Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2025 Dec 17;94(5):819–833. doi: 10.1111/jopy.70042

Causal Effect of Psychopathy on Partner Well‐Being: A National Longitudinal Study

Aaron Hissey 1,✉, Hedwig Eisenbarth 1, Matt Hammond 1, Chris G Sibley 2, Joseph Bulbulia 1
PMCID: PMC13540337  PMID: 41409007

ABSTRACT

Objective

Psychopathy is associated with harmful behaviors in romantic relationships, but its causal effect on partner well‐being remains elusive. Across two studies, we apply robust observational methods for causal inference with national‐scale dyadic data to examine how psychopathic personality and its facets affect partner well‐being over time.

Method

In Study 1 (three waves; N = 1012), we examined how increasing global psychopathy and its facets within individuals affects their partner's well‐being one year later. In Study 2 (six waves; N = 1832), we investigated how increasing global psychopathy and its facets over multiple years would affect partner well‐being.

Results

Study 1 indicates that psychopathic personality increases partners' perceived relationship conflict and anxiety in the short term but also bolsters partner self‐esteem. Study 2 reveals that, over the longer term, psychopathic personality intensifies partners' perceived relationship conflict and depression and reduces partner self‐esteem, relationship satisfaction, and personal well‐being.

Conclusions

Taken together, it is the convergence of psychopathic traits, rather than individual facets, that cause greater harm to partners, with the magnitude of harm intensifying over time. This research demonstrates the power of combining dyadic panel data with causal inference methods to address questions of psychological interest inaccessible to experiments.

Keywords: causal APIM, causal inference, psychological well‐being, psychopathy, romantic couples

1. Introduction

Psychopathic traits are associated with destructive behaviors that may harm romantic partner well‐being. However, psychopathy research has traditionally focused on clinical and forensic populations (e.g., Hare 2006) or severe outcomes such as intimate partner aggression and violence (e.g., Robertson et al. 2020). As a result, little is known about how psychopathic traits are linked with diverse relational behaviors and outcomes in the general population. Additionally, the degree to which psychopathy causally affects partner well‐being is difficult to ascertain. Notably, experimental manipulations of psychopathy are unfeasible and unethical, ruling out randomized trials. Thus, not only is it unclear how much of the association between psychopathic traits and partner well‐being in the general population is causal, but also how we might address this question.

Research suggests that romantic partners substantially impact one another's well‐being. Partners provide each other with support, buffer stress (Cohen and McKay 1984), and contribute to each other's sense of purpose and meaning (Pfund and Hill 2022). However, romantic partnership may also invite conflict, increase stress, and diminish mental and physical health (Kiecolt‐Glaser and Wilson 2017). Personality traits have been shown to predict partner relational outcomes, such as lower neuroticism and higher agreeableness, conscientiousness, and extraversion being linked to greater relationship satisfaction (Malouff et al. 2010). While psychopathy is related to broad personality traits such as the Big Five, it reflects extreme and maladaptive variants that can uniquely predict interpersonal and externalizing problems (Phillips et al. 2025). Moreover, psychopathy has been conceptualized as an emergent interpersonal syndrome (Lilienfeld et al. 2019), a configuration of interpersonally malignant personality subdimensions that interact in complex ways. Thus, it is likely that people with higher psychopathic traits negatively impact partner well‐being, which could go beyond effects found for general personality frameworks.

The causal effects of psychopathy on partner well‐being might be substantial. Psychopathy, as measured by the Elemental Psychopathy Assessment (Lynam et al. 2011), encompasses the facets of Antagonism, Emotional Stability, Disinhibition, and Narcissism. Higher Antagonism is characterized by coldness, distrust, and callousness, whereas higher Emotional Stability is marked by greater self‐contentment and immunity to stress. Higher Disinhibition is characterized by rashness, impersistence, and rebelliousness, while higher Narcissism is characterized by anger, dominance, and arrogance. In this framework, Narcissism reflects a set of arrogance and self‐assurance traits, similar to, but distinct from, broader conceptualizations of narcissism in the psychological literature (Phillips et al. 2025). Research examining links between psychopathy and romantic functioning shows that psychopathic traits are correlated with destructive relational behaviors, including heightened infidelity (Khan et al. 2017) and physical and psychological partner abuse (Carton and Egan 2017). In addition, psychopathy is associated with heightened partner psychological distress (Forth et al. 2022), increased partner aggression (Webster et al. 2016), and lower partner relationship satisfaction (Smith et al. 2014; Weiss et al. 2018). This evidence suggests that elevated psychopathic traits may be damaging to the well‐being of romantic partners.

However, a fundamental barrier to progress in psychopathy relationship research is establishing causality. Encouragingly, robust workflows have been developed in epidemiology during the past several decades that may allow investigators to obtain consistent causal effect estimates in settings where investigators are restricted to observational data (Hernán and Robins 2020). Such approaches have been further enhanced by recent progress in the machine learning literature, reducing reliance on statistical modeling assumptions and increasing estimation precision (Van der Laan and Rose 2011). Although robust methods for causal inference have come to dominate certain areas of social science (Morgan and Winship 2014), they are rarely employed in psychological science (Bulbulia 2024b). Nevertheless, rapid progress is possible because robust analytic approaches and software have already been developed (Williams and Díaz 2023).

A second gap in understanding arises from limited measurements of partner well‐being. The few studies that have investigated associations between psychopathy and partner well‐being have focused on a narrow set of well‐being measures. Given the range of destructive behaviors associated with psychopathy, it is likely to affect multiple dimensions of partner well‐being. However, we are unaware of any study reporting, for example, associations between psychopathy and partner self‐esteem or life satisfaction, fundamental dimensions of well‐being in relationships. Here too, progress is readily within reach: investigators need only model a broader array of outcomes to capture the missing dimensionality (VanderWeele et al.'s 2020).

A third gap in understanding arises from a tendency to investigate psychopathy as a single global trait (e.g., Smith et al. 2014; Weiss et al. 2018), despite evidence that psychopathy is comprised of facets. For instance, the low anxiety and stress resilience traits that comprise psychopathy show little connection to externalizing behaviors (Collison et al. 2016), while anger, social dominance, rashness, and coldness traits are differentially related to antisocial behavior, drug use, and criminal activity (Seigfried‐Spellar et al. 2017). It is therefore important to consider which facets of psychopathy affect partner well‐being, and whether these facets operate differently.

Finally, research examining the effects of psychopathic traits on partners would benefit from integrating causal methods and richer measures with dyadic data. Many previous studies have collected data from only one member of a couple rather than both members (e.g., Forth et al. 2022), providing an incomplete view of the relationship. However, romantic couples strongly influence one another's outcomes, underscoring a need to account for their interdependent characteristics and interactions (Wheatley et al. 2024). Taken together, shortcomings in present approaches offer considerable opportunity to clarify how distinct psychopathic traits causally affect multiple dimensions of partner well‐being within dyadic relationships.

2. The Current Study

Here, we integrate robust causal inference methods with large, dyadic, longitudinal, national panel data to estimate the magnitudes of causal effects of increasing psychopathic traits on multiple dimensions of partner well‐being. We pose two well‐defined causal questions: (1) “On average, how would an intervention that increases psychopathy by one measurement unit affect partner well‐being outcomes in the following year, compared to no intervention?” and (2) “On average, how would an intervention that increases psychopathy by one measurement unit over four years affect partner well‐being outcomes in the fifth year, compared to no intervention?” The target population consists of New Zealand couples over 18 years old.

In Study 1, we examined how increasing psychopathy in individuals affects their partner's well‐being one year later. To account for confounding and rule out reverse causation, we used three waves of dyadic data. In the first wave (“baseline”), we measured psychopathic traits, demographic factors, and well‐being for both members of each couple. In the second wave (“exposure”), we modeled interventions that either did not alter individuals' psychopathy scores or increased them by one unit above their observed values. A one‐unit increase on the 7‐point psychopathy scales represents a meaningful shift in trait endorsement. For example, a one‐unit increase in Antagonism reflects greater callousness, increased Emotional Stability reflects greater self‐contentment, increased Disinhibition reflects more impulsive behavior, and increased Narcissism reflects greater arrogance and social dominance. In the third wave (“end of study”), we measured partner well‐being indicators. To estimate causal effects, we compared the expected average outcomes between models with and without increased psychopathy. We hypothesized that increasing Antagonism, Disinhibition, and Narcissism traits in an individual would amplify partner psychological distress and reported relationship conflict, and diminish partner subjective well‐being, self‐esteem, life satisfaction, and relationship satisfaction. Study 1 was pre‐registered (https://osf.io/q89se/overview).

