TL;DR
This paper investigates the "convergence in physical appearance" hypothesis, which suggests long-term partners' faces become more similar over time. Using a large dataset of couple photos and both human judges and a facial recognition algorithm, the study found that while spouses are initially similar in appearance, this similarity does not increase over their marriage, aligning facial appearance with other traits that show initial similarity but no convergence.
What Problem Does This Paper Address?
For decades, a popular idea in psychology has been that long-term romantic partners' physical appearances converge over time. This "convergence in physical appearance" hypothesis suggests that shared environments, emotional mimicry (like smiling together), and synchronized activities lead to spouses looking more alike as their marriage progresses. This idea is intuitively appealing and has been widely cited. However, this stands in contrast to findings for many other personal traits (personality, intelligence, attitudes, values), where couples are found to be similar at the start of a relationship but do not become more similar over time. The paper aims to resolve this apparent contradiction by rigorously testing the facial convergence hypothesis with a large-scale study.
Methods
The researchers employed two distinct methods to assess facial similarity:
- Human Judges: 153 participants from Amazon Mechanical Turk (AMT) were shown sets of faces and asked to rank them based on how closely they resembled a target face. This approach is similar to previous studies but uses a larger sample.
- Facial Recognition Algorithm: The study utilized VGGFace2, a state-of-the-art deep neural network known for its accuracy in facial recognition tasks. This algorithm converts faces into numerical descriptors, allowing for objective, quantitative similarity comparisons.
Both methods were applied to two sets of images for each spouse: one taken near the beginning of their marriage and another taken 20 to 69 years later. This longitudinal design is crucial for testing the "convergence over time" aspect.
Datasets
The study used a dataset of 517 white married heterosexual couples. Facial images were collected from public online sources:
- 392 newspaper wedding anniversary announcements.
- 102 Google Search results.
- 23 public profiles from Ancestry.com.
For each spouse, two images were collected: one from within 2 years of the wedding and another from 20 to 69 years later. The average marriage length in the dataset was 49 years. The authors note that they were unable to find sufficiently large samples of homosexual or non-white couples for a meaningful analysis.
Key Insights and Ideas
- Homogamy Confirmed: The study strongly supports the well-established finding that people tend to marry others who are already similar to them in appearance (homogamy). Both human judges and the facial recognition algorithm showed that spouses' faces were significantly more similar at the start of their marriage than would be expected by chance.
- Facial Appearance Joins the Club: Contrary to the widely cited "convergence in physical appearance" hypothesis, this study found no evidence that spouses' faces become more similar over time. This finding brings facial appearance in line with other personal characteristics like personality, attitudes, and intelligence, which also show initial similarity but no subsequent convergence.
- The "Smile Lines" Theory Questioned: The paper directly challenges the mechanism proposed by Zajonc et al. (1987), which suggested that shared experiences, like smiling, could lead to the co-development of similar facial features (e.g., smile lines). The lack of convergence suggests this specific mechanism, or at least its impact on overall facial similarity, is not as significant as previously thought.
Key Results
- Initial Similarity: Spouses' faces were found to be significantly similar at the time of marriage. The average rank for similarity was 2.75 for human judges and 2.89 for VGGFace2, both significantly lower than the chance value of 3.5.
- No Convergence: Over 20 to 69 years of marriage, spouses' faces did not become more similar. In fact, human judges perceived a slight decrease in similarity over time (average rank increased from 2.75 to 2.90), though this effect was small and not replicated by the VGGFace2 algorithm.
- Reliability: Inter-rater reliability for human judges was consistent across different time points, indicating that the aging of faces did not significantly impact their ability to judge similarity.
Limitations and Future Work
- Publicly Available Images: The study relied on publicly available images, meaning the authors had no control over image quality, lighting, angles, or self-presentation (e.g., grooming, facial expressions). While the authors argue these factors should amplify convergence if it were occurring, they acknowledge this limitation.
- Judge Demographics: The demographics (age, ethnicity) of the human judges were not recorded or controlled, which could introduce biases. However, the authors argue that these biases are unlikely to affect the change in similarity over time.
- Sample Demographics: The study was limited to white heterosexual couples due to sample size constraints. Future work could explore these phenomena in more diverse populations.
The authors conclude that while the rejection of the facial convergence hypothesis might be less exciting than its affirmation, it provides a more consistent understanding of how long-term relationships function, aligning facial appearance with other personal traits that are selected for similarity but do not converge over time.
Summary updated