Machine learning uncovers the most robust self-report predictors of relationship quality across 43 longitudinal couples studies

Significance

What predicts how happy people are with their romantic relationships? Relationship science—an interdisciplinary field spanning psychology, sociology, economics, family studies, and communication—has identified hundreds of variables that purportedly shape romantic relationship quality. The current project used machine learning to directly quantify and compare the predictive power of many such variables among 11,196 romantic couples. People’s own judgments about the relationship itself—such as how satisfied and committed they perceived their partners to be, and how appreciative they felt toward their partners—explained approximately 45% of their current satisfaction. The partner’s judgments did not add information, nor did either person’s personalities or traits. Furthermore, none of these variables could predict whose relationship quality would increase versus decrease over time.

Abstract

Given the powerful implications of relationship quality for health and well-being, a central mission of relationship science is explaining why some romantic relationships thrive more than others. This large-scale project used machine learning (i.e., Random Forests) to 1) quantify the extent to which relationship quality is predictable and 2) identify which constructs reliably predict relationship quality. Across 43 dyadic longitudinal datasets from 29 laboratories, the top relationship-specific predictors of relationship quality were perceived-partner commitment, appreciation, sexual satisfaction, perceived-partner satisfaction, and conflict. The top individual-difference predictors were life satisfaction, negative affect, depression, attachment avoidance, and attachment anxiety. Overall, relationship-specific variables predicted up to 45% of variance at baseline, and up to 18% of variance at the end of each study. Individual differences also performed well (21% and 12%, respectively). Actor-reported variables (i.e., own relationship-specific and individual-difference variables) predicted two to four times more variance than partner-reported variables (i.e., the partner’s ratings on those variables). Importantly, individual differences and partner reports had no predictive effects beyond actor-reported relationship-specific variables alone. These findings imply that the sum of all individual differences and partner experiences exert their influence on relationship quality via a person’s own relationship-specific experiences, and effects due to moderation by individual differences and moderation by partner-reports may be quite small. Finally, relationship-quality change (i.e., increases or decreases in relationship quality over the course of a study) was largely unpredictable from any combination of self-report variables. This collective effort should guide future models of relationships.

Romantic relationship quality—a person’s subjective perception that their relationship is relatively good versus bad (1)—is a powerful psychological construct with far-reaching societal consequences and policy implications (Fig. 1). Unhappy marriages are associated with many negative stress-related outcomes (2), including poor physical health (3), high blood pressure (4), poor immune system functioning (5), mortality (2), and risk of mental health problems (6). Low marital quality spills over into people’s professional and personal lives, predicting lost work productivity (7) and lower well-being for children (8, 9).

Fig. 1.

Antecedents and consequences of relationship quality (1–9). Schematic depiction of the field of relationship science. In their work, relationship scientists use an extensive assortment of overlapping individual difference and relationship-specific constructs. These constructs predict the way couple members behave toward and interact with each other, which in turn affects relationship quality and a variety of consequential outcomes. These processes are themselves embedded in social networks as well as broader cultural and historical structures.

As the importance of relationships for health, work productivity, and parent/child well-being has entered public awareness, there has been an explosion of research attempting to explain, predict, and improve relationship quality. That is, why do some partners feel especially positively about their relationship, and why do these evaluations change (10)? Interest in this question across many disciplines—including psychology, sociology, communication, economics, and family studies—has transformed relationship quality into one of the most central and pervasive outcome variables in the social sciences, and a primary focus of applied efforts to strengthen marriages [e.g., the multimillion dollar Healthy Marriage and Relationship Education Grant program in the United States (11)]. These efforts have resulted in a wide array of constructs and concepts that—via interpersonal, behavioral processes—shape relationship quality and relationship stability (see refs. 12–15 for reviews). Some of these variables characterize individuals (e.g., age at marriage, attachment style, neuroticism) (Fig. 1, Upper Left box), whereas others characterize partners’ perceptions and experiences within the relationship itself (e.g., conflict, sex, relationship length, domestic violence) (Fig. 1, Lower Left box).

A key challenge now—more than 20 y after the emergence of relationship science as a mature discipline (16)—is to make this knowledge cumulative. In a critique of the field, Reis (17) highlights an important factor that has historically limited scholars’ ability to organize their efforts into a coherent body of knowledge: The tendency of the current academic system to reward individual contributions rather than team science. Indeed, a collectivistic approach would be particularly beneficial to relationship science for several reasons. First, couples are costly to recruit, necessarily limiting the statistical power that can be achieved in a given study by a single laboratory. Second, participants become fatigued after completing too many measures, limiting the number of constructs that can be examined in a given study. Third, traditional techniques (e.g., regression) make it easy for researchers to mistakenly overfit statistical models to individual datasets and are suboptimal for comparing the predictive importance of constructs (18, 19). The result of these practical research constraints is that no individual laboratory has the resources or means to compare the efficacy of the growing list of important constructs, much less their affiliated theoretical frameworks.

