There is a pattern we keep noticing in the way Caryl Rusbult's investment model gets cited in relationship explainers — and it has nothing to do with whether the model works. It works. The pattern is about which of the three predictors readers walk away remembering, and which one the meta-analyses say is actually doing the heavy lifting. Those two are not the same variable, and the distance between them is large enough to change the model's practical implications.
A brief orientation, because the scaffolding matters. Rusbult (1980), writing in the *Journal of Experimental Social Psychology*, proposed that commitment to a romantic relationship could be predicted by three variables: satisfaction level, quality of alternatives, and investment size. The equation is usually written as C = f(S, A, I). The model built directly on Thibaut and Kelley's (1959) interdependence theory, which already contained satisfaction and comparison-to-alternatives as its structural predictors. Rusbult's distinctive contribution was the investment term — the idea that commitment also depends on what you would lose if the relationship ended, the tangible and intangible resources you cannot take with you when you leave. Call it the sunk-cost clause of the equation, though Rusbult herself was careful to distinguish investments from pure sunk-cost reasoning.
Here is the discovery we kept running into when we went back to the primary literature: that third term, the one Rusbult added to the existing framework, is the weakest of the three predictors whenever the model is tested at meta-analytic scale. That is not an obscure methodological footnote. It is the Le and Agnew (2003) meta-analysis in *Personal Relationships*, 52 studies, roughly 11,000 participants, and it has been public for more than twenty years. Yet most introductions to the model still present the three predictors as though they were equal partners, and some go further and treat the investment term as the model's central insight. It isn't. It is the weakest leg of the tripod.
Before we push on that, the concession the model deserves. Across those 52 studies, the three variables together accounted for roughly 61% of the variance in self-reported commitment, which is a large number by the standards of this literature. Satisfaction correlated with commitment at approximately r = .68, quality of alternatives at approximately r = -.48, and investments at approximately r = .46 (Le & Agnew, 2003). A model that explains that much of that construct across that many samples is a model that has earned its citations. We are not arguing that Rusbult was wrong. We are arguing that the way the model gets taught reverses the internal weights of its own predictors, and the reversed version smuggles in advice the data do not support.
The Three-Variable Shortcut Pattern
Every time a popular relationship explainer cites Rusbult, roughly the same sentence appears: commitment depends on satisfaction, alternatives, and investments. The sentence is literally correct and yet subtly misleading, because it implies the three predictors operate as peers.
The shortcut happens because the equation is memorable. C = f(S, A, I) fits on a slide. Readers walk away with "three things predict commitment" and the three things get listed in the order Rusbult used in her original 1980 paper, which happens to place investments last — where it sounds climactic rather than subordinate. The theoretical history of the model compounds the effect. Because investments were Rusbult's distinctive addition to Thibaut and Kelley, secondary sources naturally highlight them as the novel contribution, and novelty gets confused with explanatory weight. The part of the equation Rusbult inherited disappears from the summary. The part she contributed gets the spotlight. That rearrangement is the exact opposite of what the subsequent meta-analyses show.
There is also a measurement issue underneath the shortcut. Rusbult's (1980) original study used a small undergraduate sample and operationalized investment size through subjective self-report on numerical scales. Respondents were asked how much they felt they had put into the relationship, not how much they had measurably put in. That methodological choice means the "investment" term in the original model was always closer to "perceived investment" than to any objective accounting. The shortcut citations rarely mention this. They talk about shared time, shared possessions, and shared social networks as though these were what Rusbult measured. They weren't — or rather, they became what the model measured much later, in the 1998 Investment Model Scale, which is not the same instrument as the 1980 procedure.
The Weakest-Predictor Problem
When the model is tested meta-analytically, the investment term is the term that shrinks. Le and Agnew (2003) reported satisfaction as the strongest predictor by a clear margin, quality of alternatives second, and investment size third. The three-way ordering has been stable in subsequent replications and extensions. Across samples that varied by age, relationship length, marital status, and country of origin, the internal rank order of the predictors held.
This matters because Rusbult's original theoretical argument was not that investments would be the weakest predictor. Her argument was that the investment term explained a distinctive variance that satisfaction and alternatives could not reach on their own — the path-dependent part of commitment, the stickiness that persists even when satisfaction drops and alternatives look attractive. The meta-analytic reality is that investments do predict commitment independently of the other two, but the magnitude of that independent contribution is smaller than the part of the equation Rusbult inherited from Thibaut and Kelley. The distinctive contribution turned out to be the smallest contribution.
The practical consequence is that if you strip the investment term out of the model entirely, you do not lose as much predictive power as the model's theoretical history implies you should. You lose some — the term is not zero, and its independent contribution is statistically reliable across studies — but what you are left with is something close to interdependence theory with a commitment outcome variable attached. The novelty of the Rusbult framework, measured in variance explained, is smaller than the novelty measured in citation count.
