When to get Married
Please consider taking a survey designed to help improve my newsletters. It's anonymous and consists of 5 quick questions about your Statisfied? experience. Just click here anytime.
According to one online wedding planner, "we're about to be in the thick of proposal season. By that, we mean a whopping 37 percent of engagements happen from November to February!"
I love that "whopping"! November through February is almost exactly one third of a year, so if engagements were randomly distributed over time, we would expect about 33% of them to occur during this time period. 37% doesn't seem "whoppingly" greater than 33%.
To be fair, studies do show a spike in engagements between Thanksgiving and Valentine's Day, and, in any case, the purpose of this newsletter isn't to pick on wedding planners. Rather, I want to discuss a study, released this week, which explores when it's "best" to get married.
I'll be using this study to illustrate a simple point about statistics: They're never objective. Statistics always reflect ideological assumptions. Assumptions about what is and isn't worth studying. Assumptions about what kinds of data are and aren't relevant. Assumptions about how to interpret findings. As you'll see, the study I focus on is more ideologically-driven than most.
Overview
The study, conducted University of Virginia researchers, was reported by several news organizations throughout the week. The researchers themselves described it in a Wall Street Journal article published on Saturday. Their starting point was evidence from prior studies that getting married in one's mid-20s or earlier leads to a higher divorce rate than marriage later in life (presumably because age brings greater maturity about one's choice of partners as well as how to interact with them). In their own study, the Virginia researchers found an exception to this pattern: People under 30 who married without previously living with someone (other than their future spouse) had the lowest divorce rate, even if they married in their teens or early 20s. In other words, early cohabitation predicted a higher incidence of divorce.
Why would cohabitation increase the chances of divorce later on? The researchers speculated that relationship baggage plays a role. They also suggested that people might compare their spouses to whoever they'd previously cohabited with and discover shortcomings.
I found these explanations odd. Even if "relationship baggage" made people more prone to divorce, why would cohabitation create more "baggage" than other close but non-cohabitative relationships that didn't work out? Likewise, why would you be more likely to compare your spouse to an ex-partner if you happened to have lived with your ex? (And, why would that comparison tend to make your spouse look bad?) I turned to the original study for clarification.
Study methods
The study looked at the whether age at time of marriage, cohabitation before marriage, and religious upbringing would predict marital outcomes (staying married vs. getting a divorce).
The sample consisted of over 53,000 American women, ages 15 through 49, who participated in the National Survey of Family Growth (NSFG) between 1995 and 2019. (Since 1973, the National Center for Health Statistics, a branch of the CDC, has run the NSFG in order to plan health services and health education programs, and to provide a database for studies related to families, fertility, and health.)
So far, so good. The sample is large, and the data are “objective” (in the sense that the surveys were designed by a federal agency rather than researchers with some theoretical agenda). Later though we'll see that both the sample and the data are problematic.
Study findings
Most of the key findings are presented in the figure below. (Some people like this kind of graphic; some don't. After unpacking the figure, I'll describe the results in both visual and non-visual terms.)
The y-axis in this figure is an estimate of the annual probability of divorce. For example, for any given year, the chances of divorce are about 3.8% for a woman who got married between the ages of 25 and 29.
On the x-axis are four age groups (Under 20, 20 to 24, 25 to 29, and 30 or older). These are the ages of the women when they got married. For each age group there are four bars. From left to right, the first bar is Nonreligious, direct marriage ("direct" means no cohabitation prior to marriage). The second bar is Religious, direct marriage ("Religious" refers specifically to religious upbringing). The third bar is Nonreligious, other marriage ("other" means cohabitation prior to marriage). The fourth bar is Religious, other marriage.
Here are three of the main findings:
1. Divorce rates decline as age at time of marriage increases (prior to age 30).
For women under 30, the chances of divorce decrease slightly from the youngest group (Under 20) to the oldest (25 to 29), The divorce rate then rises slightly for the 30 and older group. Visually speaking, the bars in the figure tend to get lower for each age group as you scan from left to right, until you reach the final group.
2. Religious and nonreligious women do not differ in divorce rates.
Within each age group, there were no significant differences between religious and nonreligious women in divorce rates. (Here's how to "see" this in the figure. First, estimate the midpoint between the heights of the first and third bar for each age group. That's the divorce rate for nonreligious women. Next, estimate the midpoint between the heights of the second and fourth bar for each age group. That's the divorce rate for religious women. You can then "see" that those midpoints are about the same.)
3. Women who never cohabitated have lower divorce rates (prior to age 30).
Within each of the first three age groups, divorce rates were significantly lower for women who had no cohabitation prior to marriage. For women 30 or older when they married, divorce rates are lower if they did live with someone prior to marriage. In the figure, you can see that for each of the first three age groups, the first two bars (women with no prior cohabitation) are always lower than the second two bars (women who had cohabited). This pattern is reversed for the 30 and older group.