In Study 2, we investigated how increasing psychopathy over multiple years would affect partner well‐being. Using six waves of dyadic data, we measured baseline confounding variables using data from the first wave. Over the next four waves (waves 2–5), we modeled interventions that did not alter individuals' psychopathy scores or increased them by one unit above their observed values. At the end of the study (wave 6), we measured partner well‐being and estimated causal effects by comparing the sequential interventions from a baseline model in which no intervention occurred. Study 2 was not pre‐registered. It developed from enhancements to our workflows in modeling multiple years of psychopathy exposure and in handling attrition from the baseline to first exposure wave.

3. Study 1

3.1. Materials and Methods

We used panel data from times 10, 11, and 12 of the New Zealand Attitudes and Values Study (NZAVS). The NZAVS is a national panel study exploring the social attitudes and values, personality characteristics, and health outcomes of New Zealanders. NZAVS time 10 (the baseline wave in both Study 1 and Study 2) included N = 47,948 individuals, approximately 1.5% of New Zealand's adult population. Ethics approval is reviewed every three years by the University of Auckland Human Participants Ethics Committee (#22576). Time 10 data was collected across years 2018/2019, Time 11 data was collected across years 2019/2020, and Time 12 data was collected across years 2020/2021. Using a rule set defined in Sibley (2021), romantic couples were primarily identified by finding people that lived together and had the same date of marriage.

3.1.1. Participants

Participants were identified as romantic couples who were in the same relationship at Time 10 (baseline wave) and Time 11 (intervention wave). Participants were required to have reported full data for the psychopathy scale at Times 10 and 11. Missing data were allowed for all other variables, including demographic data at baseline and well‐being variables at Time 12 (outcome wave). We imputed missing baseline data using single imputation, adding a separate column for each indicator denoting whether a missing value was imputed, following Hoffman et al. (2023). This approach allowed our semi‐parametric machine learning to incorporate information about missingness when estimating causal effects (Hoffman et al. 2023).

If either partner was lost to follow up, or if the couple separated, both members of the dyadic pair were treated as censored. Of 1012 eligible couples at baseline, 774 couples had both partners observed at the final wave (76.5% retention). Final‐wave outcomes were missing for 238 couples (23.5%). Attrition was amplified because if one partner dropped out or the couple separated before the outcome assessment, both partners were treated as censored from that point. However, by modeling censoring within our statistical models, we obtain valid inference for the full sample population at baseline under the assumption that all covariates within the model account for attrition and separation (including baseline and subsequent psychopathy). Baseline covariates were effectively complete, so missingness reflects attrition or relationship dissolution rather than item nonresponse. In estimation, we treat these cases as dyad‐level censoring and adjust using inverse probability of censoring weights within the Longitudinal Modified Treatment Policy (LMTP) framework, targeting the “no breakup/no attrition” estimand defined for the baseline cohort. To handle statistical dependencies from within couples, we estimated effects adjusting for clustering using the clustering “id” indicator in lmtp. We batched processed the models using the margot package in R (Bulbulia 2024a).

Our selection criteria resulted in 1012 participants, within 506 couples. Of these participants, 492 (49%) identified as male. Age varied between 22 and 90 years (M = 53, SD = 11), and relationship duration ranged from less than one year to 61 years (M = 23, SD = 13). In total, 425 of the 506 couples (84%) were married. Participants self‐identified as Pākehā/New Zealand European (n = 874; 87%), Māori (n = 73; 7%), Asian (n = 37; 4%) and Pacific peoples (n = 24; 2%).

3.1.2. Measures

3.1.2.1. Psychopathy

The NZAVS does not include an explicit measure of psychopathy. Psychopathic traits were assessed using a 24‐item NZAVS psychopathy scale that serves as a proxy for the Elemental Psychopathy Assessment—Short Form (EPA‐SF; Lynam et al. 2013) and includes four six‐item subscales representing Antagonism, Emotional Stability, Disinhibition, and Narcissism. Ratings for all items (e.g., “I am not interested in other people's problems”) were given on a seven‐point Likert scale, ranging from 1 (“Strongly disagree” or “Very inaccurate”) to 7 (“Strongly agree” or “Very accurate”). A Global Psychopathy score was calculated by averaging all items, while facet scores were calculated by averaging items within each subscale. Reliabilities for the scales, based on the full NZAVS sample and as indicated by McDonald's omega, were ω = 0.77 (Antagonism), 0.79 (Emotional Stability), 0.68 (Disinhibition), 0.75 (Narcissism), and 0.81 (total psychopathy). Further details regarding the NZAVS psychopathy measure can be found in Tables B1–B5 in the Supporting Information.

3.1.2.2. Personal Well‐Being

Personal well‐being was assessed using the following four items from the Personal Well‐being Index (PWI; Cummins et al. 2003): “Please rate your level of satisfaction with the following aspects of your life and New Zealand… (1) Your health (2) Your standard of living (3) Your future security (4) Your personal relationships”. Participants rated items on a 10‐point scale, ranging from 1 (“Completely Dissatisfied”) to 10 (“Completely Satisfied”). A global personal well‐being score was calculated by averaging all items. Scale reliability for the full NZAVS sample was ω = 0.73.

3.1.2.3. Psychological Distress

The six‐item Kessler‐6 (Kessler et al. 2010) was used to assess psychological distress. The six items were: “During the last 30 days, how often did… (1) you feel hopeless (2) you feel so depressed that nothing could cheer you up (3) you feel restless or fidgety (4) you feel that everything was an effort (5) you feel worthless (6) you feel nervous”. Participants rated items on a five‐point Likert scale, ranging from 0 (“None of the time”) to 4 (“All of the time”). A global distress score was calculated by averaging all items, an Anxiety score was calculated by averaging items 3 and 6, and a Depression score was calculated by averaging items 1, 2, 4, and 5. Scale reliability for the full NZAVS sample was ω = 0.85.

3.1.2.4. Self‐Esteem

Self‐esteem was assessed by using three items from the Rosenberg Self‐esteem Scale (Rosenberg 1965). Participants rated items (“I on the whole am satisfied with myself”, “I take a positive attitude toward myself”, and “I am inclined to feel that I am a failure”) on a seven‐point Likert scale, ranging from 1 (“Very inaccurate”) to 7 (“Very accurate”). A global self‐esteem score was calculated by averaging all items. Scale reliability for the full NZAVS sample was ω = 0.81.

3.1.2.5. Life Satisfaction

Satisfaction with life was assessed by using two items from the Satisfaction with Life Scale (Diener et al. 1985). Participants rated items (“I am satisfied with my life” and “In most ways my life is close to ideal”) on a seven‐point Likert scale, ranging from 1 (“Strongly disagree”) to 7 (“Strongly agree”). A global life satisfaction score was calculated by averaging both items. Scale reliability for the full NZAVS sample was ω = 0.78.

3.1.2.6. Relationship Conflict

Relationship conflict was based on an item from the NZAVS (“To what extent do you experience conflict or disagreement with your partner?”). Participants responded on a seven‐point Likert scale from 1 (“No conflict at all”) to 7 (“A great deal of conflict”).

3.1.2.7. Relationship Satisfaction

Relationship satisfaction was based on an item from the NZAVS (“How satisfied are you with your relationship?”). Participants responded on a seven‐point Likert scale from 1 (“Not satisfied”) to 7 (“Very satisfied”).

3.1.3. Statistical Analysis

We propose a causal Actor‐Partner Interdependence Model (causal APIM), which differs from statistical APIMs popular in dyadic studies (see Kashy and Kenny 1999), by explicitly specifying and targeting a causal estimand—or magnitude of causal effect under a pre‐specified intervention. Similar to APIMs, we utilize dyadic data and control for both the individual's and their partner's attributes in the baseline wave. Statistical estimation under a causal APIM follows seven main steps: (1) stating the causal quantities of interest; (2) evaluating assumptions under which the targeted causal quantities may be identified from data; (3) stating the population for whom inferences are meant to generalize; (4) defining a statistical estimator to compute the quantities of interest; (5) statistical estimation; (6) sensitivity analysis; and (7) reporting.

