Can We Predict if A Given Couple Will Make it or Not?

By Scott M. Stanley

University of Denver

 For years now, various marital researchers (including our team at the University of Denver headed by my long time colleague Howard Markman) have reported high accuracy of prediction rates (up to 90% or even higher) for which couples will do well and which will fail, using only data from early in the relationships.  There is something very exciting about these findings.  They suggest that we have new understanding about the patterns that put marriages at risk, and this knowledge adds to a growing empirical basis upon which to build approaches for helping couples build strong and happy marriages.  However, there are serious limits about the meaning and use of “prediction” research. This report is intended to clarify the issue in greater depth. 

The title above illustrates the crucial question. "Can We Predict if A Given Couple Will Make it or Not?"  The short answer is "not really."  As far as I am aware, no one can confidently predict if any given couple will divorce or be happily (or unhappily) married—though such claims exist.  Further, it is widely considered unethical in the marital counseling professions to attempt to give couples such specific feedback about their future.  That is different, though, from being able to tell a couple whether or not they are substantially greater risk than other couples. There is plenty enough good science behind an understanding of risk factors for marital distress and divorce upon which one could determine that a particular couple is at risk. The difference is in the degree of certainty one attaches to the assessment.

When it comes to the matter of prediction, there are two significant issues.  First, the often quoted "high prediction rates" are based on samples of couples, not specific individual couples.  Second, they aren't actually "prediction" rates in the normal sense of the term.  In these studies, researchers are usually assessing the degree to which they can look backward (having collected data over years) and use statistical techniques to correctly classify the couples as to their eventual outcomes based on variables collected earlier in their relationships.  In other words, researchers know the outcomes for the couples, then, using only early variables, they study how well the various groups of couples (e.g., divorced or not divorced) can be differentiated using special statistical techniques.  In such studies (as well as others) it is common for researchers to use the word “prediction,” but researchers are not actually writing down their predictions for specific couples prior to following them over some number of years.  

As one example, in our data set from the 80s, we (Mari Clements, Howard Markman, and colleagues) found we can correctly classify with about 90% accuracy who ended up divorced— and who did not— based on the premarital data for a sample of about 100 couples who went on to marry.  John Gottman and colleagues have reported a number of studies where they achieved high "predictive" rates (e.g., Gottman, & Krokoff, 1989; Gottman, Coan, Carrere, & Swanson, 998).  Matthews, Wickrama, & Conger (1996) and others have also done so.  Usually, the highest rates of classification "accuracy" are found in studies which employ objective coding of couple interaction—meaning studies where trained coders look in depth at how the couples interact.  But, in none of these studies do the researchers predict which couples will do well and which will not ahead of time.  At least not as far as I know. 

Do not get me wrong. I think these kinds of studies yield impressive information.  Researchers can argue about such things, but that is the view I and many others take.  However, one negative side effect of the attention these studies have received is that they have given the public the idea that we can predict marital outcome—happiness or divorce—accurately for an individual couple.  A true prediction rate would come from studies using methods that ask the researchers to predict, early on, which couples will make it, writing their predictions down, and following up years later to see how well the predictions hold up.  This is not what these studies do.  These kinds of studies come up with ratings of how accurate we can correctly "classify" couples based on data earlier in their relationships, and these accuracy ratings are for the entire sample, not a given couple.  Further, these kinds of studies capitalize to some degree on characteristics in given samples.  It is unlikely one could get the same high rate of  accuracy in another sample of couples using the exact same set of variables used in an earlier study.  This is somewhat like getting a tux to fit one person perfectly well.  The tux is specially “fit” to that person.  It just will not fit another person in the same manner.  Likewise, the equations of variables that are used to statistically differentiate couples in these kinds of studies become tailored to the samples of those studies.  They will not fit just the same with another sample, and it is possible they would fit quite poorly. 

Bottom line? We should NOT feel confident enough about prediction to look at one couple and say "you're not going to make it."  On the other hand, it might be wise to tell a couple something like the following: 

“You have many risk factors associated with marriages that don't make it.  You might want to think seriously about what that means for the two of you and take measures to lower these risks—work on changing these behaviors There are approaches geared to trying to lower the kinds of risks you have and to trying to help couples strengthen factors that are associated with successful marriages." 

Beyond that, the ice is thin for skating—and the water darn cold if you fall in.

The Technicality of Base Rates

        For those of you who want to go further in understanding this, I would like to explain how what are called "base rates" affect the kinds of studies that both identify marital risks and assess accuracy of classification based on those risks. 

