← Thinking, Fast and Slow
第 3 章 · 22 分钟
Anchoring and availability: two mechanisms tested repeatedly
Effect sizes, replication records, and which claims to discount.
概念地图
Anchoring
A number seen first pulls the estimate that follows toward it.
深色为本章概念，浅色为其他章
陪练模式
不想只是读？让它带你走一遍。
导师把这一章拆成小步，每一步先问你，再讲；你用自己的话答，它先认下对的部分再纠偏。用你自己的 AI，对话只存在这台设备上。
Locus01
延伸阅读
Tversky and Kahneman · 1974 · anchoring and adjustment
Proposed that people start from an initial value and adjust toward the target, adjusting too little. They argued against treating numerical estimation as independent of presentation order; the cost is that the mechanism behind insufficient adjustment was never settled, and at least two incompatible accounts followed.
Foundational
The availability heuristic · same period
Estimating frequency by how easily instances come to mind. It sits beside anchoring but has a different evidential structure: it relies on indirect manipulation of 'ease', which makes clean experiments harder.
Foundational
Open Science Collaboration · 2015 · large-scale replication in psychology
Independently repeated over a hundred published effects and found an overall replication rate far below what the original literature implied. It argued against reading a single significant result as an established fact, and split 'does the effect exist' from 'how large is it' into two questions.
Revision
Kahneman · 2011 · the chapters on anchoring and availability
Presents the two heuristics alongside a set of priming experiments. The author stated publicly in 2017 that the evidence in the priming sections was weaker than he had believed while writing. This chapter grades the evidence accordingly.
Subject of this course
机制02
Anchoring
Tversky and Kahneman gave the name in 1974 for what a randomly drawn figure does to the guess that comes after it: anchoring. Its evidential position in this book is the strongest of any mechanism: stable across tasks and populations, and it survived the large replication projects. Note that its explanation is contested — insufficient adjustment versus selective retrieval — and this course takes no side between them.
机制A number seen first, through becoming the starting point of the estimate, pulls the judgement toward itself.
可迁移性测试
Move it to a negotiation: whoever quotes first sets the band within which every counter-offer moves. The isomorphism breaks in that a negotiating anchor carries information about the other side's limit, while an experimental anchor is random — the two are not equally legitimate to follow.
机制03
Availability
The second heuristic named in the same 1974 paper, and the one this course files under chapter 1 rather than beside anchoring, is availability. Its place in the argument is that it is an instance of substitution: the target attribute is frequency, the stand-in is retrieval fluency. The distribution of news coverage therefore rewrites people's ranking of risks — not because the judge is irrational, but because the stand-in attribute really was changed.
机制Retrieval fluency, through standing in for frequency information, changes the estimated commonness of a class of events.
可迁移性测试
Move it to a search engine: the density of the first page sets your impression of how big a topic is, and the first page is decided by the ranking algorithm, not by the topic's size. The isomorphism breaks in that result counts can be looked up directly, while the true frequency of instances in memory cannot.
Derivation04
From one significant result to 'how much to discount this claim'
What is being derived is the step this book most needs from its reader: separating 'the effect exists' from 'the effect is this large', with grounds for the discount.
Published literature is biased toward significant results
Chosen, and well supported: null results are harder to publish.
In a small sample, an estimate that clears the significance threshold must be large
Chosen, though close to arithmetic: only a large estimate can clear the noise.
Replication studies use larger samples and more preregistration than the originals
Chosen, and an accurate description of replication projects since the 2010s.
推导 · 0 / 4
Worked derivation05
Same book, two grades of evidence
'Anchoring is real' and 'pictures of money in the room make people more selfish' both come from this book. Do they deserve the same credence?
Ask separately: how large were the samples, was there independent replication, did the effect size hold up in it?
裂缝06
Cracks in this framework
争议地形07
本书主张
Anchoring and availability are stable mechanisms of judgement; they recur across many experiments and predict the direction in which judgement departs.
另一种看法
The post-replication line holds that a substantial share of published effects are far smaller than the literature suggests and some do not replicate at all, so popular books built on literature reviews systematically overstate them.
分歧扎在
The disagreement is rooted in whether published evidence can be read directly as population evidence, not in any particular mechanism. Both sides agree anchoring is robust; they differ on how to treat the weakly evidenced effects.
什么证据能裁决
What would settle it is item-by-item comparison: split the experiments the book cites by whether they entered a large replication project, and compare the ratio of original to replicated effect size. Ratios near one across both groups would refute the overstatement claim; systematically larger originals in the unreplicated group would show grading is necessary.
The evidence points to grading being necessary: among effects the replication projects covered, replicated sizes are markedly smaller overall, with anchoring comparatively robust. What is missing is a treatment for effects no project has covered — they can be assumed neither true nor false.
Falsification08
Anchoring as a preregistrable test
The mechanism under test: a number seen first, through becoming the starting point of the estimate, pulls the judgement toward itself.
Data source: newly collected estimates, with anchors generated on the spot by a random number generator and logged
The anchor must be unrelated to the target, or the test measures reasonable updating instead.
Sample: a pre-computed sample size analysed in full, with no exclusion on response time or 'seriousness'
Post hoc exclusion is the commonest degrees-of-freedom entry.
Window: a single session, with no other task inserted between anchor and estimate
An intervening task introduces memory decay as an extra variable.
Threshold: the difference between the high-anchor and low-anchor groups reaches a pre-registered minimum effect size
Register a minimum effect size, not just significance — with a large sample, significance alone turns any trivial difference into 'support'.
Failure condition: the difference runs the wrong way, or the effect size falls below the registered minimum
The second also counts as refutation: an effect too small to matter is, in application, the same as none.
推导 · 0 / 3
接口09
Which slot it hangs on
挂在哪个槽位The conclusion you already hold is probably 'studies show that…' — reading 'published' as 'established'. This chapter goes straight at it.
Does this chapter (a) constrain 'studies show that…', or (b) open a new slot for grading evidence?
慢变量Register two slow variables: what share of the studies you cite you have checked for replication, and how many of your working conclusions rest on a single small-sample study. Look at the second first.
小结10
本章小结
01Anchoring is the best-evidenced mechanism in the book and survived large-scale replication.
02Availability is an instance of substitution: frequency is swapped for retrieval fluency.
03'Published effects are inflated' is an arithmetic consequence of a publication threshold; nobody has to cheat.
04Two sentences in one book can differ in credibility by an order of magnitude, and the text does not carry that information.
05Anchoring has two incompatible causal accounts; a test settles the effect, not the cause.
提取练习 · 合上书，先自己答一遍。
?Is 'the smaller the sample, the larger an observed effect must be to clear significance' (a) an identity or (b) an empirical claim?
?Is moving toward the other side's opening price (a) an anchoring bias or (b) reasonable updating on an informative anchor?
Forced choice11
B · Identify the tag
'The expectation of published effect sizes exceeds the true effect' — what kind of step is that?
二选一
Forced choice12
C · Locate the crack
In which situation does this framework give a confident and wrong answer?
二选一
读到这里，把它记为已读。