To document the most reliable predictors of relationship quality and the relative predictive power of different measurement strategies, the ideal study would combine the longitudinal and dyadic data-collection efforts of multiple independent laboratories, it would include a wide array of published and not previously published predictors, and it would use preregistered statistical procedures that permit data exploration without overfitting. This paper reports the conclusions of such a study. The project combines the efforts of 86 relationship researchers by examining 43 longitudinal datasets (funded by 39 national/university grants) with 11,196 couples (baseline n = 22,163 participants) and 2,413 (mostly self-report) measures collected at baseline. The datasets tracked couples for an average of four time points (range = 2 to 11 time points) over 14 mo (range = 2 to 48 mo). The baseline measures collected from each partner were used to predict relationship quality at baseline (the first time point collected), at follow-up (the last time point collected), and over time (i.e., each participant’s slope calculated across all available time points). This design provides initial answers to the questions of: 1) How much variance in relationship quality can researchers predict? and 2) What types of psychological measures most reliably emerge as predictors of relationship quality?

Data Solicitation Strategy

Datasets were eligible to be included in the study if they included: 1) Data from both romantic partners of each couple, 2) data collected from at least two time points that were at least 2 mo apart, and 3) a measure of relationship satisfaction collected at each time point.

The overall design and analysis plan for the project was preregistered on June 15, 2018 (https://osf.io/g9sqf/). We used listservs (Society for Personality and Social Psychology and International Association for Relationship Research), social media (Twitter), and the Open Science Framework (OSF) StudySwap platform to invite researchers with dyadic longitudinal datasets to join the project. We solicited new datasets from June 15 to October 1, 2018. A total of 48 datasets were committed to the project, of which 43 datasets were ultimately provided. Datasets were analyzed on a rolling basis from June 18, 2018 (Dataset 1) to March 25, 2019 (Dataset 43). For each dataset, coauthors provided a codebook outlining their design and measures. Each codebook was used to tailor an analysis plan, and each was preregistered prior to analysis (i.e., 43 preregistered analysis plans total).

Measures

The dependent measure was relationship quality (i.e., a person’s subjective perception that their relationship is relatively good vs. bad; a person’s evaluation of the relationship), and our primary operationalization of this construct consisted of relationship-satisfaction measures. Commitment was used as an additional operationalization of relationship quality in the datasets that included it (31 datasets). We selected satisfaction as our primary dependent measure because it is the most common dependent measure used in relationship science—we have never encountered a couples dataset that lacked it—and we selected commitment because it is theoretically central to the field and nearly as pervasive (13).

The remaining self-report measures collected at baseline were used as predictors; the specific predictors included varied from dataset to dataset. Baseline measures were categorized into two groups of predictors: Individual difference variables (judgments about the self, such as traits and characteristics) and relationship-specific variables (judgments about the relationship or the partner, and variables that are, by definition, identical for both couple members, such as relationship length). Although the major theories of relationships differ with respect to which specific individual and relationship variables they emphasize, both classes of variables are purported to make independent or interactive contributions in virtually all of them (e.g., attachment theory, interdependence theory, the interpersonal process model of intimacy, relational regulation theory, risk regulation theory, the vulnerability-stress-adaptation model) (see ref. 15 for a review). Furthermore, two versions of each predictor were available in all datasets: An actor-reported version (Amir’s individual/relationship variable used to predict Amir’s satisfaction), and a partner-reported version (Amir’s partner Alex’s individual/relationship variable used to predict Amir’s satisfaction). The distinction between actor and partner is also central to relationship science (20), and their purported joint importance is often the raison d’être of intensive dyadic data collection efforts.

Four relationship-specific variables—trust, intimacy, love, and passion—are often conceptualized as predictors of relationship quality (21–23). But alternatively, they could be conceptualized as indicators of relationship quality, as these four variables may tap relationship quality approximately as well as satisfaction and commitment do (1). It is therefore possible that retaining these measures as predictors artificially inflates the amount of variance that relationship-specific variables can collectively explain. In the models presented below, we removed the actor and partner versions of trust, intimacy, love, and passion as predictors (59 total variables across 21 of the datasets). A version of the analyses in which these predictors are retained, consistent with our preregistered analysis plan, is also presented in SI Appendix.

The initial categorization of variables into individual versus relationship variables was made by the authors of each dataset. After all 43 datasets had been compiled, the first and second author combined the predictors into a master list of individual versus relationship variables, and recategorized variables as necessary to ensure consistent categorization across datasets (see the OSF for procedural details). We next identified constructs that were measured multiple times across datasets and grouped each one using a common code. For example, the item, “How old are you?” from Dataset 1 and the item “Age in years” from Dataset 4 were each coded as “age.” This coded master list of predictors was then used to compute the predictive success rate of each construct.