Let's say (this is close enough for the point being made here) that in a premarital samples, of those who marry, 25% or so are going to divorce in 8 years.  So, let's say a researcher comes to you and says, "I can accurately predict, with 75% accuracy, which couples will divorce and which will stay together."  You think, "gee, that's pretty impressive."  In fact, the researcher further impresses you by saying this: "Yep, and I can do it just looking down the list of names."  "Wow, that's fabulous" you think.  Then you spend some time with the researcher (undoubtedly having a great time, for they are a wonderful kind of human being) and ask to be shown just how one can do this without even looking at substantive data about the couples.  The researcher says to you, "easy, I just predict that all of the couples in the sample will stay together." 

What the heck does that mean?  This researcher knows his stuff and says, "well, I know 75% of these couples will stay together.  I'll just predict all of them will, and I'll surely be right for 75% of them.  The only problem is, I'm not sure which of the 75% I'll be right about, but I'll be right about 75% of them."  You think, "oh." 

You're now meeting a major issue in social science, that of base rates.  Knowing the base rate of how many of these couples will not make it tells you a ton about just that: how many will make it and how many will not.  Predicting how well couples will do over time is not impressive until you can say that you can do this significantly better than chance (what the base rate would allow you to guess correctly). 

Now, to me, it is impressive when various researchers find that they can correctly classify couple outcomes based on early data in these long term data sets.  In these studies, the statistical methods used do tell us that we can improve upon the prediction one could obtain with base rates alone (significantly beyond chance) by using the variables and methods of the research. All of the researchers mentioned above, as well as others, have shown this ability to improve upon the accuracy rating based on chance alone.  There are some important issues researchers grapple with regarding methods (e.g., Stanley, Bradbury, & Markman, in press), but nevertheless, I think these studies do mean something.

By the way, there is a very technical point it may help to keep in mind when thinking about the degree to which we can predict outcomes for one specific couple.  Prediction rates will generally be very much better (by chance alone) for the group that is more frequently occurring (those couples who stay together over those 8 years) than for those who divorce.  So, if you look carefully at most of these studies that look at prediction (or "classification accuracy possibilities"), you'll find that, within the classifications, the groups of couples that stay together will usually be classified more correctly than those who divorce. 

So, here's how this works.  In our fictitious sample above, we know 25% eventually divorce (we have the data at this point of this research).  It would be entirely unimpressive if we could say we could—using early variables— correctly classify with 25% accuracy those who will divorce.  Heck, any of  us could do that without knowing anything about the couples.  Just predict all of them will break up and you'll have a 25% accuracy rating.  But if a researcher can say, "look, I've correctly classified 50% of those who will divorce."  That's actually quite impressive because it's twice as good as chance.  In many of these studies where you might hear of or see a 84% (for example) accuracy rate, you may also note that, often, they will show something like correct classification of 95% of those who stay together and 50% of those who don't. That will get you the overall accuracy of 84%.

So, given this point, let's think about one specific couple.  As a researcher, you know that within a given data set (where you do capitalize on chance to some degree) you can correctly classify 50% of those 25% who will divorce based on key variables early on.  If you are that one specific couple, that's not a very good prediction rate for you to base a key life decision on.  To the researcher, and to prevention science, it's an important prediction accuracy.  It's telling us a lot about various risks in marriage.  But, it's not telling this one specific couple something very important about themselves.  Hence, my admonition earlier: any of us, researcher or not, are on firmer ground when we tell a given couple they have a significant degree of patterns that put couples at great risk of not making it.  But, when we begin by saying to that one specific couple, "you are not going to make it," the science behind us is now far more tenuous.  (Though, and this is important, there may be data out there where a researcher has done the really most impressive thing of saying, early on, which couples will actually divorce, and finding out years later they were right most of the time.)

The  Important NEWS

The exciting news in all this is that, in most of these studies you hear about, researchers really can improve the base rate prediction in statistically significant ways by looking at certain variables.  Those 85% numbers or 90% numbers mean something important to prevention science and to couples.  In my view, these kinds of studies have yielded immensely helpful advances to our field.  They mean this that, for many couples, the seeds of their eventual divorce (those that will divorce) are there before they even say "I do."  That's a sobering point. 

That point is sound whether or not we can tell one specific couple with high accuracy whether or not they will make it.  Across various samples and methods, we really do know a lot more than we used to know about what puts couples at great risk of marital failure.  That hope is what led my colleague Howard Markman to believe we could develop better preventive interventions by paying close attention to such research.  And, we're learning more all the time in this field.  Many talented researchers are elucidating various kinds of risks (and potential ways to mitigate them).  Knowing risks from a more scientific standpoint can help us all think more clearly about what we may want to target in our interventions, and it can help to get couples’ attention about prevention.

