Skip to content
#small language model Open access

With No Cause at All, the Bottom Rises 1.76 Standard Deviations and the Top Falls 1.76 ── Everyone equally able, both measurements independent, and the bottom decile moves +1.756597 while the top decile moves -1.759617 ── One treatment looks opposite depending on which group was chosen ── [Paper 371]

Sep 2026 · Zenodo (CERN European Organization for Nuclear Research)
Advanced Causal Inference Techniques

Abstract

It is said that drilling the weak improves them and praising the strong spoils them. This paper asks whether that is the effect of the treatment──the answer is that it happens at the same size with no treatment at all. No new mathematical theorem and no new law is claimed. Scope of this paper (scope note): No new mathematical theorem and no new law is claimed──regression to the mean, Galton’s regression, and regression discontinuity are all standard. We do not build statistics──corrections for regression (control-group design, difference-in-differences) are not entered. They are named and no more. We use no real data──actual records from education, medicine, or sport are not used. The computation is on standard normal draws. We do not build causal inference──identification conditions for treatment effects are not the subject. Only what is seen when the treatment is zero. Correlation is not treated──the two measurements are taken as fully independent. In reality they correlate, and the regression is smaller by that much. We do not generalise──not every before-and-after comparison is false. Only those in which the group was selected on the first measurement. Relation to earlier papers: Paper 364 showed selection alone raising the mean by 1.52 standard deviations──there the issue was which group is visible; here it is the “change” when one group is measured twice. Paper 353 showed a claim turning false when its condition is dropped──here the dropped condition is how the group was chosen. Paper 300 showed whether two things share a root is decidable──by that test, a treatment effect and a regression effect have distinct roots. Paper 361 showed a classifier failing when tried on a known instance──here the known instance is a treatment of exactly zero. What is added is giving first and second means for five groups, showing that bottom and top move in opposite directions by nearly equal amounts, showing both second means returning to zero, placing a control in which making the second measurement identical to the first removes the effect exactly, and putting the separator on selection by the first measurement. First, everyone is given equal ability and the two measurements are independent noise (Section 2). Second, this is the core of the paper. The bottom decile’s second measurement rises by +1.756597 (Section 2). Third, at the same time the top decile falls by -1.759617 (Section 2). Fourth, both second means are near zero (0.008800 and -0.000170) (Section 2). Fifth, the middle half barely moves (+0.004677) (Section 3). Sixth, the separator is whether the group was chosen on the first measurement, not what the treatment was (Section 5). It is said that drilling the weak improves them and praising the strong spoils them──but the same thing happens, at the same size, with no treatment at all. Give 100000 people equal ability and make the two measurements independent noise──nothing is done between them, so any change is entirely an appearance. And still the bottom rises and the top falls──the bottom decile moves +1.756597 and the top decile -1.759617: opposite in sign and nearly equal in size. Both second means are near zero (0.008800 and -0.000170) ── both groups have simply returned to the population. So the “change” lives on the first side: that the first mean was -1.747797 makes the whole of it. One treatment looks opposite by group──applied to the bottom it looks effective, applied to the top harmful, and the treatment is “do nothing”. The observation “scolding improves and praising spoils” is fully explained by this ── and being able to explain something is not evidence