Berkson's Paradox: Why Attractive People Seem Less Wealthy, and How Dating and Marriage Markets Can Fool Us
Have you ever noticed how people in certain dating circles seem to have qualities that compensate for one another? Someone who isn't particularly attractive might be financially well-off, while someone who is exceptionally attractive might not be particularly wealthy. Similarly, in some marriage markets, candidates with modest careers might come from affluent families, while those from modest backgrounds might have stronger individual careers.
It is tempting to interpret these observations as evidence of some underlying trade-off between different human qualities. But what if the relationship we are observing does not actually exist in the general population?
There is a statistical phenomenon that can explain how such misleading relationships emerge. It is called Berkson's paradox, and it has interesting implications for how we understand dating, relationships and marriage markets.
What is Berkson's paradox?
Berkson's paradox is a form of selection bias in which two variables that are independent, or even positively associated, in the general population appear negatively associated within a selected group.
In simpler terms, selecting people based on two or more characteristics can create an apparent relationship between those characteristics that did not exist before the selection took place.
The reason is that when people are selected because they possess at least one desirable characteristic, having more of one characteristic can compensate for having less of another.
Let's understand this with a simple example.
The classic example: Attractiveness and wealth
Imagine a population of 100 people. For simplicity, divide them into two groups based on physical attractiveness and income:
| High income | Low income | Total | |
|---|---|---|---|
| High attractiveness | 25 | 25 | 50 |
| Low attractiveness | 25 | 25 | 50 |
| Total | 50 | 50 | 100 |
In this hypothetical population, attractiveness and income are completely independent. Half of the attractive people have high incomes, and half of the less attractive people have high incomes.
There is no relationship between the two characteristics.
Now imagine a dating platform or hypothetical dating market where someone is considered eligible if they have either above-average physical attractiveness OR above-average income.
The 25 people who are low in both characteristics do not make the cut.
We are now left with 75 people:
| High income | Low income | Total | |
|---|---|---|---|
| High attractiveness | 25 | 25 | 50 |
| Low attractiveness | 25 | 0 | 25 |
| Total | 50 | 25 | 75 |
Look at what has happened.
Among the attractive people, only 50% have high incomes.
Among the less attractive people, 100% have high incomes.
Suddenly, attractiveness and income appear negatively related. It almost looks as though less attractive people are richer!
But nothing changed in the original population. The relationship was created entirely by the process of selecting people into the dating pool.
The explanation is straightforward: attractive people did not need to be wealthy to qualify, while less attractive people needed some other desirable quality, such as wealth, to make the cut.
One characteristic effectively compensated for the absence of another.
That is the essence of Berkson's paradox.
How Berkson's paradox manifests in dating
Dating is particularly interesting because people select potential partners based on multiple characteristics.
Physical appearance, income, education, height, personality, social status, lifestyle and family background can all influence someone's desirability.
Imagine a dating market where people are willing to consider someone if they have at least one of the following:
Exceptional physical attractiveness.
High financial resources.
An impressive educational or professional background.
High social status.
Someone who scores highly on one dimension might still attract interest despite being relatively ordinary on others.
Someone who does not stand out in one area might need to compensate through another.
Consequently, when we examine only people who receive substantial attention or make it into a particular dating pool, we might observe negative relationships between certain characteristics.
For instance:
People who are exceptionally attractive might not need to be particularly wealthy to attract potential partners.
People who are less conventionally attractive might be disproportionately represented among those with high financial resources.
People with ordinary educational qualifications might compensate through social status or family wealth.
This could create the impression that attractiveness, wealth and other desirable qualities naturally trade off against one another.
But the same relationships might not exist in the general population.
There is an important distinction here: Berkson's paradox does not claim that these characteristics actually compensate for one another in people's lives. It explains how a selection process can make them appear to do so.
Berkson's paradox in India's arranged-marriage market
India's arranged-marriage system provides another interesting hypothetical illustration.
In arranged marriages, prospective matches are often evaluated on several dimensions. Depending on the family and community, these might include:
Individual salary and career prospects.
Educational qualifications and profession.
Family wealth and property.
Family reputation and social standing.
Community and other compatibility considerations.
Imagine a marriage market where a family considers a prospective groom acceptable if he has either an excellent individual career or a particularly strong family background.
For simplicity, assume that individual career prospects and family wealth are independent in the wider population.
Now consider the following hypothetical selection process:
| Individual career | Wealthy or established family | Modest family |
|---|---|---|
| Excellent career | Shortlisted | Shortlisted |
| Modest career | Shortlisted | Rejected |
Notice the resulting pattern.
A man with a modest salary might still make the shortlist because his family owns considerable agricultural land, property or an established business.
On the other hand, a man from a relatively modest family might make the shortlist because he has a prestigious government job, an excellent professional qualification or a high-paying career.
Among shortlisted candidates, family wealth and individual career prospects might therefore appear negatively associated.
Someone examining only these candidates might conclude that men from wealthier families tend to earn less, while high earners tend to come from less affluent families.
But that conclusion would be misleading if the two characteristics were independent in the wider population.