In short, according to the researchers, cohabitation when you're under 30 is undesirable, because it increases your chances of divorce. As they put it in their Wall Street Journal piece: "If you’re a young woman thinking about getting married but worried about divorce, our research suggests that you need not wait until you’re 30—so long as you’ve found a good partner and don’t move in with anyone until after your wedding day."
Hmmm....
The ideology behind the study
I found a number of statistical and interpretive problems with this study, all of them traceable to a single source: The study was created to serve an ideological agenda. In other words, it wasn't intended to reveal something about marital relationships. Rather, the point was to find data that fit pre-existing beliefs.
The researchers who conducted the study are part of University of Virginia's Institute for Family Studies (IFS), which claims to support "objective" research "dedicated to strengthening marriage and family life." In fact, the sole purpose of the IFS is to promote a traditional, relatively narrow view of family relationships. Specifically, IFS research, policy reports, blog posts, etc. push the agenda that people should get married and have children. They should marry once, without prior cohabitation, not too late in life, and raise children together. They should not get divorced. The marriages should be heterosexual. And, people who aren't married should not raise children. The IFS is quite serious about this agenda. As their website puts it, "the fact that roughly one in two children in America grow up outside of an intact, married family constitutes one of the most significant threats to America’s future stability and prosperity."
You could hold views like this and still conduct excellent research. However, in this particular case, the researchers' ideological biases undermined the study design, methods and statistical analyses. Here are some of the main problems:
Ideologically-driven research questions
The researchers looked at a single outcome variable: Whether married women stayed married or got divorced. The assumption was that staying married is always best, while divorce is inherently undesirable. Not all of us share this perspective. (Many of us would say, for example, that a woman whose husband abuses her is better off divorced.) In short, the study is limited by its simplistic contrast between a good outcome (staying married) versus a bad one (getting divorced).
It's telling that the researchers only examined cohabitations that preceded marriage. They should’ve also looked at outcomes for couples who choose to cohabitate as an alternative to marriage, an increasing trend in the U.S. Whether such people have a legal relationship (e.g., a domestic partnership) or just a personal agreement, they may be as committed as married couples to a lifelong connection. So, I would argue that the outcomes for these people – i.e., whether they stay together or split up – are just as important as the outcomes for married couples. The Virginia researchers chose to ignore them, owing to the ideological bias that only marriage counts.
Non-representative sampling
The sample was huge, but it's far from representative. Only women were surveyed. The focus was on heterosexual marriages. And, again, the researchers only considered outcomes for married women rather than those in committed partnerships.
Suspicious variable selection
The NSFG survey provides information on over a hundred variables, including many that are known to influence peoples' decisions about whether or not to get divorced (e.g., quality of relationship with spouse, number of children at home, financial independence, sexual attitudes and preferences, etc.) In spite of that wealth of data, the researchers focused on only three predictors of divorce (age at marriage, cohabitation, and religious upbringing). Why only these three? Presumably because they're the ones that allowed the researchers to tell the story they wished to tell.
If I seem overly suspicious, consider the researchers' own justification for why they chose to focus on religious upbringing rather than current religious affiliation:
"Current religious affiliation is not a very informative variable for understanding how religion influences family life because, for example, marriage might motivate people to become more religious (or cohabitation might motivate people to become less religious). But religious upbringing…occurs before the vast majority of marriages or cohabitations, so is not influenced by them.
I find the logic here pretty weak. Even if marriage influences one’s religious beliefs, such beliefs still have a discernible impact on how one views marriage (and whether one decides to stay married or not). Knowing a person’s current religious affiliation (e.g., Catholicism) should thus be informative. On the other hand, just because a person was raised in a certain faith doesn't mean that their decisions as an adult are affected by that faith. If you knew someone had been raised Catholic, that wouldn't necessarily tell you much about why they got divorced.
Ideally, this study would’ve considered both religious upbringing and current religious affiliation, as both are informative at least some of the time. Again, I suspect the researchers chose religious upbringing because it was the variable that best enabled their agenda.
Exaggerated conclusions
In the study itself, and in their Wall Street Journal summary, the researchers say a lot about risk of divorce. For example: "Especially for religious men and women who avoid cohabitation, our analysis of the NSFG indicates that they can marry in their 20s without serious adverse divorce risks."
What's problematic about statements like this (and there are many of them) is that differences in risk of divorce, both within and across age groups, are just a few percentage points at most. There are no "serious adverse divorce risks" in this study. Or, if there are (since what constitutes a "serious" risk is a matter of opinion), the risks are nearly the same across the board. If you look back at the figure, you can see that the largest difference between any pair of bars is only about 3 percentage points.
Large sample size
You're probably thinking: Why call this a problem? Aren't large samples a good thing? Generally, yes. But huge samples must be treated cautiously, because they make it inordinately easy to find statistically significant effects. In the Appendix I discuss the causes and implications of this phenomenon.