Data Availability

Analysis plans, final syntax files, and word files outlining any preregistration changes can be found for each dataset in the OSF (https://osf.io/d6ykr/). Analytic features of each included dataset are reported in Table 1. Demographic features of each dataset can be found in SI Appendix. Meta-analytic materials and data, including the final master list of predictors and the syntax used to compute success rates, are also available in the OSF (https://osf.io/v5e34/). The raw datasets are too ethically sensitive to make publicly available. However, S.J. will work with any professional scholar to obtain access to the raw data for any of the 43 individual datasets.

Table 1.

Analytic features of the 43 datasets

Dataset	Baseline n	Follow-up n	Changen	No. of individual predictors	No. of relationship predictors	Baseline satisfaction mean (SD)	Follow-up satisfaction mean (SD)	Baseline commitment mean (SD)	Follow-up commitment mean (SD)
1	148	133	146	97	50	6.01 (0.89)	5.56 (1.53)	5.88 (1.25)	5.63 (1.59)
2	240	228	240	98	50	5.84 (1.21)	5.59 (1.58)	6.77 (0.54)	6.49 (1.05)
3	176	156	154	13	6	6.05 (1.02)	6.00 (1.09)	NA	NA
4	166	166	166	32	71	5.31 (0.69)	5.01 (1.02)	NA	NA
5	350	316	343	42	50	69.59 (9.49)	66.18 (13.87)	6.87 (0.43)	6.71 (0.72)
6	172	90	90	9	5	131.20 (21.04)	121.48 (31.16)	NA	NA
7	201	119	116	11	9	132.05 (21.00)	122.84 (30.67)	NA	NA
8	194	157	155	9	22	5.86 (1.19)	5.74 (1.27)	6.19 (1.04)	6.11 (1.10)
9	129	126	126	4	11	6.03 (1.05)	5.93 (1.25)	6.59 (0.77)	6.38 (1.07)
10	88	61	61	7	10	7.96 (0.99)	7.79 (1.38)	6.72 (0.57)	8.26 (1.03)
11	159	117	115	23	15	6.01 (0.88)	5.68 (1.22)	6.13 (0.91)	5.98 (1.05)
12	124	124	124	9	8	6.03 (0.72)	6.02 (0.80)	NA	NA
13	200	145	192	27	18	5.92 (0.76)	5.97 (1.00)	6.48 (0.65)	6.39 (0.90)
14	122	106	106	21	21	5.97 (0.85)	5.93 (1.07)	6.34 (0.84)	6.26 (1.05)
15	239	158	206	33	20	6.84 (1.60)	6.82 (1.65)	7.48 (0.93)	7.39 (1.10)
16	450	365	410	11	5	6.45 (0.68)	6.09 (0.96)	6.81 (0.45)	6.62 (0.75)
17	345	120	195	40	21	5.98 (0.91)	5.55 (1.38)	6.11 (1.05)	5.93 (1.29)
18	245	107	192	11	29	6.78 (1.21)	6.71 (1.08)	6.75 (1.17)	6.85 (0.96)
19	80	32	51	6	11	28.95 (4.61)	27.44 (5.46)	NA	NA
20	386	278	343	37	41	42.65 (5.14)	41.26 (6.81)	NA	NA
21	255	189	189	41	32	5.97 (0.83)	5.93 (0.84)	6.47 (0.73)	6.34 (1.04)
22	347	216	283	24	22	6.02 (0.76)	5.82 (0.93)	6.48 (0.67)	6.23 (1.08)
23	318	258	289	21	19	41.89 (4.56)	41.21 (5.83)	NA	NA
24	394	230	372	17	15	4.52 (0.49)	4.50 (0.55)	4.87 (0.25)	4.86 (0.36)
25	172	118	144	32	29	70.69 (9.06)	76.63 (7.78)	6.53 (0.65)	6.44 (0.69)
26	464	322	322	32	4	−0.00 (0.97)	−0.00 (1.02)	6.53 (1.68)	6.58 (1.94)
27	254	247	247	75	69	6.16 (0.89)	5.95 (1.14)	5.45 (0.63)	5.37 (0.59)
28	206	130	158	12	14	4.45 (0.70)	4.48 (0.70)	5.98 (0.88)	5.88 (0.90)
29	564	261	478	32	19	4.46 (1.21)	4.34 (1.36)	5.61 (1.08)	6.00 (1.07)
30	237	208	205	16	19	6.11 (1.02)	5.92 (1.31)	6.64 (0.80)	6.46 (1.01)
31	203	167	167	88	28	31.23 (2.69)	31.24 (3.27)	NA	NA
32	196	136	196	8	4	5.96 (1.13)	5.85 (1.23)	6.33 (1.00)	6.19 (1.08)
33	156	156	156	9	10	17.65 (3.63)	17.99 (3.76)	NA	NA
34	323	316	316	17	11	16.90 (2.93)	16.95 (3.37)	NA	NA
35	192	161	161	20	17	5.89 (1.06)	5.74 (1.38)	6.41 (0.88)	6.29 (1.14)
36	111	139	111	44	2	117.86 (22.45)	123.06 (19.42)	NA	NA
37	97	31	72	12	19	5.22 (1.50)	5.35 (1.33)	6.19 (0.96)	6.45 (0.95)
38	12,200	7,731	9,886	63	26	5.42 (1.60)	5.89 (1.28)	1.52 (0.88)	1.57 (0.39)
39	373	190	322	58	131	5.54 (0.93)	5.49 (0.97)	6.80 (0.90)	6.84 (0.87)
40	151	109	133	39	54	6.66 (1.61)	7.00 (1.16)	6.75 (1.08)	6.74 (0.90)
41	240	181	181	38	24	7.63 (1.16)	5.92 (1.10)	7.79 (1.30)	6.05 (1.02)
42	390	351	327	13	19	41.39 (4.65)	39.98 (6.19)	6.55 (0.56)	5.14 (0.49)
43	144	73	73	14	31	5.09 (0.72)	5.09 (0.83)	7.83 (1.25)	7.95 (1.26)