The apparent trade-off would have emerged because of the marriage-selection process.
Government jobs versus family wealth
Consider another hypothetical example.
Suppose a family is willing to consider either a man with a secure government job or a man from a financially established family.
A candidate without a government job might still be considered because his family owns considerable property or agricultural land.
Conversely, a candidate from a modest family might be considered because he holds a prestigious government position.
Among the shortlisted candidates, government employment and family wealth might appear negatively associated, even if they are independent in the general population.
Education versus family background
Imagine a marriage market in which families are willing to consider either someone with an elite educational qualification and strong career prospects or someone from a particularly affluent and established family.
Within the selected group, people with less impressive educational qualifications might disproportionately come from wealthy families.
People from less affluent families might need stronger educational and professional credentials to receive similar consideration.
Again, the resulting pattern does not establish that wealthy families produce less educated individuals or that highly educated individuals generally come from modest backgrounds.
It might simply reflect how candidates were selected.
Of course, actual Indian arranged-marriage markets are considerably more complicated. Many families consider several characteristics simultaneously, and preferences can differ substantially by gender, region, caste, community and socioeconomic background. These examples illustrate a possible statistical mechanism, not an empirical claim about all Indian families.
What about married couples and divorce?
The paradox can also help explain why studying only certain groups of couples can lead to misleading conclusions.
Imagine two factors that contribute to marital breakdown:
Serious financial stress.
Serious communication problems.
For simplicity, suppose these factors are independent across all marriages.
Now imagine studying only divorced couples, under the assumption that either severe financial stress or serious communication problems can independently be sufficient to cause a marriage to end.
Among divorced couples, you might observe:
Some couples with serious financial problems but relatively good communication.
Other couples with serious communication problems but relatively little financial stress.
Very few couples without either problem.
Financial stress and communication problems could consequently appear negatively correlated among divorced couples.
A researcher might mistakenly conclude that couples experiencing financial problems tend to communicate better, or that poor communication is less common when financial stress is present.
But this apparent relationship could be an artefact of studying only marriages that ended in divorce.
It is important to remember that this is a simplified hypothetical example. Real divorces have multiple interacting causes, and studying divorced couples alone cannot establish how different marital problems relate to one another in the wider population.
The statistical explanation: Conditioning on a collider
There is a name for the mechanism behind Berkson's paradox: conditioning on a collider.
Consider two independent characteristics:
A: Physical attractiveness.
B: Wealth.
Both influence whether someone gets selected for a date.
In causal diagrams, this relationship is represented as:
Attractiveness → Gets a date ← Wealth
The variable "Gets a date" is called a collider because two separate causal arrows point towards it.
Before selection, attractiveness and wealth are independent.
But when we restrict our analysis to people who got dates, we are conditioning on the collider.
This can create an association between attractiveness and wealth even though no such relationship existed in the original population.
The crucial insight is that selection itself can create correlations.
This is why researchers have to be careful when studying populations that have already been filtered through some selection process.
Does this mean all observed trade-offs are statistical illusions?
Not at all.
There are a few important distinctions to make.
1. The selection criteria matter.
The classic Berkson's paradox example involves selection through alternative or compensatory criteria: having either characteristic can be sufficient for selection.
If someone is selected only when they have both high attractiveness AND high income, the resulting association need not be negative. Different selection mechanisms can produce different results.
2. Real-world characteristics can genuinely be related.
Income, education, attractiveness, personality and social status may have actual relationships with one another. Berkson's paradox does not imply that every negative correlation is artificial.
It simply warns us that some observed relationships might have been created or distorted by selection.
3. Dating and marriage involve mutual selection.
Arranged marriages, dating apps and conventional dating are not one-way selection processes. Both individuals, and sometimes both families, evaluate one another.
This can produce additional patterns, including assortative mating, where people tend to partner with others who resemble them in education, socioeconomic background, religion or other characteristics.
Berkson's paradox and assortative mating can operate simultaneously.
4. The selected group can be genuinely different from the general population.
A negative relationship observed among shortlisted matrimonial candidates can be statistically real within that group. The mistake is assuming that it must also describe the wider population.
Conclusion: Always ask who made it into the sample
Berkson's paradox is a useful reminder that the relationships we observe depend not only on the characteristics we measure but also on how the people we are studying were selected.
When we examine only people who receive dates, only matrimonial profiles that attract interest, only successful arranged-marriage matches or only divorced couples, we are no longer looking at a random sample of the population.
The selection process can introduce relationships that did not previously exist, conceal relationships that do exist, or even reverse the apparent direction of an association.
So the next time you observe that attractive people seem less wealthy, that financially secure candidates seem to have weaker individual credentials, or that certain qualities appear to compensate for others in marriage markets, consider another possibility.
Perhaps people are not actually trading off those qualities.
Perhaps you are simply looking at a group of people who were selected precisely because they possessed at least one of them.
The most important question is not just "What relationship do these two characteristics have?" It is also "Why are we looking at this particular group of people in the first place?"
Sometimes, understanding how a group was selected tells us more than studying the group itself.