The term Rusbult personally added to interdependence theory is the term four decades of meta-analytic evidence have found contributes the least to predicting the thing the model is supposed to predict.
The Shelter-Study Paradox
This is where the model gets ethically uncomfortable, and where two primary documents inside Rusbult's own research program say things that only fit together when you read them carefully.
The first primary document is Rusbult (1983) in the *Journal of Personality and Social Psychology*, the longitudinal test of the investment model. That paper framed commitment as a generally constructive mechanism — a process by which couples accumulate satisfaction, recognize the relative scarcity of good alternatives, and build up investments that stabilize the bond over time. The tone is mildly celebratory. Commitment, in the 1983 framing, is what healthy relationships have.
The second primary document is Rusbult and Martz (1995) in *Personality and Social Psychology Bulletin*, which applied the same model to women residing in a domestic violence shelter. The finding was that investment size and poor quality of alternatives predicted whether women returned to abusive partners. The model worked — the equation described the decision cleanly. But the normative valence is inverted. In the 1995 paper the same mechanism that 1983 framed as constructive becomes the mechanism that keeps people in relationships that are actively harming them.
These two papers are not contradictory at the level of prediction. They are contradictory at the level of what the reader is supposed to do with the prediction. The 1983 paper implies that building investments is how couples weather rough patches. The 1995 paper shows that sufficient investments will hold a couple together even when one partner is hitting the other. Both readings are correct descriptively, and the contradiction unwinds once you notice that Rusbult's model was always agnostic about whether the relationship being predicted was a good one. The shortcut citations almost always import the 1983 normative framing and skip the 1995 evidence, which is how a descriptive social-psychological model quietly becomes relationship advice.
The Measurement Drift Pattern
The fourth pattern is a slow drift between the 1980 theoretical construct and the 1998 Investment Model Scale (Rusbult, Martz, & Agnew, 1998, *Personal Relationships*). The scale is well-validated and is now the standard way the investment model gets tested. It is also not quite measuring what the 1980 paper described.
The scale's investment subscale includes items about shared possessions, shared memories, shared friends, and shared future plans. Those are reasonable operationalizations, but they collapse a theoretical distinction that Stanley and Markman (1992) made explicit in the *Journal of Family Psychology*: the distinction between constraint commitment (things that make leaving costly) and dedication commitment (things that make staying desirable). Rusbult's 1980 investment construct was closer to pure constraint — the resources you would forfeit by exiting. The 1998 scale, by including forward-looking items like shared plans, partly absorbs dedication into the investment term. When researchers report "investment predicts commitment," they are sometimes reporting that the constraint-plus-dedication aggregate predicts commitment, which is closer to a tautology than the original model contained.
This is not a fatal problem for the scale — it performs well psychometrically and its reliability across cultures is strong — but it is a reason to read any post-1998 study of the investment model with a specific question in mind: which investment, the 1980 theoretical one or the 1998 measured one. The two have drifted enough that a careful reader should know which is on the page before they generalize from the results.
So What Do You Actually Do
If you are reading the research literature: treat the investment model as a three-variable equation in which the predictors are ranked, not equal. Satisfaction first, alternatives second, investments third. When you see a paper citing Rusbult, check which version of the model is being invoked — the 1980 theory, the 1983 longitudinal test, the 1995 shelter application, or the 1998 scale — because they do not all operationalize investments the same way. And when a popular summary lists the three predictors as peers, mentally re-weight them using Le and Agnew's (2003) correlations before you decide what the model is telling you.
If you are reading the model as a person thinking about a real relationship: the most important implication of the literature is probably the one the shortcut citations bury. The investment term is a descriptive variable, not a prescriptive one. The fact that leaving would be costly is information about the cost of leaving, not information about whether leaving is the right decision. The 1995 Rusbult and Martz shelter study is the cautionary data point that every reading of this model should keep visible. If you catch yourself using "we have invested so much" as the primary argument for staying, you are reasoning inside the part of the equation meta-analysis says has the smallest independent effect, and you are doing it in a way the model's own authors documented can keep people in danger.
We would revise this reading if a longitudinal study with a follow-up period longer than twenty-four months, drawn from a non-WEIRD sample, found investment-term effect sizes that matched or exceeded satisfaction in predicting commitment. That study would genuinely reopen the weighting question and might recover the distinctive explanatory role Rusbult originally claimed for the investment term. Until a result at that scale contradicts Le and Agnew (2003), the internal rank order holds, and the investment model's distinctive contribution remains its smallest one.