[richmarr]

> You can still avoid most of the effects of a worst-case bias by adding two additional measurements... given 3 who all had the same experience

You're right that it's an advantage, but it reduces noise, not bias.

Bias by definition skews systemically in the same direction so the positive effect of taking multiple measurements is minimal.

[sokoloff]

Each bias skews the same direction for each person, but not every bias is in the same direction for individual people. (Some people are biased in favor of Harvard/Ivy League graduates. Other people are biased against those exact same candidates. Bias is not by definition unidirectional for all people.)

The YC partners are trying to be similarly biased against entrepreneurs who (they believe) will not be successful in the program.

They are much less likely to be similarly biased against irrelevant factors like accents, mannerisms, backgrounds, etc.

[richmarr]

> They are much less likely to be similarly biased against irrelevant factors like accents, mannerisms, backgrounds, etc.

They're not less biased, they just average out their biases over the group.

Your assumption is that three people chosen from a fairly homogenous pool are going to cancel out each others biases, which is... optimistic.

I don't know from this conversation what they're actually doing, but what they should be doing is using a diverse set of opinions to create a fixed set of questions and a fixed marking scheme, and then sticking to it for that round of interviews. Then looking back over time at every interview question and analysing how well it predicted later outcomes.

[random42]

A huge part of it is resulting discussion that happens between different observers which puts the onus to check for biases and map the signals on objective parameters.

[richmarr]

If I understand you correctly I think this is misleading.

Discussing candidates after an interview allows social dynamics within the group to distort the signal so you reduce the value of taking independent data points. Not only will it not reduce bias in the way you seem to suggest, but you'll also lose some of your ability to reduce random noise as the noise from more dominant interviewers will be amplified.

I don't have time to dig out citations, but a good starting point would be "What Works - Gender Equality By Design" by Iris Bohnet. She's one of the world's leading academics studying how biases are affected by different hiring techniques.