Looks like my hypothesis is supported by the evidence: On page 4, you can see their regression results. In fact, they don't estimate the statistical significance of any occupation as a factor, not even the law enforcement occupation that they focused on.
The "media and communications equipment workers" category had a reported 0.00% divorce rate! It seems implausible that we should infer causality from this.
There are a few other studies on this. Maybe I'm reading these incorrectly so would appreciate your analysis. As far as I can interpret, the pattern is legit. What is especially counter-intuitive (unless you have a negative view of the profession, as I do) is that so-called professionals in mental health have a higher divorce rate even than respondents who score high on the "anger scale". So what's that all about?
> Survey results from 2008 through 2013 included responses from more than 40,000 physicians; 200,000 other health professionals—dentists, pharmacists, nurses and health care executives; and more than 6 million other adults who reported currently being employed and ever being married. While 24 percent of physician respondents had ever been divorced, the probability of being divorced was 25 percent among dentists, 31 percent among health care executives and 33 percent among nurses. Only pharmacists, at 23 percent, were less likely than physicians to have been divorced. Lawyers had a 27 percent probability of being divorced, and *in all non-health-care occupations, the probability of ever being divorced was 35 percent.*
> A subsequent, larger study of 1118 medical graduates of Johns Hopkins University found cumulative rates of divorce of 29% *with rates higher among psychiatrists (50%)* and surgeons (33%) but was limited by its analysis of physicians from a single institution.
> The choice of specialty was significantly associated (P>0.001) with the risk of divorce (Fig. 1 and Table 2). The cumulative incidence of divorce was highest
for *psychiatrists (50 percent)*, followed by surgeons (33 percent) and “other” physicians (31 percent). Among internists, pediatricians, and pathologists, the incidence of divorce was similar (22 to 24 per-
cent).
> When we examined psychological variables, we found that physicians in the highest quartile for the anger scale had a higher risk of divorce than those
scoring in the lower three quartiles. *Moreover, the cumulative incidence of divorce among the physicians with the highest anger scores was higher than
that of any other subgroup, with the exception of those practicing psychiatry.*
Looking at the "Medical Specialty and the Incidence of Divorce" article, my first thought on reading "Scores on these [psychological characteristic] scales were grouped according to quartiles for analysis" is that the researchers aren't experts in statistical methods, or else are hiding something by bucketing arbitrarily. It's a widespread problem in life sciences. They don't explain (not by this point in the article, at least) why they chose quartiles. I'm skeptical of any analysis that reduces the fidelity of the data without explanation. It smells like p-hacking.
My second thought regarding their model is wondering why they didn't first build a model predicting divorce for the general population. I've heard (received wisdom) that financial stress is the primary cause. It seems unlikely that medical specialty is unrelated to household finances. Ignoring a likely confounding variable is sloppy. It makes the effort seem like a novelty article intended for amusement rather than serious research. It's in the "occasional notes" section of the periodical, which suggests less rigor.
They do little to address the question of causality. They muse that, "One explanation is the longer work hours required in some specialties," but never bothered to analyze the hours worked.
My conclusion: That article is literally a joke, published to give the community something to laugh about over beers after the conference.
----
If I were to perform the analysis, I'd first build a model of causes of divorce for the general population. Some candidate variables: household income, household income as a ratio to each person's childhood household income, cost of living where the household lives, frequency of attending religious services, number of children, number of hours worked per week, age when married, etc.
I'm skeptical of psychological characteristic surveys, so I'd spend some time considering alternatives, but I'd want to include some measure of characteristics that might drive the choice of a medical specialty which might also be correlated with divorce. We want to isolate the choice of specialty as the cause as separate from confounders that cause both the choice and divorce.
If medical specialty holds up as a factor after controlling for everything else, then I'd investigate further.
Thanks. My first reaction is that this one is much better. Larger sample, better methods.
Again, I hate the arbitrary bucketing. There's no benefit to creating age buckets, as if there's some magical change that happens at 40, 50, and 60. I did have an atypically large birthday party, compared to other years, but I don't think that had a dramatic effect on my likelihood to divorce. Maybe some people have such spectacularly intense decade-related birthday parties that it increases divorce incidence for those years? Income doesn't benefit from bucketing, either. Having $199k annual income isn't much different from $201k annual income.
They do a decent job of controlling for confounders. Considering, among other things, that "if the annual rate of divorce was identical across occupations but physicians marry later in life, then at any given time physicians would be less likely to report ever having divorced compared with people in other occupations, simply because they were at risk for less time."
However, their inclusions of state and year fixed effects could have been more considered. These are proxies for other things, like cultural characteristics and neighborhood income levels. I'd like to see some discussion of why they chose to use state and year proxies instead of searching for more specific explanatory variables. Especially state, because some kind of urbanization measure might be more helpful. However, because state medical licensing regulation might affect the choice of occupation, the use of that variable is easily defensible.
Ugh! Bucketing again! Hours worked should be a continuous variable. Inexcusable, unless they feared misreporting. Perhaps they saw some banding at 40, 45, and 50 hours, so they figured it's not really a continuous variable anyway. Again, that needs explanation. Any rationale for bucketing should be thoroughly discussed.