that the explanation is right. The size is set by how extreme the selection is──the middle half moves only +0.004677, one three-hundred-and-seventy-sixth of the bottom decile’s. A design that picks an extreme group and compares before with after produces the strongest false effect ── and picking the extreme group is the natural decision to start where the trouble is. Make the second measurement identical to the first and the effect is exactly 0, naming the noise as the cause. One thing separates them──whether the group was chosen on the first measurement; not what the treatment was. The treatment is identical in all three rows and only the selection differs, yet the apparent effect runs from +1.76 to -1.76. Which is why randomisation works; and where one is forced to choose the group in trouble, choosing the control by the same criterion makes the regression cancel in the difference. To be said honestly──real pairs of measurements are correlated and the regression is smaller by that much; the 1.756597 here assumes independence and is an upper bound. And this differs in root from Paper 364: the data here have no gaps, so the records show nothing wrong and only the design reveals it. One last thing──one treatment of exactly zero can produce both “it worked” and “it backfired”. Choose the group according to the conclusion you want. A design whose conclusion can be chosen does not support that conclusion, and looking at results without looking at the design cannot tell the two apart. On the making of this work: The ideas and content of this work stem from the author's own considerations. Assistance from an AI (a large language model) was used for structuring, English translation, and checking the algebra. Any remaining errors or misinterpretations are solely the author's. Feedback and corrections are sincerely appreciated. ----- 成績の悪い者を鍛えると伸び、良い者を褒めると落ちる、と言われる。本稿が問うのはそれが処置の効果かである──答えは処置が何も無くても、同じ大きさで起きる。新しい定理も法則も主張しない。 本稿の射程(射程注記):新しい定理も法則も主張しない──平均への回帰・ゴルトンの回帰・回帰不連続はすべて既知である。統計学を作らない──回帰への補正法(対照群の設計・差の差法)には立ち入らない。名前を挙げるにとどめる。実データを扱わない──教育・医療・スポーツの実際の記録は用いない。標準正規乱数の上で計算する。因果推論を作らない──処置効果の識別条件は主題ではない。処置がゼロの場合に何が見えるかだけを扱う。相関を扱わない──二回の観測は完全に独立とする。実際には相関があり、回帰はその分だけ小さくなる。一般化しない──全ての前後比較が偽であるとは主張しない。一回目で選別した場合だけである。既刊との関係:論文364 は選別だけで平均が 1.52 標準偏差上がることを示した──そこでは「見える集団」が問題だった。ここでは同じ集団を二度測ったときの「変化」が問題である。論文353 は条件を落とすと主張が偽になることを示した──ここでは「選び方」という条件が落ちている。論文300 は同根か別根かが判定できることを示した──処置の効果と回帰の効果は、その基準で別根である。論文361 は既知の一件で試すと分類が落ちることを示した──ここでは処置ゼロという既知の一件で試している。加えたのは、五つの群で一回目と二回目の平均を数に出したこと、下位と上位で符号が逆になり大きさがほぼ同じであることを示したこと、二回目の平均がどちらも 0 に戻ることを示したこと、二回目を一回目と同一にすると効果が厳密に消える対照を置いたこと、分離子を「一回目で選んだか」に置いたことである。 第一に、実力を全員同じにし、二回の観測を独立な雑音とする(第2節)。 第二に、これが本稿の芯である。下位 10 パーセントの二回目は +1.756597 上がる(第2節)。 第三に、同時に上位 10 パーセントは -1.759617 下がる(第2節)。 第四に、二回目の平均はどちらも 0 に近い(0.008800 と -0.000170)(第2節)。 第五に、中央 50 パーセントはほとんど動かない(+0.004677)(第3節)。 第六に、分離子は「一回目で選んだか」であって、処置の内容ではない(第5節)。 成績の悪い者を鍛えると伸び、良い者を褒めると落ちる、と言われる──だが処置が何も無くても、同じことが同じ大きさで起きる。100000 人の実力を全員同じにし、二回の観測を独立な雑音とする──二回目には何もしていないのだから、変化があればそれは全部が見せかけである。それでも下位は上がり、上位は下がる──下位 10 パーセントの二回目は +1.756597 上がり、上位 10 パーセントは -1.759617 下がる。符号が逆で、大きさはほぼ同じ。そして二回目の平均はどちらも 0 に近い(0.008800 と -0.000170)──どちらの群も母集団に戻っているだけである。つまり「変化」は一回目の側にあり、一回目が -1.747797 だったことが変化の全部を作っている。同じ処置が、選んだ群で逆に見える──下位に施せば効いたと見え、上位に施せば害があったと見える。同じ「何もしない」という処置がである。「叱ると伸び、褒めると落ちる」という観察が、これで説明できてしまう──そして説明できることは、その説明が正しいことを意味しない。大きさは選別の極端さで決まる──中央 50 パーセントは +0.004677 しか動かず、下位 10 パーセントの 376 分の 1 である。極端な群を選んで前後を比べる設計が最も強い偽の効果を生み、しかも極端な群を選ぶのは「困っているところから手をつける」という自然な判断である。二回目を一回目と同一にすれば効果は厳密に 0 になる──雑音が原因であることの名指しである。分けているものは一つ──群を一回目の測定で選んだかどうかであって、処置の内容ではない。三行とも処置は同じ(何もしていない)で、違うのは群の選び方だけなのに、見かけの効果は +1.76 から -1.76 まで動く。だから無作為化が効く。現実に「困っている群」を選ばざるをえないなら、対照群を同じ基準で選ぶ──選び方をそろえれば回帰は両群で同じだけ起き、差には現れない。正直に言えば──実際の二回の観測には相関があり、あれば回帰はその分だけ小さくなる。本稿の 1.756597 は独立を仮定した上限側の値である。論文364 が扱った「見える集団の平均」とは根が違い、こちらはデータに欠けが無いぶん、記録を見ても異常が分からない。最後に一つ──「効果があった」と「効果が無かった」の両方を、同じ処置ゼロが作れる。どちらの結論を得たいかによって、選ぶ群を変えればよい。結論を選べる設計は、結論を支えない。設計を見ずに結果だけを見ると、その区別がつかない。 作成にあたって:本稿の着想と内容は、著者自身の考察に基づくものです。文章の構成整理や英訳、数式の確認には AI(大規模言語モデル)の助力を得ました。最終的な内容の解釈や誤りがあれば、それらはすべて著者の責に帰します。お気づきの点があれば、ご教示いただければ幸いです。