Note: The three n columns refer to the number of usable participants in the models predicting baseline, follow-up, and change in satisfaction, respectively. See SI Appendix, Table S2 for dataset authorship details. Note that for datasets with more than two time points, change scores could still be calculated for some participants whose data were missing at the final wave. NA, not applicable.

Analysis Strategy

Machine Learning.

Each dataset was analyzed using Random Forests (24), a machine-learning method designed to handle many predictors at once while minimizing overfitting (i.e., fitting a model so tightly to a particular dataset that it will not replicate in other datasets). The Random Forests method builds on classification and regression trees (25). Specifically, using a random subset of predictors and participants, the Random Forests method tests the strength of each available predictor one at a time through a process called recursive partitioning. It builds a decision tree out of the strongest available predictors and tests the tree’s overall predictive power on a subset of data that were not used to construct the tree (also called the “out of bag” sample). The Random Forests method does this repeatedly, separately bootstrapping thousands of decision trees and then averaging them together. Results reveal how much variance in the dependent measure was predictable and which predictors made the largest contributions to the model. Random Forests are nonparametric—they do not impose a particular structure on the data—and as such they are able to capture nonlinear relationships, including interactions among the predictors (26). For example, a model with actor- and partner-reported predictors would detect any robust actor × partner interactions (e.g., moderation, attenuation effects, matching effects) that could not be captured in a model featuring actor- or partner-reported predictors alone.

Each model was conducted using the “randomForest” package for R, with the same tuning parameters that we have used in previous research (27). Specifically, we set “ntree” to 5,000 for all analyses, meaning that each Random Forests model was constructed from 5,000 regression trees, and we left “mtry”—the number of predictors available for splitting at each tree node—at its default value of one-third of the total number of predictors. Variable selection was conducted using the “VSURF” package for R, such that models were constructed using only the predictors that meaningfully contributed to the model (i.e., the “interpretation” step). Procedural details on how the VSURF R package selects predictors can be found in papers published by Genuer et al. (28, 29). Each model revealed the total amount of variance explained by the model, and the specific variables that emerged as predictors. We conducted 21 Random Forests models on each dataset with satisfaction as the dependent variable (i.e., 7 predicting baseline satisfaction, 7 predicting follow-up satisfaction, and 7 predicting change in satisfaction). Similarly, we conducted 21 Random Forests models on each dataset that contained commitment (i.e., our secondary dependent variable), for a total of 42 Random Forests models (maximum) per dataset. Results across the 43 datasets were then combined using random-effects meta-analysis. Results for each individual dataset can be found at https://osf.io/4pbfh/.

Meta-Analysis.

Each of the 42 models was examined as a separate random-effects meta-analysis; the 21 satisfaction meta-analyses each contained k = 43 effect sizes, and the 21 commitment meta-analyses each contained k = 31 effect sizes. We performed the basic analyses using comprehensive meta-analysis (30). To calculate each of the effect sizes, we transformed the “% variance accounted for” outcome of the Random Forest model into effect size r (by taking the square root); we then administered the Fisher zr transformation, and we used N-3 as the inverse variance weight (31, 32), where N equals the number of observations used in the Random Forests analysis. We transformed the outcomes of the meta-analyses back to percent variance accounted in Results (by squaring the values). The meta-analytic data files for satisfaction and commitment can be found at https://osf.io/v5e34/.

Moderation Analyses

We examined 12 possible meta-analytic moderators. Ten were features of the datasets: Study length, length between time points, number of time points, average relationship length of the sample, average age of the sample, the year data collection began, country, publication status (≥1 publication vs. not previously published), sample type (community vs. college student), and relationship status (dating vs. married). We also examined two features that were specific to each meta-analytic datum: Number of predictors used in the Random Forests model and number of predictors selected in the final model by VSURF. We used David Wilson’s SPSS macros (http://mason.gmu.edu/∼dwilsonb/ma.html) to perform the moderator analyses (i.e., ANOVA for country, regression for the other 11 moderators).

Results

Primary Meta-Analytic Results.