Finally, the effects. First, with this population size, I'd be surprised if these weren't "statistically significant". I'm looking more for practical significance. Check out the estimates for dentists. Dentists appear to have lower incidence of divorce, based on the last year, yet higher prevalence of divorce. Strange. That means that dentists in past years had a higher annual incidence, but the rate has been declining, or declining relative to physicians. Has the practice of dentistry, relative to general medicine, changed that much over those years? This suggests spurious results, at least for the physician vs dentist comparison.
Hispanics were more likely to divorce? Bogus. I'll chalk it up to sampling weirdness. The CI includes 1 anyway, so they're saying it doesn't matter. They should put some asterisks in to highlight the variables we should pay attention to. It looks like they're including Black and Hispanic just to explain what Other means, because Other is the only significant one.
Wow! Income is irrelevant. Weird again. Maybe bucketing at work, turning 1 continuous variable into 4 binary variables, diluting the impact. It should have been log(dollars).
If you work more than 60 hours, you're more likely to get divorced. Makes sense. Again, log(hours) would have been better, though maybe a threshold at 40 hours would have been useful.
I enjoyed the article, but if I were the journal editor, I'd have returned it with some suggestions for improvement rather than publishing it.
Anyway, I hope my commentary was interesting/useful. This article doesn't say anything about psychology, so we've gone off on a bit of a tangent.
Yes, very interesting, thank you for putting in the time. I appreciate the detailed analysis. Were you able to make any inference at all about psychiatry and divorce from these data or no? It's interesting that you are using some background knowledge to evaluate the findings (e.g., hispanic divorce). I'm curious where that comes from and how it fits into our discussion. Is it that you have data about hispanic background divorce rates or is it because you know hispanics are largely Catholic and making a logical inference? As far as dentistry is concerned, it may be a hidden variable, such as makeup of those practicing dentistry. Maybe a change in composition of male/female overall. Or from different cultural or ethnic backgrounds with different background rates of divorce. My overall impression of your analytical tools is that you are willing to hypothesize causes with the caveat that they be subject to further investigation. If you reach such places in analysis and stop, do you just reserve judgment from that point? I would think not. Rather, your priors change and your probabilities change so you can go about life without perfectly constructed and complete statistical evidence, as all of us must.
To return to the topic under discussion, it sounds like you are saying, "there may or may not be a correlation between psychiatry and divorce, but these particular studies can't provide the answer." I assure you I am not basing my opinions about psychiatrists on these studies. Rather, I expect that properly constructed studies that meet your standards would bear out what I know from my own experience and encounters with people in the profession. Others in the thread provided anecdotal data that supports my own. And, I'm not making my judgment based solely on experience. My experience confirms an intellectual analysis based on the history of the profession. Those are my priors and probabilities. I would not be surprised if the data backed me up but I would be surprised if the data refuted my suppositions (and would question the study). It seems that my sort of reasoning doesn't have much place in your toolbox. Is that not the case?
I appreciate your reply and explanation. However, I think we'll probably be speaking at cross-purposes because your description of the kind of variables you would choose for your own model strike me as (necessarily) limited and reflective of the data requirements of your preferred methods (McNamara fallacy?). (I don't actually believe you live your life via models generally, but maybe you do.)
My opinion is there is no room in such models for all the many things that are part of the rich fabric of psychological experience, without which all you get is a kind of significant/not-significant binary according to available imperfect data. I mean, yes, it preserves the null hypothesis, but it feels sterile to me to attribute so much to randomness when the model itself is so obviously curated to work with available data. Or even to hidden variables that will require further study, at some point, in the future, maybe... Not to mention the cases where the null hypothesis was subsequently rejected (smoking, ulcers,...). Yes, Maybe you believe these variables are the principal components of the theoretical complete data? The don't appear orthogonal to me.
Meanwhile, life has to be lived and if you've encountered "types" of people in the world but won't allow yourself to acknowledge their existence unless one can build a rigorous predictive model to verify the existence of those types and to be sure they are not noise or sample bias or whatever, then you are enjoying a life that I would find almost barren. There are so many locked doors in this way of seeing the world. You have problems with datasets, which you have to decide whether to trust. Will you apply statistics or heuristics to those problems? Absent trustworthy data will you just decide to defer judgment? It's a form of not trusting oneself as well, which I reject on principle, and of making oneself unbiased to the point of being inconsequential. In other words, a form of nihilism (perhaps nullism is the appropriate term).
No, I don't think I can make any inferences about psychiatry from those articles. My Bayesian update is the opposite of yours: because the evidence presented was weak, I'm more skeptical of any relationship between psychiatry and divorce. I was quite ready to believe it, and was disappointed to see such weak studies. It suggests to me that it's difficult to find positive results with stronger analysis.
You seem to be suggesting I'm philosophically a Frequentist. Somewhat true, but I am, like basically everyone, a Bayesian when it comes to practical decisions. Also, I have no fear of logical deduction when statistical inference is infeasible.
Nullism is a good term. I'm of the opinion that most things are random and that humans' imaginations frequently mislead us. Absent better evidence, I'm unlikely to believe any link between psychiatry and divorce.
https://sci-hub.st/https://doi.org/10.1007/s11896-009-9057-8