Berkson's Paradox

You screen a universe of hedge funds for your portfolio, keeping only funds that clear a bar on either strong historical returns or low volatility — funds need to be good on at least one dimension to survive your filter. Looking only at the funds that passed, you notice something odd: returns and low volatility appear negatively correlated among your survivors — the high-return funds tend to be volatile, and the low-volatility funds tend to have mediocre returns. It's tempting to conclude chasing returns costs you stability. But this "trade-off" might be a complete illusion, manufactured entirely by the fact that you only looked at funds that passed a selection filter. That illusion is Berkson's paradox.

An analogy: dating pool logic

Suppose someone will only date a person who is either very kind or very attractive — they don't need both, just one is enough to clear the bar. Looking only at the people they actually date, kindness and attractiveness seem negatively correlated — the very attractive dates tend to be less kind, and vice versa. In the general population the two traits might be completely unrelated, or even positively related. But once you filter to "cleared the bar on at least one trait," people weak on both traits get excluded, and among survivors, being strong on one trait makes it less necessary to also be strong on the other — manufacturing an apparent trade-off that reflects the selection rule, not reality.

The logic, one piece at a time

Let $A$ and $B$ be two traits (say, high returns and low volatility) that are independent, or even positively associated, in the full population of funds. Suppose a fund is included in your dataset only if $A$ or $B$ (or both) hold — call this the selection event $S$. Within the selected group, conditioning on $S$ changes the relationship: a fund with a weak $A$ that made it into $S$ must have had a strong $B$ (that's the only way it could have cleared the bar), and vice versa. Formally, $P(A, B \mid S)$ can show negative association between $A$ and $B$ even when $P(A, B)$ (unconditional) shows none. In plain English: filtering a population on "at least one of two things is true" mechanically creates an apparent trade-off between those two things among the survivors, because failing on one makes passing on the other necessary, and that necessity shows up as correlation even with no real underlying relationship.

Worked example 1: a stylized fund screen

Suppose among 1,000 candidate funds, returns and volatility are truly independent: 50% have high returns, 50% have low volatility, drawn independently, so 25% have both, 25% have high returns only, 25% have low volatility only, and 25% have neither. Your filter keeps a fund if it has high returns OR low volatility — that's $25+25+25=750$ funds, excluding only the 250 with neither. Look at the conditional structure among survivors: those who do NOT have high returns (250 funds, all "low volatility only") are 100% low-volatility. Those who DO have high returns (500 funds) are only 50% low-volatility. Having high returns dropped the rate of low volatility from 100% to 50% — a stark negative association, purely from the selection filter, despite zero true correlation in the original 1,000 funds.

Worked example 2: seeing it vanish without selection

Rerun the same numbers on the full 1,000-fund population before filtering: among funds without high returns (500 of them), the fraction with low volatility is $25/50 = 50\%$ — the same 50% as among funds with high returns. No association at all. The negative correlation only appeared in worked example 1 because the "neither" funds (250 of them) were dropped by the OR-filter, and it's specifically the removal of "weak on both" cases that manufactures the illusion. This confirms the effect isn't in the data-generating process — it's entirely a byproduct of which rows you kept.

Drag the slider above toward a strongly negative $\rho$ and picture it as "survivors only" — this is the shape returns and volatility take on among funds that passed an OR-filter, even though the true, unfiltered relationship (rho near zero) never looked like this at all.

What this means in practice

Berkson's paradox lurks in almost every curated financial dataset: index membership rules, fund databases that only include funds meeting a minimum AUM or track record, credit datasets that only contain applicants who were approved. Any time a dataset was assembled using an "at least one of several conditions" filter, apparent trade-offs or correlations among the filtered variables should be treated as suspect until checked against the unfiltered population, if it's available at all — which it often isn't, making Berkson's paradox one of the harder biases to catch after the fact.