For baseline satisfaction, actor-reported individual variables (19%) were approximately four times as powerful as partner-reported individual variables (5%), and combining actor and partner individual variables (21%) added no predictive power beyond actor individual variables alone (Fig. 2). Actor-reported relationship variables predicted baseline satisfaction quite powerfully (45%), much more so than partner-reported relationship variables (15%). Combining actor- and partner-reported relationship variables (46%), and combining all individual and relationship variables (44%) added no predictive power beyond actor-reported relationship variables alone. In essence, these findings revealed that any variance in satisfaction explained by information about actor-reported individual differences, partner-reported individual differences, and partner-reported relationship-specific variables could be explained by information about the actor’s relationship-specific variables.

Fig. 2.

Meta-analytic results predicting relationship satisfaction. Meta-analytic effect sizes (and 95% CIs) from k = 43 datasets predicting satisfaction at baseline, at follow-up, and over time. The dependent measure is the percentage of variance accounted for in the Random Forests model that used the set of predictors indicated on the x axis.

When predicting follow-up satisfaction, the pattern of findings was similar, although not surprisingly, all estimates were smaller. Analyses predicting change in satisfaction were generally poor. No analyses accounted for more than 5% of the variance, and the confidence intervals for all estimates overlapped substantially. Self-report variables may be ill-equipped to reliably predict future changes in satisfaction, at least as operationalized here (typically over a span of 1 to 2 y) (SI Appendix).

Results for commitment were generally smaller across models (the average estimate was 3% smaller), but the pattern of findings mirrored those of satisfaction (Fig. 3). Actor-reported variables were at least twice as powerful as partner-reported variables, partner variables did not contribute beyond actor variables alone, individual variables did not contribute beyond relationship variables alone, and change in commitment was generally unpredictable.

Fig. 3.

Meta-analytic results predicting relationship commitment. Meta-analytic effect sizes (and 95% CIs) from k = 31 datasets predicting commitment at baseline, at follow-up, and over time. The dependent measure is the percentage of variance accounted for in the Random Forests model that used the set of predictors indicated on the x axis.

Meta-Analytic Moderators.

Each of the 12 moderators was examined across each of the 21 meta-analytic models for satisfaction and the 21 meta-analytic models for commitment [12 × (21 + 21) = 504 total tests] (SI Appendix, Tables S6 and S7). We only interpreted a moderator substantively if 4 or more of a set of 21 tests achieved significance: The binomial probability of at least 4 of 21 tests achieving significance under the null is P = 0.019 (33).

Three of the 12 moderators exhibited meaningful effects. Effects were generally larger for 1) baseline and follow-up satisfaction in datasets in which the couples were older, and 2) baseline commitment in datasets that had smaller lags between time points. Furthermore, individual difference variables performed better for studies that were conducted relatively recently. None of the moderators affected our (in)ability to reliably predict change in satisfaction or commitment. See SI Appendix for details.

Predictor Restriction Effects.

To what extent are the current results dependent on which variables are removed or retained as predictors? In total, we conducted three versions of the current analyses: A version in which no predictors were excluded except for satisfaction and commitment (“none”; i.e., our preregistered analysis plan); a version in which trust, intimacy, love, and passion were removed as potential predictors (“moderate”); and a version in which eight more variables were removed as suggested by a reviewer (affection, appreciation, conflict, empathy, investment, perceived partner responsiveness, sacrifice motives, and sexual satisfaction; “stringent”). The moderate version is presented above and the two alternative versions are presented in SI Appendix. The relative performance of all three analytic strategies is depicted in Fig. 4.

Fig. 4.

Effects of predictor restriction on meta-analytic results. Note: Blue bars and upper 95% CIs indicate percentage of variance accounted for by actor-reported relationship variables alone. Purple bars indicate the additional percentage of variance explained by the addition of partner-reported relationship variables. Red bars indicate the additional percentage of variance explained by the addition of actor-reported individual differences, partner-reported individual differences, and partner-reported relationship variables. All analyses are averaged across commitment and satisfaction meta-analytic effect sizes.

In Fig. 4, the blue bars indicate the variance accounted for by actor-reported relationship variables at baseline (Left) and follow-up (Right), averaged across the satisfaction and commitment analyses. This figure addresses two key questions: Do models that include partner- and actor-reported relationship variables explain more variance than actor-reported relationship variables alone (stacked purple bars), and do models that include all actor- and partner-reported individual difference and relationship variables explain more variance than models including actor-reported relationship variables alone (Fig. 4, stacked red bars)? The answer in both cases is: Not by much. The total amount of variance explained declines as more potential predictors are excluded from the analyses. However, the individual difference and partner-reported variables consistently explain only an additional 0.0 to 1.9% of the variance at baseline and 0.9 to 3.5% of the variance at follow-up. In other words, regardless of which actor-reported relationship variables are retained or removed, individual differences and partner-reports collectively explain very little additional variance in relationship quality.