The key causal quantities of interest are psychopathic traits and partner well‐being indicators. Using longitudinal modified treatment policies (LMTP) via the lmtp package in R (Díaz et al. 2023; Williams and Díaz 2023), our intervention involved increasing each individual's psychopathy score by one measurement unit (up to the maximum scale value of 7) while maintaining all other variables at their observed values. We contrasted this hypothetical scenario against the observed state where no intervention occurred (i.e., psychopathy scores remained at their actual values). For example, if an individual reported a psychopathy score of 3, we modeled the intervention where that score would instead be 4. This shift was applied universally across all participants regardless of their baseline score, preserving the original variation in the data. We selected a one‐unit shift rather than a standard deviation‐based change for two reasons. First, the contrast assists with interpretability: each estimate reflects the expected change in partner well‐being for each one‐point increase from the original response across the population. Second, such an intervention helps to satisfy the positivity assumption of causal inference, which requires that intervention is possible for everyone subjected to it.

We examined multiple indicators of partner well‐being and facets of psychopathy. Partner well‐being indicators included personal well‐being, psychological distress, anxiety, depression, self‐esteem, life satisfaction, relationship conflict, and relationship satisfaction. For each of the five psychopathy variables (Global Psychopathy, Antagonism, Emotional Stability, Disinhibition, Narcissism), we ran a separate set of intervention models for each well‐being indicator, producing 40 intervention models in total. We ran a parallel set of “null” models that estimated partner outcomes without intervention, producing 40 null models. Under the assumptions of no unmeasured confounding (i.e., that all variables influencing both psychopathy and partner well‐being have been measured), causal consistency and common support (Hernán and Robins 2020), these contrasts provide consistent causal effect estimates of partner well‐being. To address participant attrition, we implemented inverse probability of censoring weights (IPCW) as part of the LMTP estimation procedure, weighting observations by the estimated probability of remaining in the study at each wave. This approach ensures our estimates remain representative of the original sample and accounts for potential selection bias due to dropout.

The study population consists of couples aged 18 years or older who participated in the NZAVS at Time 10 (the baseline wave). We define the causal effect for this baseline cohort as the mean difference in the partner's outcome at the subsequent wave(s) if, counterfactually, at Time 10 the individual's psychopathy was shifted according to our pre‐specified modified treatment policy, versus a reference strategy, with averaging over the Time‐10 covariate distribution of couples. As the censoring intervention fixes breakup and attrition to “off,” the estimated effects pertain to pathways that do not operate through dissolution or loss, which is the estimand aligned. Regarding generalizability, our results generalize (or “transport”) to the cohort population from which this sample was selected and to other populations that resemble this cohort population. Although the New Zealand Attitudes and Values Study is a national probability sample with close representation of the general adult New Zealand population in 2018/2019 (the baseline wave, NZAVS Time 10), we cannot be certain that results generalize to all couples in this population as there is likely some selection bias as individuals agree to enter the study. If psychopathy affects participation of the individual or the individual's partner then, without further adjustment, estimates will likely be biased for the general population. In the absence of further information about the departure between the participating and general populations, we restrict focus to the participating population.

Couple dissolution and loss to follow‐up can truncate observations before the outcome is recorded. Whereas selection into the study is known as left‐censoring bias, ongoing selection through drop‐out and breakups will further distort results, given the select sample of couples at the end of study may differ from the baseline cohort in qualities that affect outcomes. Again, selection from psychopathy may be expected. However, here we are on better footing for adjusting for such bias given that we measure psychopathy along with a rich set of covariates in both partners that may affect dissolution and attrition. We therefore define a dyad‐level censoring indicator that switches on at the first time either member is lost or the couple dissolves, and we target the effect under an intervention that prevents censoring through the outcome wave. We model the censoring mechanism and apply inverse probability of censoring weights within lmtp so that the resulting parameter is interpretable as the effect for the Time‐10 cohort under complete follow‐up (i.e., were no restriction to have occurred). Identification rests on standard conditions adapted to this setting: (i) consistency of outcomes under the specified intervention; (ii) conditional exchangeability for both the exposure and the censoring process given measured baseline covariates (including both partners' histories); (iii) positivity of receiving exposure values under the modified treatment policy and of remaining uncensored given covariates; and (iv) partial interference restricted to members of the same couple.

Statistical estimation was performed using Targeted Minimum Loss‐based Estimation (TMLE), a semi‐parametric estimator (Van der Laan and Rose 2011, 2018). TMLE combines machine learning with traditional statistical models to provide valid estimates and measures of uncertainty. The estimation, performed using the lmtp package in R, involves a two‐step process: first, machine learning algorithms model the relationships between exposures, covariates, and outcomes, producing initial estimates; second, these estimates are refined through an iterative process that incorporates observed data, guided by the efficient influence function, to improve accuracy. lmtp draws on the SuperLearner library, which comprises various machine learning algorithms (Polley et al. 2023). Given the large number of baseline control variables and the relatively small sample size, we used a random forest estimator Ranger package (Wright and Ziegler 2017) and the xgboost estimator (Chen et al. 2023), which are well‐suited to this setting. We also included the expected mean estimator as a sanity check. Both random forests and gradient boosting are non‐parametric estimators distinguished by their resilience against overfitting (Díaz et al. 2023; Van der Laan and Rose 2011; Wager and Athey 2018). Non‐parametric machine learning estimators reduced reliance on restrictive assumptions of parametric models to the functional forms of the variables that were modeled and to their distributions, enhancing the robustness of our causal inferences (for further details see Díaz et al. 2023; Hoffman et al. 2023; Van der Laan and Rose 2011).

To avoid bias in estimation and to recover valid confidence intervals, we evaluated models on data not observed during training. Specifically, we performed five‐fold cross‐validation, partitioning the dataset into five subsets of approximately equal size. In each iteration, we used four subsets to train the model and the remaining subset for testing. Every subset served as the test set exactly once, ensuring that each observation was used for both training and validation. Finally, we averaged the five accuracy estimates to obtain a comprehensive measure of model performance.

Missing data at baseline were imputed using predictive mean matching, implemented in the mice package (Van Buuren 2018). 1.5% of the variables were missing at baseline. To handle missingness from attrition and non‐response in the final wave, we employed inverse probability of censoring weighting. These weights were computed automatically as a feature of non‐parametric estimation in lmtp, allowing consistent causal inferences for outcomes at the end of study (NZAVS Time 12), on the assumption that missingness is random conditional on all covariate and exposure history prior to censoring (Williams and Díaz 2023).

We accounted for interdependence within couples by treating dyads as cluster‐level units, matching each individual to their partner, and clustering couples within dyads. Dependencies arising from this clustering were modeled using the “id” clustering parameter in lmtp (see Williams and Díaz 2023). At baseline (NZAVS Time 10), we controlled for Age, Gender, Ethnicity, Relationship Duration, Education, Deprivation, Church Attendance, Conservative Political Orientation, Urban/Rural status, Parent status, Hours of Work, Hours of Housework, as well as each psychopathy facet and each of the well‐being measures using indicators for these values. For each individual, we included indicators both for that individual and that individual's partner, on the grounds that either might serve as a common cause of that individual's psychopathic trait and that individual's partner's well‐being at the end of study. Interventions were modeled over all individuals, again controlling for both their own and their partners' baseline demographic measures, baseline exposure measures, and baseline outcome measures. Thus, for any unmeasured confounder to explain away an observed association it would need to be orthogonal to these measures.

We refer to our approach as a Causal Actor Partner Interaction Model or causal APIM. While we are unaware of other applications of causal inference in relationship research, our causal APIMs are not particularly novel. The model we report in Study 1 is a modest generalization of VanderWeele et al.'s (2020) three‐wave outcome‐wide design for causal investigations of multi‐dimensional well‐being in individuals. However, rather than estimating effects on individuals, we utilize dyadic data and estimate effects on their partners. The six‐wave model in Study 2 extends this three‐wave design within the ambit of longitudinal modified treatment effects, a well‐defined literature in causal inference (Díaz et al. 2023). lmtp models were estimated using wrappers in the margot package, which also produced the figures and tables (Bulbulia 2024a).

As a sensitivity analysis for unmeasured confounding, we computed E‐values for each causal‐contrast (VanderWeele and Ding 2017). An E‐value calculates the minimum strength of an unmeasured confounder in its association with both the exposure and the outcome to explain away an observed result. E‐values are computed on the risk‐ratio scale.

We report 95% confidence intervals for all causal contrasts along with E‐values, which provide a measure for robustness of findings to unmeasured confounding. Specifically, the E‐value quantifies the minimum strength of association (on the risk ratio scale) that an unmeasured confounder would need to have with both the exposure and the outcome to explain away the observed exposure‐outcome association (Linden et al. 2020; VanderWeele and Ding 2017). For example, if the E‐value were 1.3, it would mean an unmeasured confounder associated with both the exposure and outcome by a risk ratio of 1.3 each (or a 30% increase in risk) could explain away the observed effect. We also report the 95% interval bound of the E‐value closest to 1, the bound at which the finding is not reliable.