View source

Similar papers

#small language model Dataset Open access Oct 2026

Socratic guiding questions in synthetic arithmetic data: matched LoRA runs (revision v2)

Supporting data, adapters, predictions and code for the article *Low-Cost LoRA Fine-Tuning of Small Language Models for Multi-Step Arithmetic Reasoning* by Jake O'Grady, Asena Isik Gürhan, Chee Fong Ting and Effirul Ramlan (University of Galway). We generated 20,000 GSM8K-derived arithmetic problems with step-by-step s...

O'Grady, Jake, Gürhan, Asena Isik, Chee, Fong Ting et al. · 465 citations
#computer vision Open access Jun 2016

Software Development in Startup Companies: The Greenfield Startup Model

The results are packaged in the Greenfield Startup Model (GSM), which explains the priority of startups to release the product as quickly as possible, and the need to shorten time-to-market, by speeding up the development through low-precision engineering activities.

Carmine Giardino, Nicolò Paternoster, M. Unterkalmsteiner et al. · 178 citations · ⚡14
#computer vision Open access Oct 2016

Software Startups - A Research Agenda

Software startup companies develop innovative, software-intensive products within limited timeframes and with few resources, searching for sustainable and scalable business models.

M. Unterkalmsteiner, P. Abrahamsson, Xiaofeng Wang et al. · 157 citations · ⚡17
#machine learning Review Open access Oct 2016

“Failures” to be celebrated: an analysis of major pivots of software startups

This study conducts a case survey study based on the secondary data of the major pivots happened in 49 software startups, and demonstrates that customer need pivot is the most common among all pivot types.

Sohaib Shahid Bajwa, Xiaofeng Wang, Anh Nguyen-Duc et al. · 127 citations · ⚡15
#computer vision Review Open access May 2015

A survey study on major technical barriers affecting the decision to adopt cloud services

The comparison of adopter and non-adopter sample reveals three potential adoption inhibitor, security, data privacy, and portability, which underlines the importance of the technical and security perspectives for research investigating the adoption of technology.

Nattakarn Phaphoom, Xiaofeng Wang, S. Samuel et al. · 111 citations · ⚡8
#computer vision Open access Feb 2018

Lean Internal Startups for Software Product Innovation in Large Companies: Enablers and Inhibitors

This study investigates how Lean internal startup facilitates software product innovation in large companies and identifies its enablers and inhibitors, and shows the potential of the method-in-action framework to investigate the Lean startup approach in non-startup context.

Henry Edison, Nina M. Smørsgård, Xiaofeng Wang et al. · 78 citations · ⚡6

Related blog posts

MIT News · Artificial Intelligence Oct 15, 2026

AI model decodes the language cells use to communicate

A new computational approach reveals how cells communicate to assemble into tissues. This could accelerate stem cell engineering, improve disease models, and inform therapies.

Microsoft Research Blog Oct 15, 2026

Proof-based general-purpose computing with Magnetar

Zero-knowledge proofs can verify computing results without rerunning workloads or exposing sensitive data. Magnetar helps bring them into the mainstream with support for standard programming languages, major speed gains, and quantum-resistant security. The post Proof-based general-purpose computing with Magnetar appeared first on Microsoft Research.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.