So, the trick is in knowing what these studies mean and what they do not mean.  What we are seeing in study after study of this sort is similarity of some of the variables that go into these high accuracy rates.  Many of the most replicated and potent risk factors identified in these studies have to do with negative interaction.  For example, dynamics like escalation, invalidation, and withdrawal show up time after time after time.

We, in the fields of marital education or therapy, can use these kinds of  studies (summary findings) to both motivate couples and direct couples attention.  On the motivation point, we think it's very useful to tell couples that there are studies that really do show these highly accurate classification rates using early data.  Despite all the complex science about replicability and the exact meaning of those numbers, the phenomena do mean that couples have risks that may hurt them later that exist from early in their relationships.  That, in an of itself, is a very important concept for a couple to understand, especially if we are to get them interested in doing things to prevent marital distress before it develops into something serious. Prevention makes sense precisely because we do believe risk factors are there early and that people can do something about some of them (a far harder point to demonstrate in research, by the way). 

So, we use such studies in part to educate couples on the very basic, major point, that there are such things as risk factors and they might want to think about such things.  For a specific couple, the question is “what risks do you have?”  “What are they doing about them?”  Further, we use such studies to highlight the kinds of patterns within these kinds of studies that have been well replicated (e.g., escalation, invalidation, and withdrawal are dangerous for couples).  While we cannot predict accurately what any one couple may do in years to come (divorce or not, etc.), we can tell couples that such dynamics put couples at risk.  As a point of specific application, we can say that they are likely better off if they can learn ways to reduce the expression of such negative dynamics as early as possible (the heart of prevention).  There are many other risks that studies can and have identified as well (e.g., Karney & Bradbury, 1995; Kurdek, 1993, Larson & Holman, 1994; Rogge & Bradbury, 1999), and a good number of these risks point to strategies that might be employed in interventions.

Summary

It is not likely, at present, that any one person can tell any one couple just how great their odds of marital success or failure.  However, the kinds of studies we have tended to call "prediction" studies have been very useful for identifying a handful of risk factors that are very potent in how marriages do over time.  Couples that really want to fight for their marriages can make use of knowing of such patterns because, in knowing, they can take measures that may increase their odds of building life long, happy marriages.

Related References

Clements, M., Stanley, S.M., & Markman, H.J. (1997).  Predicting Divorce. Manuscript, University of Denver.

Gottman, J.M., Coan, J, Carrere, S., & Swanson, C.  (1998).  Predicting marital happiness and stability from newlywed interactions.  Journal of Marriage and the Family, 60, 5-22. 

Gottman, J.M., & Krokoff, L.J.  (1989).  Marital interaction and satisfaction: A longitudinal view.  Journal of Consulting and Clinical Psychology, 57, 47-52.

Karney, B.R., & Bradbury, T.N.  (1995).  The longitudinal course of marital quality and stability: A review of theory, method, and research.  Psychological Bulletin, 118, 3-34.

Kurdek, L.A. (1993).  Predicting marital dissolution: A 5-year prospective longitudinal study of newlywed couples.  Journal of Personality and Social Psychology, 64, 221-242.

Larson, J.H. & Holman, T.B. (1994). Premarital predictors of marital quality and stability. Family Relations, 43(2), 228‑237.

Markman, H.J., Stanley, S.M., & Blumberg, S.L.  (1994).  Fighting for Your Marriage: Positive Steps For A Loving and Lasting Relationship.  San Francisco:  Jossey Bass, Inc.

Matthews, L.S., Wickrama, K.A.S., & Conger, R.D. (1996).  Predicting marital instability from spouse and observer reports of marital interaction.  Journal of Marriage and the Family, 58, 641-655 

Rogge, R.D., & Bradbury, T.N. (1999).  Recent Advances in the Prediction of Marital Outcomes.  In R. Berger & M.T. Hannah (Eds.) Handbook of preventive approaches in couples therapy. New York: Brunner/Mazel.  In R. Berger & M. Hannah, (Eds.), Handbook of preventive approaches in couple therapy.  New York: Brunner/Mazel. 

Stanley, S.M., Blumberg, S.L., & Markman, H.J.  (1999).  Helping Couples Fight for Their Marriages: The PREP Approach.  In R. Berger & M. Hannah, (Eds.), Handbook of preventive approaches in couple therapy.  New York: Brunner/Mazel.  Pp. 279-303.

Stanley, S. M., Bradbury, T.N., & Markman, H.J.  (2000).  Structural flaws in the bridge from basic research on marriage to interventions for couples: Illustrations from Gottman, Coan, Carrere, and Swanson (1998).  Journal of Marriage and the Family, 62(1), 256-264.