We estimated the Average Treatment Effect (ATE), or marginal effect, on partner well‐being resulting from interventions on individual psychopathy. Again, our causal estimand is the difference in expected partner outcomes if everyone in the population were shifted one unit higher on the psychopathy scale compared with their observed scores. We organized the data in wide format, with each row representing an individual. We recorded measurements for both individuals and their partners, linking them with relationship identifiers. Because each person appears as both an “individual” and another person's “partner”, tracking these identifiers enabled us to estimate how interventions on each individual affect their partner's outcomes.

It is worth pausing to provide a more detailed explanation of our causal approach for readers who may be unfamiliar with these methods. Here, we extend traditional partner‐interaction models by estimating how intervening on one member causally affects their partner's well‐being. Unlike conventional Actor Partner Interdependence Models that describe statistical associations within observed data, causal inference requires estimating population‐level contrasts under different interventions. In experiments, randomization ensures these contrasts reflect causal effects by design, not by estimating individual‐level effects but by making treatment groups comparable at the population level. With observational data, however, associations alone can mislead, as we lack randomization and experimental control. Even when statistical control is attempted, the associations track changes over time in the observed population. These associations generally lack causal interpretations. However, when data are structured to mimic an experimental design and three key assumptions hold—(1) conditional exchangeability (treatment is “as if random” after adjusting for measured confounders), (2) positivity (everyone has a non‐zero probability of receiving each treatment), and (3) consistency (measured treatments correspond to the interventions being compared)—causal effects can be consistently estimated using repeated measures data.

Our study involves a population of couples observed over time, where dependencies between partners must be accounted for. We handle this by clustering statistical dependencies at the couple level, ensuring our semi‐parametric estimators (implemented via the lmtp package) adjust for within‐couple correlations while avoiding reliance on correctly specified parametric models – an assumption that cannot be verified. We modeled psychopathy as a one‐unit increase on the 1–7 Likert scale (capped at 7) from each individual's observed score. This shift intervention is more realistic than setting everyone to a fixed level; for example, individuals with low psychopathy are nudged up by one point rather than to one standard deviation above the sample mean. This approach also strengthens satisfaction of the positivity assumption, as the intervention is by definition possible for all participants.

We contrasted partner outcomes under this intervention with partner outcomes under the reference condition where no change was made to observed psychopathy scores. This difference, standardized across outcomes to enable comparison, captures the expected change in partner well‐being under the psychopathy treatment regime versus no treatment. To control for confounding, we included both individual and partner psychopathy measures at baseline (collected in the year prior to the exposure wave). We assume these measures are sufficient to capture all common causes of the exposure and of the outcomes (measured at the end of study, after the final exposure wave). Because causal assumptions cannot be statistically verified, we report sensitivity analyses that quantify the minimal amount of unmeasured confounding that could explain away a result (E‐values). Figure 1 presents a causal directed acyclic graph summarizing our causal assumptions and confounding control strategy. Our semi‐parametric estimators (implemented via the lmtp package) can estimate these causal effects while adjusting for statistical dependencies within couples. They furthermore avoid the assumption of a properly specified statistical model—an assumption that cannot be verified for parametric models (Williams and Díaz 2023).

FIGURE 1.

FIGURE 1

Open in a new tab

Directed acyclic graph indicating the dependencies between the key variables. U = unmeasured confounding; Superscript 0 = NZAVS Time 10 (baseline); Superscript 1 = NZAVS Time 11 (intervention); Superscript 2 = NZAVS Time 12 (end of study, when outcomes were measured). We do not draw an edge from the exposure to the outcomes because the purpose of a causal directed acyclic graph is to assess d‐separation—or that all “back door paths” are closed. To reduce confounding biases, in addition to controlling for demographic variables that might affect the exposure and outcome, or that serve as proxies for such variables, we also control for baseline indicators of all partner outcomes and as well as baseline values of all psychopathy facets and global psychopathy. This strategy follows VanderWeele et al.'s (2021, 2020) three‐wave outcome‐wide design. This identification strategy is robust because for any unmeasured confounder to bias a result, its effects would need to be orthogonal to baseline measures of the exposure and outcomes in both agents and their partners. The dotted arrows indicate potential residual confounding from unmeasured variables whose effects are orthogonal to baseline confounders, including the baseline exposure measurement, and all outcomes measured at baseline. The threat of residual confounding cannot be fully removed when treatment assignment is not random. As we cannot ensure that all confounding has been controlled, we compute and report E‐values as a sensitivity analysis.

3.2. Results

A series of models were conducted to assess the extent to which increases in each of the four psychopathic traits, as well as the combined construct, affect partner well‐being one year later compared with no intervention. As shown in Table 1 and Figure 2, one‐unit increases in Antagonism, Disinhibition, Emotional Stability, and Narcissism did not suggest statistically reliable causal effects on partner well‐being (see Tables A1–A5 in the Supporting Information for full results). For example, a one‐unit increase in Antagonism and an associated increase in callousnesswere related to minimal changes in partner psychological distress and relationship satisfaction. However, a one‐unit gain in Global Psychopathy increased partner anxiety and perceived relationship conflict. Global Psychopathy also increased partner self‐esteem. Therefore, although we did not find support for our hypotheses that individual facets of psychopathy reliably affect partner well‐being, we find that Global Psychopathy is causally related to several dimensions of partner well‐being. Specifically, in the short term, intervening on Global Psychopathy increases an individual's callousness, self‐contentment, impulsivity, and arrogance. This then increases partner perceived relationship conflict and, to a lesser degree, partner anxiety. However, in the short term, we also observe that such an intervention increases partner self‐esteem.

TABLE 1.

The effect of a unit measurement gain in Global Psychopathy and psychopathy facets on indicators of partner well‐being.

Partner well‐being indicator Antagonism Disinhibition Emotional stability Narcissism Global psychopathy
ATE 95% CI ATE 95% CI ATE 95% CI ATE 95% CI ATE 95% CI
Relationship conflict 0.01 −0.07, 0.09 0.06 −0.03, 0.15 −0.00 −0.07, 0.06 0.06 −0.04, 0.16 0.12 0.05, 0.19
Relationship satisfaction 0.05 −0.02, 0.12 −0.03 −0.10, 0.04 0.02 −0.04, 0.08 0.00 −0.08, 0.08 0.05 −0.02, 0.12
Psychological distress 0.03 −0.04, 0.10 0.04 −0.03, 0.10 −0.03 −0.08, 0.03 −0.03 −0.10, 0.03 0.06 −0.00, 0.12
Depression 0.01 −0.06, 0.08 0.05 −0.02, 0.12 −0.02 −0.08, 0.03 −0.03 −0.11, 0.04 0.01 −0.05, 0.07
Anxiety 0.03 −0.04, 0.10 0.02 −0.06, 0.10 −0.04 −0.09, 0.02 −0.01 −0.09, 0.07 0.07 0.01, 0.12
Self‐esteem 0.01 −0.06, 0.08 0.02 −0.05, 0.08 0.01 −0.04, 0.06 0.06 −0.01, 0.13 0.07 0.02, 0.12
Personal well‐being −0.01 −0.08, 0.05 −0.04 −0.12, 0.03 0.05 −0.00, 0.09 0.05 −0.02, 0.12 −0.02 −0.07, 0.04
Life satisfaction 0.04 −0.03, 0.10 −0.02 −0.08, 0.05 0.05 −0.00, 0.19 0.03 −0.03, 0.09 0.05 −0.01, 0.11
Open in a new tab

Note: Confidence intervals that do not cross zero are bolded.

Abbreviations: ATE, average treatment effect; CI, confidence interval.

FIGURE 2.

FIGURE 2

Open in a new tab

Results showing the effects of increases in Antagonism, Emotional Stability, Disinhibition, Narcissism and Global Psychopathy on partner well‐being indicators the following year (z‐transformed). Confidence intervals that cross zero are in black, and those positive and not crossing zero are in orange.

