The core of relationship science is well supported: strong relationships predict greater wellbeing and even longer life, similarity predicts attraction better than opposites, and certain conflict patterns predict a higher risk of breakdown. Less settled is how stable attachment styles are over a lifetime and how accurately outcomes can be predicted for a given couple. And some famous claims, particularly eye-catching divorce-prediction accuracy figures, are more overstated than the underlying patterns warrant. The honest picture is confident at the centre and careful at the edges.
How relationships are actually studied
Because relationships are personal and hard to run experiments on, the field leans on several methods, each with its own strengths and blind spots. Claims are on the firmest ground when these different approaches point the same way.
Longitudinal studies
Following the same people or couples for years, or decades, as in the Harvard Study of Adult Development, reveals how relationships and outcomes unfold over time. This is the best tool for questions about what predicts a good life, though it cannot by itself prove cause.
Observational studies
Watching and coding how couples actually interact, as in Gottman's laboratory work on conflict, captures real behaviour rather than self-report. Its limit is that behaviour in a lab, or knowing you are watched, is not always ordinary life.
Surveys and self-report
Large questionnaires measure satisfaction, attachment, and attraction across many people. They scale well and reach diverse samples, but they depend on how honestly and accurately people report their own feelings and behaviour.
Meta-analysis
Statistically combining many studies, as with the reviews linking social ties to longevity, produces more reliable estimates than any single study. It is only as good as the studies it pools, but it is the strongest evidence for broad, repeated effects.
No single study settles a question. Confidence comes from convergence: when long-term cohorts, observation, surveys, and meta-analyses agree, a claim is solid. When they conflict or run short, honesty means calling the question open.
A few robust figures
These are among the more stable numbers in the field. Even so, read them as well-supported ranges rather than precise constants, since they shift with the populations and methods studied.
Settled: what the evidence firmly supports
These claims rest on multiple methods pointing the same way, over many studies and many years. They are not seriously disputed among mainstream researchers, even where the exact numbers still move.
Strong relationships predict greater wellbeing and longer life. This is one of the best-supported findings in all of social science. Large meta-analyses show that people with stronger social ties have substantially better odds of survival, an effect on mortality that stands comparison with well-known physical risk factors, and decades-long studies like the Harvard Study of Adult Development find relationship quality among the strongest predictors of late-life health and happiness. These are population-level associations rather than individual guarantees, and untangling cause is genuinely hard, but the link itself is robust and repeatedly replicated.
Similarity predicts attraction better than opposites do. Despite the popular slogan, the evidence consistently favours similarity. Across studies of initial attraction and of long-term couples, shared values, interests, background, and outlook predict who we are drawn to and who we stay with better than difference does. Complementary contrasts exist and can be enjoyable, but they sit on a base of common ground. The "opposites attract" idea is a durable myth rather than a finding.
Certain conflict patterns are linked to a higher risk of breakdown. Observational research, especially Gottman's, consistently associates habitual contempt, criticism, defensiveness, and stonewalling with worse relationship outcomes, with contempt the most corrosive. That these destructive patterns matter, and that repair and a healthy balance of positive interactions protect a relationship, is well established. The pattern-to-risk link is solid, even where the precise predictive power for a given couple is not, as the contested section explains.
Mixed or evolving: where the science is still moving
These questions are actively researched and partly answered, but the evidence is incomplete or the estimates shift with method. Here the honest posture is confidence about the broad shape and caution about the details.
How stable attachment styles are across a lifetime. There is real continuity between early attachment and later patterns, but also real change: styles can shift with new relationships, major life events, and therapy. How much of attachment is enduring versus changeable, and how strongly childhood attachment predicts specific adult outcomes, is still debated, with estimates varying by how attachment is measured. The safe summary is that attachment is influential and somewhat stable, but not fixed for life.
How well outcomes can be predicted for an individual couple. Research identifies factors linked, on average and across groups, to satisfaction and breakdown. Turning those group-level patterns into an accurate forecast for one specific relationship is much harder. People change, circumstances shift, and averages do not decide individual cases. The predictors are meaningful at the population level while remaining rough guides, not verdicts, for any single couple.
The precise size of specific effects, like the positive-to-negative ratio. The broad principle that stable relationships carry a clear surplus of positive interactions over negative ones is well supported. The exact figure, often quoted as five to one, is better treated as a memorable rule of thumb than a precise constant, since it depends heavily on how interactions are defined and counted. The direction is settled; the exact number is softer than its fame implies.
Contested: where claims outrun the evidence
These are not questions with a hidden right answer so much as places where popular retellings have run ahead of what the data support. The underlying research is often sound; the confident headlines built on it are not.
The "predicting divorce with over 90 percent accuracy" claim. This famous figure is widely misunderstood. Such accuracy comes largely from statistical models fitted to couples whose outcomes were already known, which flatters the result. Genuine forward prediction, forecasting the future of new couples not used to build the model, is considerably less accurate, and some of the striking numbers have been criticised on exactly these grounds. The conflict patterns are real and useful; the precise accuracy figures are overstated and should be read with scepticism.
Rigid, formulaic pop-relationship rules. Much popular advice packages genuine findings into tidy laws: exact ratios, fixed personality-match rules, or the idea that people fall into permanent unchangeable types. The kernels are often real, but the certainty and precision are added later for appeal. Relationships are variable and context-dependent, and formulas that promise to sort them neatly usually oversell modest, probabilistic patterns as firm rules.
Strong causal claims from correlational data. Many relationship findings are associations, strong ties go with longer life, certain habits go with satisfaction, but the direction of cause is often unclear. Happier, healthier people may find it easier to build good relationships, as much as the other way round, and both likely feed each other. Headlines that leap from "linked to" to "causes" claim more than the evidence, which is usually observational, can support.
A common misreading
Because the contested pile exists, it is tempting to dismiss the whole field as soft. That is the mistake to avoid.
Ask which pile a claim belongs in. The overstated accuracy figures and rigid formulas sit at the edges; they do not undo the settled core. That strong relationships support wellbeing and longevity, that similarity beats opposites, and that contempt and its cousins corrode a bond are well established. When you meet a bold relationship claim, notice whether it distinguishes a group-level pattern from an individual certainty, whether it cites the weight of evidence or a single striking study, and whether it treats a probabilistic tendency as an iron law. The most trustworthy sources are neither dismissive nor breathless, and they tend to sit closest to the truth.
Where to go next
If this page has weighed the evidence, the other pages put it to use. Start with the overview for why connection matters, read what makes relationships work for the habits the settled findings support, and see conflict and repair for the patterns behind the breakdown research.
Sources
- Holt-Lunstad J, Smith TB, Layton JB. Social relationships and mortality risk: a meta-analytic review. PLoS Medicine. 2010;7(7):e1000316.
- Waldinger RJ, Schulz MS. The Good Life: Lessons from the World's Longest Scientific Study of Happiness. 2023.
- Heyman RE, Slep AMS. The hazards of predicting divorce without crossvalidation. Journal of Marriage and Family. 2001;63(2):473-479. (On the limits of divorce-prediction accuracy.)
This page is educational and summarises general patterns in relationship research. It is not therapy or personalised advice, and findings about groups do not determine any individual relationship. If your relationship is causing you distress, a qualified counsellor can help.