We place too much weight on redundant information

The present work identifies a so-far overlooked bias in sequential impression formation. When the latent qualities of competitors are inferred from a cumulative sequence of observations (e.g., the sum of points collected by sports teams), impressions should be based solely on the most recent observation because all previous observations are redundant. Based on the well-documented human inability to adequately discount redundant information, we predicted the existence of a cumulative redundancy bias. Accordingly, perceivers’ impressions are systematically biased by the unfolding of a performance sequence when observations are cumulative. This bias favors leading competitors and persists even when the end result of the performance sequence is known. We demonstrated this cumulative redundancy bias in 8 experiments in which participants had to sequentially form impressions about the qualities of two competitors from different performance domains (i.e., computer algorithms, stocks, and soccer teams). We consistently found that perceivers’ impressions were biased by cumulative redundancy. Specifically, impressions about the winner and the loser of a sequence were more divergent when the winner took an early lead compared with a late lead. When the sequence ended in a draw, participants formed more favorable impressions about the competitor who was ahead during most observations. We tested and ruled out several alternative explanations related to primacy effects, counterfactual thinking, and heuristic beliefs. We discuss the wide-ranging implications of our findings for impression formation and performance evaluation.

That is from a new paper by Hans Alves and André Mata, via the excellent Kevin Lewis.


This emergence of the " bias- as- a type" kind of paper dovetails with the growing interest of biases in the workplace. I wonder if there are particular biases that are more significant in specific situations- than others.

Kahneman, the nobel-prize winning economist, has most notably identified five core biases- fundamental attribution error, amongst other categories of biases. FAE conflates and attributes the key success ( or lack thereof ) of the success with the person, rather than the situation of the factors.

I wonder if sequential information formation is part of our need to make sense of the morass of information, to impose some form of structure on the surrounding contexts.

Perhaps the inability to hold cognitively dissonant information is antithetical- reject information that does not conform to our understanding of trends, in favour of a more coherent, consistent story line.

Kahneman is a psychologist who just happens to be better at economics than most economists.


i'll stop reading this blog then

Why not bring the excellent Kevin Lewis on board?

He seems to be making someone redundant, after all.

Does MR use up all the newsprint alloted to them? https://en.wikipedia.org/wiki/Internet

(Disclaimer: this paper is gated, so I haven't read the paper itself.)

"...impressions should be based solely on the most recent observation because all previous observations are redundant."

This is an exceptionally strong statement to open with, and to be honest I don't think it's true. Or it may only be true in some cases, and generalising from a view that is true in many cases, especially domains in which you have little direct prior experience with, is a good bayesian strategy to forming a guess. I don't know whether that should be referred to as a 'bias', or better described as 'typical uninformed belief forming'.

For instance, any Elo model weights prior performance at a decreasing rate (depending on the K score you've set). These models, with some level of historical averaging, are typically much better than guessing, and are often the basis for pricing markets for gambling businesses. They're not perfect, but they're not bad. Many models also explicitly use a more sensitive k-score in early rounds as an earlier round may be more informative, since it's pairing up teams that you've never observed an a match before - and the players are learning strategies as they compete and adapting as they go.

Have the authors tested the validity of their original assertion in reality?

So trade redundancy bias for recency bias?

Some good comments here, and this is the pithiest good one.

Ever notice how Trump repeats himself all the time? He deeply understands redundancy bias. He's redundant. The most redundant.

Classic speech making, Tell the audience what you are going to tell them, tell them, and then tell them what you told them.

This kind of question is addressed by correlograms, or autocorrelation plots for measurements spread in time.

For continuous processes, what happens after time T is related to what happens now. Beyond the correlation time, which is greater than T, the observed parameter will hold no correlation to what happens now or happened in the past.

Think of a racing car at 100 MPH right now. In the following 0.1 seconds its speed can increase or decrease but even in the case of maximum braking being applied, it's going to be closer to 100 MPH than zero. 10 seconds after, anything between 0 and 200 MPH is possible. You can tell with higher confidence what a car will do in the following second, beyond that it's hard.

The performance of sport teams should also have a correlation time. If the team is good today, it will be probably be good in 1 week. Next season? Less probable.

With enough good quality and relevant data the time correlation of several things can be estimated. That would yield a limit beyond which past data is not so relevant.

Agreed. Observations over time establish a trend. Part of the analysis has to include determining when previous observations no longer contribute value to the trend.

Regardless, in many situations a trend is more valuable than a single observation, showing that previous observations are not completely redundant.

Then stop asking me for 20 page papers! I can answer you in 3

So like, Trump is bad, and people don't want to hear it, because it is redundant information, so they never actually do anything?

Trump is living in my head, making a constant racket and he won't leave!

He's living in our White House, which is actually worse!

And if Trump keeps making a constant racket, then I'm going to keep making a constant racket!

Dude, seriously.

We just put "one foot in, one foot out" on war with Iran, and your claim is that we are thinking too much about the performance of this administration?

Maybe your comments are a placeholder for the actual problem, that too many dodge responsibility in a democratic republic.

See also the Trump administration's argument that kids in US care require less than prisoners under the Geneva Convention.

Cubs won yesterday. Don't see how they lose again.

When the latent qualities of competitors are inferred from a cumulative sequence of observations (e.g., the sum of points collected by sports teams), impressions should be based solely on the most recent observation because all previous observations are redundant.

I don't understand this at all. Is the cumulative sequence the points scored in a single competition, like the runs scored in each inning of a baseball game? Or is it the cumulative result of a number of games?

If the former, OK, all it seems to say is we should pay attention to the final score, and ignore the information about which innings they were scored in.

If the latter, it seems obviously wrong.

Maybe if they wrote in English it would be clearer.

Hi, Are you looking for a regular investment partner to grow your business? We would like to be your marketing partner and we will not charge you full fixed services fee.

Do you want me to send you the details?


- This is an ad.
- Address: ST 1X, Raw, PK
- Unsubscribe: https://goo.gl/forms/xVqY6pof3ozkk6F62

Comments for this post are closed