4. Study 2

4.1. Materials and Methods

For Study 2, we used NZAVS panel data from waves 10 to 15. Time 13 data were collected during 2021–2022, Time 14 during 2022–2023, and Time 15 during 2023–2024. We identified romantic couples in the same way as in Study 1. To account for attrition across these additional waves, we applied inverse probability of censoring weights, using the censoring modeling built into the lmtp package. As in Study 1, our censoring model included full exposure and covariate histories up to the point of censoring, assuming that censoring is random given these histories (Hernán and Robins 2020). When a couple separated, we treated both partners as censored. Under these assumptions, our results generalize to the population at baseline had no one been censored (Williams and Díaz 2023).

4.1.1. Participants

Participants were all identified romantic couples who were in a relationship at Time 10 (baseline wave). As in Study 1, couples that broke up or where one or both partners dropped out of the study at any time after the baseline wave were censored. However, whereas in Study 1 full information was required for the exposure wave, in Study 2 censoring was allowed for the first and subsequent exposure wave, with censoring weights used to upweight those participants in the baseline wave who were more likely to drop out. These decisions were made in advance of any analysis with a view to obtaining inferences that more closely approximated the population of couples in New Zealand. Because psychopathy may be associated with attrition, the approach in Study 2 captured a broader population of couples in which partners may be subject to the effects of psychopathy. Such a mismatch between the analytic sample and target population is of concern in a setting of psychopathy, which is associated with short‐term mating strategies (Tsoukas and March 2018). Although attrition prior to exposure will not lead to confounding, it may lead to systematic differences between the analytic sample population and the target population (Bulbulia 2024b). Of the full 1832 couples at baseline, 1400 remained uncensored and eligible for analysis at the first intervention wave, which we treat as the analytic baseline. Of these, 332 had both partners observed at the sixth wave (23.7% retention). In contrast to Study 1, we permitted attrition (censoring) at the first intervention wave, allowing the study to better address left‐censoring bias or bias arising from selective inclusion into the analytic sample. Missingness in partner outcomes increased monotonically over time: approximately 10% at wave 2, 34% at wave 3, 55% at wave 4, and 77% at wave 5. This reflected dyad‐level censoring due to separation or partner‐specific dropout, rather than item nonresponse within observed couples. As in Study 1, we model censoring as an intervention and apply inverse‐probability‐of‐censoring weights in LMTP so that estimates are interpretable for the full baseline cohort under complete follow‐up.

Missing data in Study 2 were imputed in the same manner as the main analysis, with binary indicators created to denote whether an observation was missing. A key challenge in dyadic data is obtaining valid estimation under attrition, where if any individual is missing an observation, both the individual and their partner must be regarded as censored. To address this challenge, where data were missing in any of the four treatment waves, observations were carried forward from the previous wave if (and only if) an observation for this indicator was observed in a future wave, and an NA indicator was created from which machine learning algorithms could learn that an observation was imputed. This method, which tends to be conservative for estimating the causal effects of change in treatment values, follows the handling of missing data reported in Hoffman et al. (2023).

Our selection criteria resulted in 916 couples (N = 1832). A total of 888 (49%) participants identified as male, and age ranged between 19 and 90 years old (M = 50, SD = 12). Most participants were married (n = 1435, 78%), with relationship duration ranging from less than a year to 61 years (M = 22, SD = 13). Participant self‐identified ethnicity was as follows: Pākehā/New Zealand European (n = 1519; 84%), Māori (n = 165; 9%), Asian (n = 88; 5%) and Pacific peoples (n = 36; 2%).

4.1.2. Measures

The measures used for Study 2 are the same as described in Study 1.

4.1.3. Statistical Analysis

For Study 2, we replicated the main analysis in Study 1 with the following modifications. First, instead of analyzing three waves, we conducted a six‐wave investigation. After controlling for baseline confounders measured in the initial baseline wave, psychopathy was either increased by 1 measurement unit from an individual's observed value (up to a maximum of 7) at each of the following four waves, or—in the contrast condition—we computed partner well‐being under no intervention. Each partner well‐being outcome was estimated using data from the final (sixth) wave. For this analysis, NZAVS Time 10 provided baseline control measures, Times 11–14 captured sequential interventions on psychopathy, and Time 15 measured the outcomes. This temporal structure ensured that reverse causation could not confuse results, as exposures of interest always preceded outcomes, and baseline confounders were measured before both exposures and outcomes.

Second, having defined our causal estimands and ordered our data to reflect the temporal order of causality, we employed a Sequential Doubly‐Robust Estimator from the lmtp package, which is designed for repeated treatments across multiple waves (Díaz et al. 2023; Hoffman et al. 2023). This estimator strengthens robustness to misspecification in either the outcome or treatment model. To estimate effects, we used the “Ranger” library, which employs efficient non‐parametric causal forest estimators that are resistant to overfitting (Wright and Ziegler 2017). Machine learning estimators allowed us to impose such strict confounding control without over‐parameterizing estimation. Moreover, our doubly‐robust estimation approach avoids over‐conditioning biases arising from conditioning on mediators and allows us to obtain interpretable results for our pre‐specified causal estimand (Robins and Greenland 1992).

Third, we accounted for time‐varying confounding. Where sequential treatments are estimated, there is a prospect that results may be invalidated or induce bias if the confounders of subsequent exposures are affected by past treatments. Notably, standard statistical models such as regression (or multi‐level regression) and structural equation modeling cannot obtain consistent causal effect estimates in the presence of time‐varying confounding (Hernán and Robins 2020; Rohrer and Murayama 2023). However, the sequentially doubly robust estimators we employed using the lmtp package are capable of assessing causality in this setting, under assumptions of “sequential ignorability.” Identification of our sequential treatment regime relies on sequential ignorability for the exposure and censoring, positivity, consistency, and partial interference within couples. Specifically, for sequential ignorability, at each wave and among couples uncensored to that wave, the partner's potential outcome under our fixed policy with no censoring is independent of both the index partner's current psychopathy level and whether the couple remains in the study, conditional on both partners' complete observed histories. These conditions allow future covariates to depend on current psychopathy and do not assume independence among future covariates. Under these assumptions, and with positivity and consistency, our estimates have the usual interpretation for the Time‐10 cohort under complete follow‐up. To address time‐varying confounding, at each exposure wave we included lagged values of all partner outcome variables. Lagged outcomes are among the strongest confounders of any psychopathy‐outcome association at the end of study (VanderWeele 2021). We also included indicators both of individual and partner subjective health and disability measured coincidentally with each exposure wave, on the assumption that these values may also serve as common causes of the exposure and outcome and may vary over time.

Fourth, we accounted for potential attrition during the treatment waves to estimate the per‐protocol effect of the intervention on the baseline population of couples that would have stayed together until the end of study had they not been censored. That is, by a per‐protocol effect, we refer to the causal effect on partner well‐being assuming all participants remained in relationships for four successive waves after the baseline wave and provided well‐being information for each of the well‐being outcomes in the final wave. Notably, the causal effects we estimate assume that our censoring model is adequate for explaining the causes of attrition.

4.2. Results

A series of models were conducted to investigate whether the observed effects from Study 1 evolve under a four‐year fixed sequential exposure regime and whether new effects would emerge. Results are presented in Table 2 and Figure 3 (see Tables A6–A10 in the Supporting Information for full results). Increasing Antagonism by one unit on a Likert scale (1–7, max score = 7) increased partner perceived relationship conflict and decreased partner personal well‐being and life satisfaction. Increasing Disinhibition by one unit reduced partner relationship satisfaction, personal well‐being, and life satisfaction. Increasing Emotional Stability by one unit decreased partner psychological distress, depression, and anxiety, as well as increased partner life satisfaction and relationship conflict. Increasing Narcissism by one unit increased partner depression and decreased partner self‐esteem. Finally, increasing Global Psychopathy by one unit decreased partner personal well‐being, self‐esteem, and relationship satisfaction, and increased partner psychological distress, depression, and relationship conflict. These findings indicate that increases in an individual's callousness, impulsivity, and arrogance worsen partner well‐being, while increases in self‐contentment and stress immunity improve partner well‐being. Overall, we observe substantial evidence that prolonged exposure to psychopathic traits diminishes partner well‐being, with the strongest and most reliable results observed in the escalation of partner's perceptions of relationship conflict. Notably, in contrast to the short‐term boost to self‐esteem from psychopathy observed in Study 1, sustained exposure to Global Psychopathy diminishes partner self‐esteem.

TABLE 2.

The effect of a unit measurement gain in Global Psychopathy and psychopathy facets for four years (Times 11–14) on indicators of partner well‐being (Time 15).

