention that these rates will likely change over time, due to changing cultural and economic circumstances.

If anything, it seems like the more recent, more comprehensive study should be more persuasive. But that isn't the way public conversations about this issue have played out.

I think there are some lessons here.

First, important details about scientific research often get lost as members of the public take it up for discussion. What's usually preserved is "the gist" – in this case, that more divorces occur when the wife gets sick than the husband (regardless of whether that's 21% divorce rate vs 3% for couple where a spouse contracted cancer or a 6% vs 0% increased probability of divorce where a spouse contracts a serious illness). The gist is attached to other cultural ideas – that husbands are less faithful than wives – which creates a composite claim: husbands tend to leave their wives when they get sick (at least more so than wives leave their husbands). Of course, even taking the "one in five" statistic from the 2009 study at face value, most husbands (80%) would remain with their sick wives. But it's almost never framed that way.

What's ignored are methodological considerations that would influence how much weight we should put on various studies. Or details about the effect size or the practical importance of the findings. Or the considerable qualifications that researchers must make with any of these findings.

Second, the speed at which the media reports new findings creates an ongoing trail of misinformation. During a window of a couple of months, the retracted information became "fixed" on the internet. Search engines will continue to serve up articles that never corrected their information, post-retraction. And people will continue to read those articles. The Washington Post's original article, for instance, remains up without any correction or notice that it's not accurate. At least without further examination, we will continue to give claims a scientific imprimatur they don't deserve.

Third, in the modern era, quantitative data analysis is vulnerable to coding errors. It just is. Many errors are caught because they lead to absurd results. Errors that don't, however, may never be caught. Here, the original authors might have suspected something was odd because the overall divorce rate among these older couples was so high (>30% before the correction; 6% after). But that never triggered their alarm bells – there's a delicate balance between being open to the surprise of a new finding and being suspicious of things that don't make sense based on what you already know. No doubt improved research practices could reduce these kinds of errors. But it was just luck that another lab happened to be looking at the same data set and attempting to replicate the original authors' numbers. In most cases, this does not happen.

Fourth, transparency about the record of events is critical for correcting things later on. SagePub's decision to replace the retracted article – with nary a pop-up or a highlighted box or a simple text message noting the previous retraction – creates a gap in the record of what went on; a gap that caused considerable confusion for me, when I started looking into this situation, and a gap that, I think, contributes to continued misperception of the state of the evidence.

This whole episode is just one example of the breakdown between scientific research and public discussion. But it represents some of the complexities and pitfalls of the research-to-public-opinion pipeline, even when everything about the retraction process "goes right".

References

The original article is here: Karraker, A., & Latham, K. (2015). In sickness and in health? Physical illness as a risk factor for marital dissolution in later life. Journal of health and social behavior, 56(1), 59–73. https://doi.org/10.1177/0022146514568351 (Retraction published J Health Soc Behav 2015:doi/10.1177/0022146515595817)

The retraction notice is here: Karraker, A., & Latham, K. (2015). Authors’ Explanation of the Retraction. Journal of Health and Social Behavior, 56(3), 417-419. https://journals.sagepub.com/doi/full/10.1177/0022146515595817

The (fixed) article is here: Karraker, Amelia and Kenzie Latham. 2015. “In Sickness and in Health? Physical Illness as a Risk Factor for  Marital Dissolution in Later Life.” Journal of Health  and Social Behavior 56(1):59–73. Available here: https://journals.sagepub.com/doi/full/10.1177/0022146515596354 and

Retraction Watch articles from 2015 are here: https://retractionwatch.com/2015/07/21/to-our-horror-widely-reported-study-suggesting-divorce-is-more-likely-when-wives-fall-ill-gets-axed/ and https://retractionwatch.com/2015/09/10/divorce-study-felled-by-a-coding-error-gets-a-second-chance/.

The Reddit thread in question: https://www.reddit.com/r/TwoXChromosomes/comments/18s1zn1/more_than_one_in_five_men_leave_their_wives_if/

The 2009 study: Glantz, M. J., Chamberlain, M. C., Liu, Q., Hsieh, C. C., Edwards, K. R., Van Horn, A., & Recht, L. (2009). Gender disparity in the rate of partner abandonment in patients with serious medical illness. Cancer, 115(22), 5237-5242. https://acsjournals.onlinelibrary.wiley.com/doi/pdfdirect/10.1002/cncr.24577

Other research related to divorce rates after the onset of illness: Carlsen, K., Dalton, S. O., Frederiksen, K., Diderichsen, F., & Johansen, C. (2007). Are cancer survivors at an increased risk for divorce? A Danish cohort study. European Journal of Cancer, 43(14), 2093-2099. https://www.sciencedirect.com/science/article/pii/S0959804907004376

For another null finding in this area, see: Syse, A., & Kravdal, Ø. (2007). Does cancer affect the divorce rate?. Demographic Research, 16, 469-492. https://www.demographic-research.org/Volumes/Vol16/15/16-15.pdf