Finally, relationship quality change again proved difficult to predict. The ability to predict change was similar regardless of whether the low (mean = 2.4%), moderate (mean ≤ 2.5%), or severe (mean = ≤ 2.2%) restriction strategy was implemented.

Predictive Success of Specific Constructs.

We also compiled and categorized the success of specific predictors. Constructs were sorted according to their prediction success rates: The number of measures of the construct that emerged as a contributing predictor for at least one of the three time points (baseline, follow-up, or change over time), divided by the number of measures of the construct that were tested. The results for the most commonly measured constructs—those that were measured at least 10 times across datasets—are presented in Table 2 (relationship predictors) and Table 3 (individual predictors).

Table 2.

Success rates of the most commonly measured relationship-specific constructs across datasets

Construct	No. of predictors tested		Percent of actor versions successful		Percent of partner versions successful		Overall success rate, %
	Predicting satisfaction	Predicting commitment	Predicting satisfaction, %	Predicting commitment, %	Predicting satisfaction, %	Predicting commitment, %
Perceived partner commitment	10	10	90	70	100	80	85
Intimacy	12	9	92	92	67	67	81
Appreciation	10	10	90	80	60	60	72
Love	17	17	88	53	76	65	71
Sexual satisfaction	20	13	90	75	54	54	71
Perceived partner satisfaction	11	9	91	64	78	44	70
Conflict	29	28	90	79	57	50	69
Perceived partner responsiveness	14	13	93	57	69	54	69
Trust	15	15	87	60	73	53	68
Investment	13	13	77	62	92	38	67
Support general	12	9	67	42	89	67	64
Capitalization	16	10	81	62	40	30	58
Normative attachment	13	13	69	38	69	54	58
Relationship length	54	41	59	67	44	56	57
Passion	14	13	64	50	54	46	54
Alternatives	12	12	58	33	67	50	52
Sexual frequency	11	8	73	36	25	50	47
Inclusion of the other in the self	24	23	54	33	65	35	47
Affection	10	7	50	50	29	43	44
Empathy	11	11	45	36	45	45	43
Intimate partner violence	26	17	27	62	47	35	43
Conflict strategies	23	15	52	30	27	27	36
Power	13	13	31	31	31	23	29
Relationship status	27	21	26	22	38	29	28
Cohabiting	15	14	27	20	29	36	28
Sacrifice motives	22	22	18	18	14	14	16
Children	32	23	16	6	4	13	10

Note: Success rate percentages can be interpreted as the strength of the variable relative to the other variables of this class, but it does not have any independent meaning or effect size. Random Forests do not specify the size or direction of the effect; only that the variable meaningfully contributes to the total variance explained in a given model. Some studies included multiple measures of the same construct, and thus the number of predictors tested can be higher than the total number of datasets. Boldfaced rows correspond to four constructs excluded from the primary models reported in the main text, because they are debatably indicators (not predictors) of relationship quality (1). The values for these four constructs derive from alternative models reported in SI Appendix.

Table 3.

Success rates of the most commonly measured individual difference constructs across datasets

Construct	No. of predictors tested		Percent of actor versions successful		Percent of partner versions successful		Overall success rate, %
	Predicting satisfaction	Predicting commitment	Predicting satisfaction, %	Predicting commitment, %	Predicting satisfaction, %	Predicting commitment, %
Satisfaction with life	12	12	100	83	92	75	88
Depression	28	18	82	68	72	72	74
Negative affect	10	3	90	70	33	67	73
Anxious attachment	38	29	71	74	62	76	71
Avoidant attachment	34	25	71	65	80	68	70
Age	37	25	59	70	72	72	68
Anxiety	11	8	73	82	50	50	66
Self-esteem	16	15	56	50	67	60	58
Agreeableness	20	18	50	60	50	56	54
Positive affect	17	10	53	59	40	60	54
Psychological well-being	19	9	53	53	44	44	50
Religiosity	16	16	38	44	69	44	48
Stress	34	27	38	50	59	41	47
Conscientiousness	19	17	47	26	53	47	43
Income	26	21	46	50	43	29	43
Neuroticism	20	18	65	40	33	22	41
Openness	20	18	20	40	44	44	37
Relationship beliefs	19	19	37	32	53	26	37
Empathy	18	13	28	22	46	38	32
Sexism	21	21	38	24	29	38	32
Health	30	24	40	27	29	29	31
Extraversion	20	18	40	30	28	11	28
Alcohol use	17	14	18	24	43	29	27
Family history	12	12	17	25	42	17	25
Political orientation	10	10	20	20	30	30	25
Education	36	24	22	19	29	25	23
Employed	18	16	33	17	12	25	22
Aggression	13	13	15	38	0	31	21
Race/ethnicity	54	46	20	22	15	17	19
Gender	31	25	13	16	24	20	18
Own traits	35	35	9	20	23	17	17
Religious affiliation	15	14	20	20	14	7	16
Parents' relationship	13	13	8	15	31	0	13
Ideal standards	39	39	10	3	18	8	10

See legend to Table 1.