Partner well‐being indicator Antagonism Disinhibition Emotional stability Narcissism Global psychopathy
ATE 95% CI ATE 95% CI ATE 95% CI ATE 95% CI ATE 95% CI
Relationship conflict 0.08 0.01, 0.15 0.02 −0.03, 0.07 0.05 0.00, 0.09 0.04 −0.01, 0.08 0.20 0.15, 0.24
Relationship satisfaction −0.02 −0.07, 0.03 −0.06 −0.11, −0.01 −0.04 −0.08, 0.00 0.01 −0.03, 0.05 −0.05 −0.10, −0.01
Psychological distress 0.04 −0.03, 0.10 −0.01 −0.06, 0.03 −0.07 −0.10, −0.03 0.03 −0.01, 0.08 0.05 0.01, 0.09
Depression 0.04 −0.03, 0.12 −0.02 −0.07, 0.02 −0.06 −0.10, −0.02 0.05 0.00, 0.10 0.09 0.05, 0.14
Anxiety 0.01 −0.06, 0.07 −0.02 −0.06, 0.03 −0.07 −0.10, −0.03 −0.00 −0.05, 0.04 0.04 −0.00, 0.08
Self‐esteem −0.04 −0.10, 0.03 −0.03 −0.08, 0.01 −0.00 −0.04, 0.04 −0.06 −0.10, −0.02 −0.07 −0.10, −0.03
Personal well‐being −0.10 −0.17, −0.04 −0.07 −0.12, −0.02 0.03 −0.01, 0.07 −0.03 −0.07, 0.02 −0.06 −0.10, −0.02
Life satisfaction −0.05 −0.10, −0.00 −0.07 −0.12, −0.02 0.08 0.05, 0.12 0.03 −0.02, 0.08 −0.01 −0.05, 0.03
Open in a new tab

Note: Confidence intervals that do not cross zero are bolded.

Abbreviations: ATE, average treatment effect; CI, confidence interval.

FIGURE 3.

FIGURE 3

Open in a new tab

Results showing the effects of increases over four years in Antagonism, Emotional Stability, Disinhibition, Narcissism and Global Psychopathy on partner well‐being indicators in the fifth year (z‐transformed). Confidence intervals that cross zero are in black, those positive and not crossing zero are in orange, and those negative and not crossing zero are in blue.

5. Discussion

It has long been observed that psychopathy is associated with antisocial and destructive relational behaviors. However, whether psychopathic traits cause lower romantic partner well‐being is under‐researched. In two studies, we offer the first systematic attempt to quantify the magnitudes of causal effect of psychopathic traits across eight dimensions of partner well‐being.

Study 1 focused on the causal effects on partner well‐being in the year following interventions on an individual's psychopathy. Contrary to our hypotheses, we did not find evidence that increasing Antagonism, Emotional Stability, Disinhibition, and Narcissism alone reliably affects partner well‐being for any of the eight outcomes we measured. These results imply that short‐term increases in individual psychopathy traits are insufficient to produce reliable, measurable effects on partner well‐being. However, we find that intervening on Global Psychopathy, which increases callousness, self‐contentment, impulsivity, and arrogance, reliably increases both partner anxiety, perceived relationship conflict, and self‐esteem the following year. Not only does this result show that the aggregated construct of psychopathy differs in its causal effect on partner well‐being compared to the individual traits, it also reveals potential boosts to self‐esteem from a one‐year exposure to Global Psychopathy.

Our finding that Global Psychopathy is causally related to anxiety aligns with earlier correlational findings that male psychopathy and female partner psychological distress are associated (Forth et al. 2022). Regarding the effect on partner self‐esteem, prior research and theory hint at a possible explanation. Psychopathy has been linked with greater attractiveness (Brazil and Forth 2020), which may correspond with increased charm, social dominance, and perceived status and power. According to Social Comparison Theory (Suls and Wheeler 2012), partners may experience greater validation by association, resulting in an initial boost in self‐esteem. Although there is a basis for explaining the short‐term effect of Global Psychopathy on partner self‐esteem, further research is required to clarify the mechanism underlying this finding.

Study 2 examined the effects of sustained exposures to psychopathy over four years on partner well‐being, and results both support and extend findings from Study 1. First, the causal effects of individual facets of psychopathic personality on partner well‐being are not uniformly negative in one domain: increasing Emotional Stability by one unit for each of the four exposure waves leads to higher partner life satisfaction and lower partner psychological distress, depression, and anxiety. Nevertheless, apart from Emotional Stability, interventions on other individual facets of psychopathy tend to negatively affect partner well‐being.

Second, comparing the results of the first and second studies reveals that prolonged exposures to psychopathy are more detrimental to partner well‐being. A one‐unit increase in Antagonism for each of four waves causes greater partner‐perceived relationship conflict and lower partner personal well‐being and life satisfaction, while a one‐unit increase in Narcissism reduces partner self‐esteem and increases partner depression. Furthermore, a one‐unit increase in Disinhibition reduces partner relationship satisfaction, personal well‐being, and life satisfaction. These effects, which again are not evident in single wave treatments, appear to arise from longer exposure to psychopathic personality shifts. Notably, these findings align with prior research showing that relationship duration exacerbates the link between psychopathy and argumentativeness in couples (Webster et al. 2016) and that psychopathy contributes to declining partner relationship satisfaction over time (Weiss et al. 2018).

Finally, sustained exposure to modest increases of Global Psychopathy, corresponding with greater callousness, self‐contentment, impulsivity and arrogance, more strongly diminishes partner well‐being than do sustained exposures to psychopathy facets. Specifically, increasing Global Psychopathy for individuals by one unit above each individual's observed value over four waves leads to an average diminishing of partner self‐esteem, relationship satisfaction and personal well‐being. The Global Psychopathy intervention also amplifies partner relationship conflict, distress and depression. These effects were greater than in the short‐term intervention investigated in Study 1, and again, the strongest result is for augmented relationship conflict. We infer that Global Psychopathy over a sustained period is considerably more destructive on partner well‐being than it is over a shorter period. Moreover, we infer it is the coming together of psychopathic personality traits that particularly harms a partner's well‐being. This sustained harmful impact parallels prior research indicating that enduring antagonistic tendencies erodes relationship functioning over time (Weiss et al. 2018). The findings also align with conceptualizations of psychopathy as an emergent interpersonal syndrome (Lilienfeld et al. 2019), a malignant configuration of personality traits whose joint expression is distinct from and greater than the sum of its parts. Psychopathy, then, may be best conceptualized as a dynamic interpersonal syndrome with cumulative negative consequences for partner well‐being.

5.1. Limitations

Our study has some limitations. First, we utilized self‐report measures, which introduce possibilities for both systematic and random measurement errors. Self‐report measures of psychopathy may be particularly susceptible to a person's dishonesty and lack of insight (Sellbom et al. 2018). Even if many of the outcome variables measured, such as life and relationship satisfaction, are inherently subjective constructs, future research could include partner‐reported measures of psychopathic traits or other more objective measures where possible. Relatedly, we used short versions of some scales, such as a 3‐item version of the Rosenberg Self‐Esteem Scale (Rosenberg 1965). Although brief measures can be less precise, research suggests the short‐form version in the NZAVS retains reliability and validity comparable to the full scales (Sibley et al. 2025). Second, our causal estimates rely on a set of assumptions, including sequential ignorability and no unmeasured confounding. While these assumptions cannot be formally tested, we have included extensive covariates to account for confounding and reported sensitivity analyses (E‐values; see Tables A1–A10 in the Supporting Information) to indicate how strong unmeasured confounding would need to be to explain away the observed effects. Third, we modeled separation or partner‐specific dropout as dyad‐level censoring and targeted a “no breakup/no attrition” estimand. This choice addresses treatment‐dependent censoring (i.e., psychopathy may influence attrition, without extrapolating unobserved exposure paths after dissolution). The resulting estimates are interpretable for the Time‐10 cohort under complete follow‐up and, by design, exclude pathways operating through breakup or loss to follow‐up; they should not be read as total effects in the observed population. Identification and consistency rely on standard conditions (consistency, conditional exchangeability of both exposure and censoring given measured covariates, and positivity) and on partial interference restricted to members of the same couple. Future work should examine the joint dynamics of psychopathy, dissolution, and partner well‐being by analyzing breakup as an outcome, considering total effects that allow pathways through dissolution, and conducting sensitivity analyses for outcome‐dependent censoring. Moreover, we caution against generalizing findings outside the New Zealand population from which the data were sourced. Although our study consists of a national probability sample of couples in New Zealand who participated in the NZAVS, the results can only be extrapolated to populations and sociocultural settings that resemble New Zealand at the time these studies were conducted. Notably, there are some indications that suggest psychopathic traits may differ across cultures and ethnicities (Fanti et al. 2018); whether and how these patterns of variation modify the effects of psychopathy on partner well‐being is a question for future comparative research.