For three of the four age groups in the Virginia study, women who cohabited prior to marriage had higher divorce rates than women who did not. However, the difference in chance of divorce was less than about 1.5% for any age group. That's a tiny percentage, but statistically significant thanks to a total sample of over 53,000 women.
What's hidden here are the many other variables known to influence the probability of divorce. Quality of relationship with spouse, financial independence, number of children at home, etc. It's a long list. If the researchers had considered these variables, the effects of cohabitation would've probably disappeared. (Stats people: Logistic regression techniques would be suitable here.) Even if those analyses weren't run, researchers who were less ideologically driven wouldn't have made much of the cohabitation effect. They might’ve noted it, but they'd also acknowledge that it's readily explained by something other than cohabitation. I'll close with one example of what that "something" might be:
The researchers labeled each woman in the study as either having or not having cohabitated prior to marriage. They didn't measure the number of prior cohabitations. Clearly, they should have done so. Although it seems implausible that living with one person prior to marriage could increase the risk of divorce, if you've lived with, and then broken up with, five different people prior to marriage, you might be more likely than most people to end up divorced, owing to whatever it was that caused the break-ups with those five people you'd previously lived with (i.e., your difficult personality, your desire for independence, or whatever). Thus, in the Virginia study, the tiny difference in divorce rates between the cohabitation and no-cohabitation groups could simply be attributable to a higher divorce rate among a subset of the former group (i.e., women with many cohabitations). If that's the case, we can't conclude that cohabitation before marriage increases the risk of divorce. Rather, we would infer that multiple cohabitations before marriage reflect some underlying characteristic that increases the risk of divorce.
Conclusion
Although this study was conducted by researchers at a prestigious R1 institution (University of Virginia is currently ranked #25 among national universities) and reported in a prominent news outlet (The Wall Street Journal boasts a circulation of over 2.8 million and 37 Pulitzer Prizes), in the end, the study reveals almost nothing about marital relationships, because it was carried out merely to advance an ideological agenda. In doing so, the study takes its place among other studies, policy papers, educational materials, etc. on the University of Virginia IFS website, all of which label as undesirable practices such as cohabiting without marriage, being a single parent, getting divorced, and participating in any form of non-heterosexual marriage and/or parenting.
There are lots of studies on predictors of family well-being. The data are voluminous, complex, and not always consistent, but one thing is clear: No single demographic category is automatically desirable or undesirable. For instance, a long line of research (and casual observation) tells us that couples tend to stay together when they listen to each other, treat each other with affection and respect, negotiate conflict thoughtfully, and so on. These variables predict relationship success, regardless of the couples' prior dating history, demographics, and so on. In other words, the success of your marriage doesn't depend on whether or not you lived with someone else prior to marriage.
If your religion, your code of ethics, and/or your personal taste tells you not to cohabitate before marriage, fine, there's no compelling reason that you should. You can "practice" living with someone without sharing a lease. But if you do choose to cohabitate, you shouldn't worry that it will increase your chances of divorce. All it will do is to aggravate scholars who wish all women would be June Cleaver.
Happy Valentines Day!
Appendix: Large samples and statistical hypersensitivity
Here I’ll explain, using an informal example, why significance testing can be risky when sample size is unusually large.
Imagine that we administer the same IQ test to two groups of adults who differ in some way that's unrelated to intelligence. For example, one group consists of 50 adults born on a Tuesday, while the other group consists of 50 adults born on a Thursday. We'll almost surely find that the groups differ slightly in mean IQ, but that the difference isn't significant. For example, the Tuesday group mean might be 101.3, while the Thursday group mean might be 100.9. Now, if we sample a different group of 50 Tuesday adults and 50 Thursday adults, we'll almost surely get different means – maybe a Tuesday group mean of 101.6, and a Thursday group mean of 100.8 – but the difference probably still won't be significant. However, if we increased the sample size for each group to 20,000, the difference between 101.6 and 100.8 would be significantly different. In fact, with such large samples, even a difference of one tenth of an IQ point (e.g., 101.6 vs. 101.7) would probably turn out significant. And yet, (a) one tenth of an IQ point isn't cognitively meaningful, and (b) this tiny difference must be due to something other than being born on different days (e.g., chance).
In short, enormous samples create statistical hypersensitivity. If you go fishing in the enormous datasets these samples yield, you'll eventually find significant IQ differences between people born on Tuesday vs. Thursday, or between people whose favorite color is red vs. blue, and so on. (The preferred alternative to fishing is to only run analyses justified by prior research and/or theory.)
Because the Virginia researchers had a sample size of over 53,000 to work with, as well as a dataset consisting of more than a hundred variables, it was almost guaranteed that they’d be able to find significant effects consistent with their ideological agenda. I suspect that some of the quirks of their study (e.g., ignoring known predictors of divorce; focusing on religious upbringing rather than current religious affiliation) can be attributed to an initial fishing expedition that identified the desired patterns of significant findings.