The most reliable (top five) relationship variables were perceived partner commitment (e.g., “My partner wants our relationship to last forever”), appreciation (e.g., “I feel very lucky to have my partner in my life”), sexual satisfaction (e.g., “How satisfied are you with the quality of your sex life?”), perceived partner satisfaction (e.g., “Our relationship makes my partner very happy”), and conflict (e.g., “How often do you have fights with your partner?”). Many of these successful predictors have been emphasized by interdependence theory and related models [e.g., the interpersonal process model (34), the investment model (35), communal and exchange perspectives (36)], although most theories are not specific enough to generate hypotheses about which relationship variables should function as better predictors than others. Relatively objective relationship variables (e.g., cohabiting status, dating versus married relationship status, having children) generally mattered little, with the exception of relationship length. Finally, the predictors trust, intimacy, love, and passion generally performed quite well in the SI Appendix analyses that included them as predictors (see boldface rows in Table 2).

The most reliable individual difference variables were satisfaction with life (e.g., “The conditions of my life are excellent”), negative affect (e.g., “distressed,” “irritable”), depression (e.g., “feelings of hopelessness”), attachment anxiety (e.g., “I worry a lot about my relationships with others”), and attachment avoidance (e.g., “I prefer not to be too close to romantic partners”). Attachment theory (37) was well-supported in that its two central individual difference constructs were the fourth and fifth most robust predictors. Variables from personality psychology (agreeableness, conscientiousness) and clinical psychology (negative affect, positive affect, depression, anxiety) also proved relevant; these results are consistent with a large body of research on the strong, likely bidirectional connection between relationship quality and well-being (38). Demographic variables, such as sex/gender, race/ethnicity, and education mattered little.

Discussion

How predictable is relationship quality, and which variables predict it best? This project aimed to answer these questions by applying machine-learning techniques to 43 datasets consisting of 11,196 couples. Results revealed that variables capturing one’s own perceptions of the relationship (e.g., conflict, affection) predicted up to 45% of the variance in relationship quality at the beginning of each study and up to 18% of the variance in relationship quality at the end of each study. Individual differences—variables capturing features of the self, such as neuroticism, age, or gender—predicted a smaller but still meaningful amount of variance: Up to 21% at baseline and up to 12% at follow-up. Furthermore, individual differences did not predict relationship quality above relationship-specific predictors alone, partner-reports did not predict relationship quality beyond actor-reports alone, and relationship-quality change was largely unpredictable. That is, our results suggest that if Amir and Alex each complete many questionnaires about themselves and their relationship, all of the predictable variances in their relationship quality will be explained solely by their own perceptions of that relationship. Amir’s reports about his own traits and other characteristics, Alex’s reports about her characteristics, and Alex’s perceptions of the relationship will not explain any additional variance in Amir’s relationship quality. Furthermore, changes in Amir’s relationship quality over subsequent months or years are unlikely to be predictable by any of these self-report measures.

Explaining the Relative Success of the Models.

The finding that relationship-specific variables are more predictive of relationship outcomes than individual difference variables is consistent with existing meta-analyses. In reviews of marital (12) and dating relationships (13), relationship-specific variables are strong predictors of divorce and nonmarital break-ups, respectively, whereas individual difference variables have lower predictive utility. However, meta-analyses are broadly limited to the effects already published in existing literature and tend to reflect the publication biases of that literature (see ref. 39 for discussion). In particular, relationship variables may emerge as stronger predictors than individual differences across published studies because some prominent relationship theories [e.g., interdependence theory (40)] tend to emphasize dyadic and contextual features over stable individual differences. This project addresses this limitation by conducting new, preregistered analyses on raw datasets, such that every measured variable had a similar chance to contribute to the models.

Why did the addition of individual differences and partner reports to the models fail to improve upon the predictive power of actor-reported relationship variables alone? Had these variables functioned as robust and consistent moderators of actor relationship-specific variables (e.g., individual-difference × relationship-specific variable interactions; actor × partner interactions), the addition of individual differences and partner-reported variables to the Random Forest models should have accounted for more variance (24). One possibility is that the actor-reported relationship variables are redundant with each other (and with the satisfaction/commitment-dependent measures), and their collective inclusion leads to model misspecification. This concern surely seems intuitive for scholars familiar with typical problems caused by collinearity in multiple regression contexts, in which the simultaneous inclusion of many correlated predictors causes estimates to become erratic. Critically, Random Forests models are specifically designed to overcome this issue through recursive partitioning: The iterative sampling of random sets of participants and predictors (24, 25). In light of the way Random Forests models work, then it makes sense that our additional analyses that relaxed and restricted the specific predictors available did not strongly affect these conclusions.