5.2. Conclusion

Correlational studies have provided many valuable insights in psychopathy and relationship research. However, the next step is to evaluate evidence for causality. Here, for the first time, we used a robust causal‐inference workflow with longitudinal, real‐world data from couples to assess whether—and how strongly—increasing psychopathic traits affects partner well‐being. In the short term, we find that higher Global Psychopathy increases partner self‐esteem but also amplifies partner anxiety and perceived relationship conflict. We do not observe reliable effects on partner well‐being from individual psychopathy facets. In the longer term, we find stronger evidence of partner harm from interventions on individual facets of psychopathy, and much stronger harm to partners from interventions on Global Psychopathy. Together, these findings underscore the power of combining longitudinal observational data with rigorous causal inference methods to address questions that have previously remained elusive to inquiry.

Author Contributions

The first author was responsible for the study conceptualization, project administration, measurement design, measurement analysis, writing the original draft, and leading the review and editing of subsequent drafts. The second author provided guidance on the measurement analysis, design, and relevant psychological theories, and reviewed and edited the manuscript. The third author provided supervision and reviewed and edited the manuscript. The fourth author managed data collection and reviewed and edited the manuscript. The last author was responsible for developing the causal model and its statistical analysis, and reviewed and edited the manuscript. All authors approved the final version of the manuscript for submission.

Funding

The New Zealand Attitudes and Values Study is funded by a grant from the Templeton Religion Trust (JB TRT‐2021‐10418).

Ethics Statement

All authors declare no conflicts of interest. Ethics approval for the NZAVS is reviewed every three years by the University of Auckland Human Participants Ethics Committee (#22576). Informed consent was obtained from all individual participants included in the study. The study hypotheses, methods and analysis plan for Study 1 were preregistered (https://osf.io/q89se/overview). All analysis scripts are publicly available (https://osf.io/ce4t9/overview).

Supporting information

Data S1: jopy70042‐sup‐0001‐Supinfo.docx.

JOPY-94-819-s001.docx (61.6KB, docx)

Data Availability Statement

The data described in the paper are part of the New Zealand Attitudes and Values Study (NZAVS). A de‐identified dataset containing the variables analyzed in this manuscript is available upon request from any member of the NZAVS advisory board for the purposes of replication or checking of any published study using NZAVS data.

