Counterfactual analysis lets you mentally test decisions before you make them: "If we had set this boundary differently three months ago, where would we be now?" or "What would happen if we actually committed to couples therapy?" This kind of clear-eyed modeling helps separate imagination from reality.
Counterfactual analysis asks "what would have happened if X had been different?" In AI terms, it means using data from similar relationships to model alternative outcomes for your own decisions. If you're deciding whether to move for your partner's job, counterfactual analysis can examine relationships that made that choice and didn't, extracting patterns about outcomes based on relationship factors you can measure.
Relationship decisions like relocation, cohabitation, or starting a family involve irreversible commitments with long-term consequences. You can't run A/B tests on your own life. Counterfactual analysis approximates this by finding "matched" couples (similar age, income, attachment style, communication patterns, values alignment) and analyzing what happened when they made similar decisions versus alternative choices.
The value isn't prediction; it's revealing hidden dependencies. You think moving for a partner's job might strain your career. Counterfactual analysis of 50 similar couples who did versus didn't move shows the actual relationship outcomes correlated with move—not just career impact, but divorce rates, satisfaction trajectories, and which couples recovered career momentum and which didn't.
The technical approach: aggregate anonymized relationship data (from surveys, therapy notes, or published studies) on couples who share your baseline characteristics (duration, income gap, children, etc.). Identify a subset who made your decision (moved for partner's job) and a control subset who didn't. Use causal inference techniques—specifically counterfactual inference models—to estimate what outcomes the "control" couples would have experienced if they'd made the "treatment" choice.
These models use statistical matching to find couples similar to you across dozens of dimensions, then estimate the treatment effect: "For couples like you, moving for a partner's job correlates with X% increased divorce risk, Y% career satisfaction reduction, and Z% relationship satisfaction change." Critically, this isolates the causal effect of the move itself, not confounds (like "people who move are generally more flexible").
Under the hood, systems use propensity score matching—calculating the probability that a couple like you would make the decision you're considering, then matching you with couples in the opposite group who have similar propensity scores. This creates an apples-to-apples comparison. Then causal directed acyclic graphs (DAGs) model which variables influence which outcomes, separating causation from correlation.
For example: moving correlates with divorce. But does moving cause divorce, or do couples already in weak relationships move for a "fresh start"? A DAG that includes prior relationship satisfaction helps distinguish these. Counterfactual inference then estimates: "If a couple with your satisfaction level, income distribution, and communication pattern were to move, what outcome shifts would likely occur?"
Counterfactual reliability depends entirely on data quality and size. Small samples produce high variance estimates. If your specific combination of factors (high income, mixed cultural backgrounds, three kids, rural to urban move) is rare in the dataset, the counterfactual estimate becomes speculative. Also, historical data doesn't account for how your particular relationship dynamics differ from the matched sample.
Another critical limitation: unmeasured confounds. If the dataset captures income, satisfaction, and values but not, say, your partner's conflict avoidance or your own ambivalence about career, those unmeasured factors could reverse the counterfactual estimate. Counterfactual analysis can't prove causation; it can only suggest likely associations given the variables in the model.
Before a major decision, articulate the choice clearly (move or stay, have children or don't, cohabitate or maintain separate homes). Research whether published relationship studies or anonymized datasets exist covering couples who made similar decisions. Ask an AI system trained on relationship data to perform counterfactual matching: identify couples similar to you who chose differently, and estimate outcome differences. Use this not as prediction, but as a reality check on your intuitions.
Try this: Identify one major decision you're facing with your partner. Search for published relationship studies that compare outcomes for couples who made that choice versus alternatives (e.g., "relationship satisfaction outcomes of long-distance vs. cohabiting couples"). Use Claude to extract the study design, sample characteristics, and outcome data. Ask: "For couples with our characteristics (age, income, relationship length), what outcomes did this study associate with each choice?" This gives you empirical reference points rather than pure speculation.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.