Another plausible, more theoretically interesting possibility is that individual differences and partner reports exert their effects not via moderation but via mediation. That is, individual differences and partner effects are important, but they exert their influence on relationship quality indirectly, via interpersonal processes that are adequately captured by the actor-reported relationship variables. The “all predictors” models do not predict more variance than the “actor-reported relationship” models because actor-reported relationship variables fully mediate the effects of the other predictors (Fig. 5). To better understand how individual differences might shape relationship dynamics and in turn relationship quality, research is needed on the early stages of relationships when these relationship-specific dynamics first emerge (41).

Fig. 5.

Mediational pathway implied by current findings. This figure depicts the mediational model implied by the equivalent predictive power of the “all predictors” vs. “actor relationship” models in Figs. 2 and 3. That is, any effects of self-reported individual differences or partner-reported relationship variables on relationship quality are likely mediated by the actor-reported relationship variables. Individual-difference × relationship variable interactions and actor × partner interactions are not depicted because they are likely quite small. Other constructs in Fig. 1 (e.g., broader contextual forces) are not depicted because they were not examined in this study.

Also notable was the underperformance of the models predicting change in relationship quality. In other words, any nascent signal of whether a relationship is going to become better or worse over time does not seem to be detectable in self-reported variables at baseline. Surely, change in relationship quality can be explained by baseline variables in conjunction with time-varying predictors [e.g., stressful life events, the transition to parenthood (42, 43)]. However, models that attempt to account for future change entirely from contemporaneously assessed self-report variables may not prove robust. These results are consistent with another recent large collaboration showing that life trajectories are generally difficult to predict, even with complex machine-learning methods (44).

Limitations and Future Directions.

Why did demographic variables underperform as predictors of relationship quality? One possibility is that, reflecting a common limitation of psychological samples more broadly (45), the present samples may have been overly affluent, White, and college-educated, and were thus too homogeneous to reveal the predictive power of variables such as ethnicity and education. This possibility seems unlikely, however, because more than half of the couples tested (n = 6,298) were recruited as part of the Supporting Healthy Marriages Project (Dataset 38), which intentionally oversampled low-income couples. This sample varied considerably on ethnicity (both spouses were White in 21% of couples), education (at least one partner had a college degree in 27% of couples), and income (42% of couples reported income levels below the poverty line). Yet, the pattern of results from this sample mirrored the results of the other 42 datasets (SI Appendix, Fig. S3).

All of the current datasets were sampled from Western countries (the United States, Canada, Switzerland, New Zealand, The Netherlands, and Israel). Future work should examine whether the current effects generalize beyond the Western context. Our conclusions are also specific to baseline self-report predictor variables; of the 1,149 relationship-specific variables tested in this project, 99.4% were explicit self-report rating scales (and similar numerical response scales) rather than independent observations that directly captured participants’ real-time behavior (i.e., variables directly assessing the interpersonal behavior arrow in Fig. 1). Future work should explicitly solicit observational and other nonself-report data and compare their predictive utility to self-reports. These results similarly do not apply to nonself-report measures of contextual variables, such as income and debt (e.g., which could be measured instead via tax returns), stress (e.g., diurnal cortisol patterns, neighborhood-level crime statistics), or the role of social networks (e.g., informant reports). In this project, such variables were measured with self-reports—for example, self-reported income, stress, or network support—and were thus categorized as individual differences. However, drawing on evidence that context can matter a great deal for relationship quality (11), another good future direction would be to test contextual variables as their own category of predictors, ideally using nonself-report measures. Finally, this collaboration included more datasets from the laboratories of psychologists than sociologists, communications scholars, or family studies scholars; datasets in these disciplines may commonly include variables that reveal different conclusions.

This study—which represents the largest and most integrative data analytic effort in the study of romantic relationships—suggests the following four constraints on future theories and models of relationship dynamics. First, constructs self-reported by the partner are unlikely to predict the actor’s relationship quality beyond the actor’s own (contemporaneously assessed) individual-difference and relationship-specific variables. Second, individual differences are unlikely to predict relationship quality beyond (contemporaneously assessed) relationship-specific variables. Third, change in relationship quality was not predictable from baseline self-report measures, so change is likely a function of external context, behavioral processes, or other factors that are themselves changing over time. Fourth, models should posit larger effect sizes for the variables that fared well (vs. poorly) in Tables 2 and 3, regardless of whether those models emphasize main effects or interactions. Of course, the occasional study may report findings that run contrary to these constraints. Our collaborative effort does not necessarily overturn such findings, but rather suggests that scholars may want to raise the standard for attaining high confidence in them (e.g., await the independent replication of the finding in datasets that are notably distinct from those we meta-analyze here).

Conclusion

From a public interest standpoint, this study provides provisional answers to the perennial question “What predicts how satisfied and committed I will be with my relationship partner?” Experiencing negative affect, depression, or insecure attachment are surely relationship risk factors. But if people nevertheless manage to establish a relationship characterized by appreciation, sexual satisfaction, and a lack of conflict—and they perceive their partner to be committed and responsive—those individual risk factors may matter little. That is, relationship quality is predictable from a variety of constructs, but some matter more than others, and the most proximal predictors are features that characterize a person’s perception of the relationship itself.