References

  1. Brazil, K. J. , and Forth A. E.. 2020. “Psychopathy and the Induction of Desire: Formulating and Testing an Evolutionary Hypothesis.” Evolutionary Psychological Science 6, no. 1: 64–81. 10.1007/s40806-019-00213-0. [DOI] [Google Scholar]
  2. Bulbulia, J. A. 2024a. “Margot: MARGinal Observational Treatment‐Effects (0.2.1.62).” 10.5281/zenodo.12671605. [DOI]
  3. Bulbulia, J. A. 2024b. “Methods in Causal Inference. Part 2: Interaction, Mediation, and Time‐Varying Treatments.” Evolutionary Human Sciences 6: e41. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Carton, H. , and Egan V.. 2017. “The Dark Triad and Intimate Partner Violence.” Personality and Individual Differences 105: 84–88. 10.1016/j.paid.2016.09.040. [DOI] [Google Scholar]
  5. Chen, T. , He T., Benesty M., et al. 2023. xgboost: Extreme Gradient Boosting (R Package Version 1.7.6.1). CRAN. https://CRAN.R‐project.org/package=xgboost. [Google Scholar]
  6. Cohen, S. , and McKay G.. 1984. “Social Support, Stress and the Buffering Hypothesis: A Theoretical Analysis.” In Handbook of Psychology and Health, edited by Baum A., Singer J. E., and Taylor S. E., 253–267. Lawrence Erlbaum Associates, Inc. [Google Scholar]
  7. Collison, K. L. , Miller J. D., Gaughan E. T., Widiger T. A., and Lynam D. R.. 2016. “Development and Validation of the Super‐Short Form of the Elemental Psychopathy Assessment.” Journal of Criminal Justice 47: 143–150. 10.1016/j.jcrimjus.2016.09.002. [DOI] [Google Scholar]
  8. Cummins, R. A. , Eckersley R., Pallant J., Van Vugt J., and Misajon R.. 2003. “Developing a National Index of Subjective Well‐Being: The Australian Unity Well‐Being Index.” Social Indicators Research 64: 159–190. [Google Scholar]
  9. Díaz, I. , Williams N., Hoffman K. L., and Schenck E. J.. 2023. “Nonparametric Causal Effects Based on Longitudinal Modified Treatment Policies.” Journal of the American Statistical Association 118, no. 542: 846–857. 10.1080/01621459.2021.1955691. [DOI] [Google Scholar]
  10. Diener, E. D. , Emmons R. A., Larsen R. J., and Griffin S.. 1985. “The Satisfaction With Life Scale.” Journal of Personality Assessment 49, no. 1: 71–75. [DOI] [PubMed] [Google Scholar]
  11. Fanti, K. A. , Lordos A., Sullivan E. A., and Kosson D. S.. 2018. “Cultural and Ethnic Variations in Psychopathy.” In Handbook of Psychopathy, edited by Patrick C. J., 2nd ed., 529–569. Guilford Press. [Google Scholar]
  12. Forth, A. , Sezlik S., Lee S., Ritchie M., Logan J., and Ellingwood H.. 2022. “Toxic Relationships: The Experiences and Effects of Psychopathy in Romantic Relationships.” International Journal of Offender Therapy and Comparative Criminology 66. 10.1177/0306624X211049187. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Hare, R. D. 2006. “Psychopathy: A Clinical and Forensic Overview.” Psychiatric Clinics 29, no. 3: 709–724. 10.1016/j.psc.2006.04.007. [DOI] [PubMed] [Google Scholar]
  14. Hernán, M. A. , and Robins J.. 2020. Causal Inference: What if. Chapman & Hill/CRC; 2020. [Google Scholar]
  15. Hoffman, K. L. , Salazar‐Barreto D., Rudolph K. E., and Díaz I.. 2023. “Introducing Longitudinal Modified Treatment Policies: A Unified Framework for Studying Complex Exposures.” ArXiv Preprint ArXiv 2304: 09460.
  16. Kashy, D. A. , and Kenny D. A.. 1999. “The Analysis of Data From Dyads and Group.” In Handbook of Research Methods in Social Psychology, edited by Reis H. T. and Judd C. M.. Cambridge University Press. [Google Scholar]
  17. Kessler, R. C. , Green J. G., Gruber M. J., et al. 2010. “Screening for Serious Mental Illness in the General Population With the K6 Screening Scale: Results From the WHO World Mental Health (WMH) Survey Initiative.” International Journal of Methods in Psychiatric Research 19, no. S1: 4–22. 10.1002/mpr. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Khan, R. , Brewer G., Kim S., and Centifani L. C. M.. 2017. “Students, Sex, and Psychopathy: Borderline and Psychopathy Personality Traits Are Differently Related to Women and Men's Use of Sexual Coercion, Partner Poaching, and Promiscuity.” Personality and Individual Differences 107: 72–77. 10.1016/j.paid.2016.11.027. [DOI] [Google Scholar]
  19. Kiecolt‐Glaser, J. K. , and Wilson S. J.. 2017. “Lovesick: How Couples' Relationships Influence Health.” Annual Review of Clinical Psychology 13, no. 1: 421–443. 10.1146/annurev-clinpsy-032816-045111. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Lilienfeld, S. O. , Watts A. L., Murphy B., et al. 2019. “Psychopathy as an Emergent Interpersonal Syndrome: Further Reflections and Future Directions.” Journal of Personality Disorders 33, no. 5: 645–652. [DOI] [PubMed] [Google Scholar]
  21. Linden, A. , Mathur M. B., and VanderWeele T. J.. 2020. “Conducting Sensitivity Analysis for Unmeasured Confounding in Observational Studies Using E‐Values: The Evalue Package.” Stata Journal 20, no. 1: 162–175. 10.1177/1536867X20909696. [DOI] [Google Scholar]
  22. Lynam, D. R. , Gaughan E. T., Miller J. D., Miller D. J., Mullins‐Sweatt S., and Widiger T. A.. 2011. “Assessing the Basic Traits Associated With Psychopathy: Development and Validation of the Elemental Psychopathy Assessment.” Psychological Assessment 23, no. 1: 108. 10.1037/aOO21146. [DOI] [PubMed] [Google Scholar]
  23. Lynam, D. R. , Sherman E. D., Samuel D., Miller J. D., Few L. R., and Widiger T. A.. 2013. “Development of a Short Form of the Elemental Psychopathy Assessment.” Assessment 20, no. 6: 659–669. 10.1177/1073191113502072. [DOI] [PubMed] [Google Scholar]
  24. Malouff, J. M. , Thorsteinsson E. B., Schutte N. S., Bhullar N., and Rooke S. E.. 2010. “The Five‐Factor Model of Personality and Relationship Satisfaction of Intimate Partners: A Meta‐Analysis.” Journal of Research in Personality 44, no. 1: 124–127. 10.1016/j.jrp.2009.09.004. [DOI] [Google Scholar]
  25. Morgan, S. L. , and Winship C.. 2014. Counterfactuals and Causal Inference: Methods and Principles for Social Research. Cambridge University Press. [Google Scholar]
  26. Pfund, G. N. , and Hill P. L.. 2022. “Correlated Change in Sense of Purpose and Romantic Relationship Quality.” Personal Relationships 29, no. 4: 875–893. 10.1111/pere.12453. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Phillips, N. L. , Rose L., Lynam D. R., and Miller J. D.. 2025. “Elemental Psychopathy Assessment's Nomological Net: A Meta‐Analytic Review.” Personality Disorders: Theory, Research, and Treatment 16. 10.1037/per0000728. [DOI] [PubMed] [Google Scholar]
  28. Polley, E. , LeDell E., Kennedy C., and van der Laan M.. 2023. “SuperLearner: Super Learner Prediction.” https://github.com/ecpolley/SuperLearner.
  29. Robertson, E. L. , Walker T. M., and Frick P. J.. 2020. “Intimate Partner Violence Perpetration and Psychopathy: A Comprehensive Review.” European Psychologist 25: 134–145. 10.1027/1016-9040/a000397. [DOI] [Google Scholar]
  30. Robins, J. M. , and Greenland S.. 1992. “Identifiability and Exchangeability for Direct and Indirect Effects.” Epidemiology 3, no. 2: 143–155. [DOI] [PubMed] [Google Scholar]
  31. Rohrer, J. M. , and Murayama K.. 2023. “These Are Not the Effects You Are Looking for: Causality and the Within−/Between‐Persons Distinction in Longitudinal Data Analysis.” Advances in Methods and Practices in Psychological Science 6, no. 1: 25152459221140842. 10.1177/25152459221140842. [DOI] [Google Scholar]
  32. Rosenberg, M. 1965. “Rosenberg Self‐Esteem Scale (RSE). Acceptance and Commitment Therapy.” Measures Package 61, no. 52: 18. [Google Scholar]
  33. Seigfried‐Spellar, K. C. , Villacís‐Vukadinović N., and Lynam D. R.. 2017. “Computer Criminal Behavior Is Related to Psychopathy and Other Antisocial Behavior.” Journal of Criminal Justice 51: 67–73. 10.1016/j.jcrimjus.2017.06.003. [DOI] [Google Scholar]
  34. Sellbom, M. , Lilienfeld S. O., Fowler K. A., and McCrary K. L.. 2018. “The Self‐Report Assessment of Psychopathy: Challenges, Pitfalls, and Promises.” In Handbook of Psychopathy, edited by Patrick C. J., 2nd ed., 211–258. Guilford Press. [Google Scholar]
  35. Sibley, C. G. 2021. “Sampling Procedure and Sample Details for the New Zealand Attitudes and Values Study.” NZAVS Technical Documents: e01. http://www.nzavs.auckland.ac.nz. [Google Scholar]
  36. Sibley, C. G. , Stronge S., Lilly K. J., et al. 2025. “Comparative Reliability of 108 Scales and Their Short‐Form Counterparts.” New Zealand Journal of Psychology 53, no. 2: 57–76. 10.63146/001c.138416. [DOI] [Google Scholar]
  37. Smith, C. V. , Hadden B. W., Webster G. D., Jonason P. K., Gesselman A. N., and Crysel L. C.. 2014. “Mutually Attracted or Repulsed? Actor–Partner Interdependence Models of Dark Triad Traits and Relationship Outcomes.” Personality and Individual Differences 67: 35–41. 10.1016/j.paid.2014.01.044. [DOI] [Google Scholar]
  38. Suls, J. , and Wheeler L.. 2012. “Social Comparison Theory.” In Handbook of Theories of Social Psychology, vol. 1, 460–482. SAGE Publications Ltd. [Google Scholar]
  39. Tsoukas, A. , and March E.. 2018. “Predicting Short‐and Long‐Term Mating Orientations: The Role of Sex and the Dark Tetrad.” Journal of Sex Research 55, no. 9: 1206–1218. 10.1080/00224499.2017.1420750. [DOI] [PubMed] [Google Scholar]
  40. Van Buuren, S. 2018. Flexible Imputation of Missing Data. CRC Press. [Google Scholar]
  41. Van der Laan, M. J. , and Rose S.. 2011. Targeted Learning: Causal Inference for Observational and Experimental Data. Vol. 4. Springer. [Google Scholar]
  42. Van der Laan, M. J. , and Rose S.. 2018. Targeted Learning in Data Science. Springer International Publishing. [Google Scholar]
  43. VanderWeele, T. J. 2021. “Can Sophisticated Study Designs With Regression Analyses of Observational Data Provide Causal Inferences?” JAMA Psychiatry 78, no. 3: 244–246. 10.1001/jamapsychiatry.2020.2588. [DOI] [PubMed] [Google Scholar]
  44. VanderWeele, T. J. , and Ding P.. 2017. “Sensitivity Analysis in Observational Research: Introducing the E‐Value.” Annals of Internal Medicine 167, no. 4: 268–274. 10.7326/m16-2607. [DOI] [PubMed] [Google Scholar]
  45. VanderWeele, T. J. , Mathur M. B., and Chen Y.. 2020. “Outcome‐Wide Longitudinal Designs for Causal Inference: A New Template for Empirical Studies.” Stastical Science 35, no. 3: 437–466. 10.1214/19-STS728. [DOI] [Google Scholar]
  46. Wager, S. , and Athey S.. 2018. “Estimation and Inference of Heterogeneous Treatment Effects Using Random Forests.” Journal of the American Statistical Association 113, no. 523: 1228–1242. 10.1080/01621459.2017.1319839. [DOI] [Google Scholar]
  47. Webster, G. D. , Gesselman A. N., Crysel L. C., et al. 2016. “An Actor‐Partner Interdependence Model of the Dark Triad and Aggression in Couples: Relationship Duration Moderates the Link Between Psychopathy and Argumentativeness.” Personality and Individual Differences 101: 196–207. 10.1016/j.paid.2016.05.065. [DOI] [Google Scholar]
  48. Weiss, B. , Lavner J. A., and Miller J. D.. 2018. “Self‐ and Partner‐Reported Psychopathic Traits' Relations With Couples' Communication, Martial Satisfaction Trajectories, and Divorce in a Longitudinal Sample.” Personality Disorders: Theory, Research, and Treatment 9, no. 3: 239–249. 10.1037/per0000233. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Wheatley, T. , Thornton M. A., Stolk A., and Chang L. J.. 2024. “The Emerging Science of Interacting Minds.” Perspectives on Psychological Science 19, no. 2: 355–373. 10.1177/17456916231200177. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Williams, N. , and Díaz I.. 2023. Package ‘lmtp’ (1.3.2). Lmtp. www.cran.r‐project.org/web/packages/lmtp/lmtp.pdf. [Google Scholar]
  51. Wright, M. N. , and Ziegler A.. 2017. “Ranger: A Fast Implementation of Random Forests for High Dimensional Data in C++ and R.” Journal of Statistical Software 77, no. 1: 1–17. 10.18637/jss.v077.i01. [DOI] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Data S1: jopy70042‐sup‐0001‐Supinfo.docx.

JOPY-94-819-s001.docx (61.6KB, docx)

Data Availability Statement

The data described in the paper are part of the New Zealand Attitudes and Values Study (NZAVS). A de‐identified dataset containing the variables analyzed in this manuscript is available upon request from any member of the NZAVS advisory board for the purposes of replication or checking of any published study using NZAVS data.

Articles from Journal of Personality are provided here courtesy of Wiley

RESOURCES