重大oj
You might see me twitch whenever I hear a lecturer, presenter, or researcher say “significant” when they clearly mean “statistically significant” (which in fact has nothing to do with being “clinically significant” or “scientifically significant”; see Wasserstein and Lazar, 2016). These cognitive hiccups snap me out of their otherwise careful scientific narrative — and make me worry that they might have fundamentally misunderstood their analysis results.
您可能会看到，每当我听到一个演讲，演示者或研究者说“显著”时，他们明确表示“统计显著”(我抽搐这实际上无关与被“临床显著”或“科学显著” ;见瓦瑟斯坦和拉扎(Lazar)，2016年 )。 这些认知上的困扰使我脱离了他们本应谨慎的科学叙述，并使我担心它们可能会从根本上误解了他们的分析结果。
This certainly isn’t something you want to hear from your statistician colleagues! So here’s my short attempt to help correct two common phrases about research findings. They’re wrong because of how they are commonly misinterpreted.
您当然不想从统计学家的同事那里听到这件事！ 因此，这是我的简短尝试，旨在帮助纠正有关研究发现的两个常见短语。 它们是错误的，因为它们通常会被误解。
I’ll share each wrong phrase, its interpretations, and a better phrase. If you’re interested, I provide detailed reasoning on Wrong Phrase 1 at the end. (As an exercise, you should try applying similar reasoning to Wrong Phrase 2.)
我将分享每个错误的短语，其解释和一个更好的短语。 如果您有兴趣，请在结尾处提供有关错误短语1的详细推理。 (作为练习，您应该尝试对错误短语2应用类似的推理。)
词组错误1：减少/增加 (Wrong Phrase 1: Decrease/Increase)
There was a significant decrease of D in the outcome.
结果D显着降低。
释义 (Interpretations)
Common Misinterpretation: There was a scientifically or clinically significant, important, meaningful, or useful decrease of D in the outcome.
常见的误解： D的科学或临床意义，重要性，意义或有用的降低。
A Correct Interpretation*: There was significant statistical evidence that a true, unknown change (i.e., decrease/increase) in the mean outcome exists that is not null (e.g., not zero). If the study was properly designed and executed, then this evidence suggests that the decrease of D we’d calculated from the sample is a good estimate of the true, unknown change of C.
正确解释*：有大量统计证据表明，平均结果中存在真实的，未知的变化(即减少/增加)，但不为零(例如，不为零)。 如果研究的设计和执行正确，则有证据表明我们从样本中计算出的D的减少是对C真实，未知变化的良好估计。
The relevant statistical test result says absolutely nothing about whether or not C is scientifically or clinically significant, important, meaningful, or useful (Wasserstein and Lazar, 2016). It only tells us that there was enough statistical evidence (which increases with sample size) to detect that C may not be null because our estimate D, which we assume is a good estimate of C, indicates a decrease.
有关的统计测试结果完全没有说明C在科学或临床上是否有意义，重要，有意义或有用 ( Wasserstein和Lazar，2016年 )。 它仅告诉我们，有足够的统计证据(随着样本量的增加而增加)来检测 C可能不为空，因为我们假设的估计值D(我们认为是对C的良好估计)表示减少了。
更好的短语 (A Better Phrase)
There was a detectable decrease of D in the outcome.
结果中D 明显减少。
Specifically, there was a statistically detectable decrease of D (which we assume is a good estimate of C) in the outcome sample average.
具体来说，结果样本平均值中D的统计可检测到的下降(我们假设是C的良好估计)。
错误短语2：关联 (Wrong Phrase 2: Association)
There was no significant association between variables X and Y.
变量X和Y之间没有显着关联。
释义 (Interpretations)
Common Misinterpretation: There was no scientifically or clinically significant, important, meaningful, or useful relationship between variables X and Y.
常见的误解：变量X和Y之间没有科学或临床上重要，重要，有意义或有用的关系。
Correct Interpretation: There was no significant statistical evidence that the true, unknown statistical association between variables X and Y is not null (e.g., zero correlation, a linear association). If the study was properly designed and executed, then this evidence suggests that the association Z we’d calculated from the sample (basically zero association), is a good estimate of the true, unknown association A.
正确解释：没有显着的统计证据表明变量X和Y之间的真实，未知统计关联不为空(例如，零关联，线性关联)。 如果研究的设计和执行正确，则有证据表明我们从样本中计算出的关联Z(基本上为零关联)是对真实，未知关联A的良好估计。
The relevant statistical test result says absolutely nothing about whether or not A is scientifically or clinically significant, important, meaningful, or useful (Wasserstein and Lazar, 2016). It only tells us that there wasn’t enough statistical evidence (which increases with sample size) to detect that A may not be null because our estimate Z, which we assume is a good estimate of A, indicates basically zero association.
相关的统计测试结果完全没有说明A在科学或临床上是否有意义，重要，有意义或有用 ( Wasserstein和Lazar，2016年 )。 它仅告诉我们没有足够的统计证据(随着样本量的增加而增加)来检测 A可能不为空，因为我们假设的估计值Z(假定为A的良好估计值)基本上表示零关联。
更好的短语 (A Better Phrase)
There was no detectable association between variables X and Y.
在变量X和Y 之间 没有 可检测的 关联 。
Specifically, there was no statistically detectable association measured as Z (which we assume is a good estimate of A) between variables X and Y.
具体来说，变量X和Y之间没有以Z为度量单位(我们认为是A的良好估计)的统计可检测关联。
详细推理：错误短语1 (Detailed Reasoning: Wrong Phrase 1)
The original phrase invokes three key quantities:
原始短语调用了三个关键量：
The true change of C in the mean outcome is unknown. This is what we would simply calculate if we had all relevant data from the target population.
平均结果中C的真实变化是未知的。 如果我们拥有目标人群的所有相关数据，这就是我们将简单计算的结果。
Instead, we had to estimate this quantity using a sample from that target population. Our particular estimate D may or may not approximate C.
取而代之的是，我们必须使用来自该目标人群的样本来估算此数量。 我们的特定估计值D可能会或可能不会接近C。
The amount of statistical evidence for a non-null (e.g., non-zero) C is defined like a “proof by contradiction” as follows. If C really was null (i.e., if there was really no change at all in the mean outcome), what is the probability of observing our estimate D or a more extreme value? This probability is called a p-value, and is generally approximated using your sample.
非零(例如，非零)C 的统计证据的数量如下定义为“矛盾证明”。 如果C确实为零(即，如果平均结果实际上根本没有变化)，那么观察我们的估计值D或更极端值的概率是多少？ 该概率称为p值 ，通常使用您的样本进行近似 。
For a given estimate like D, this p-value decreases as the sample size increases; i.e., an increasing “contradiction” of the null hypothesis. “Statistical significance” is loosely defined as some inverse of this p-value: The lower the p-value, the higher the statistical significance. That is, if we’d calculated that same exact value D from a larger sample, D would have been more statistically significant. If we’d calculated that same exact value D from a smaller sample, D would have been less statistically significant.
对于给定的估计值(例如D)，此p值随样本大小的增加而减小； 即，原假设的“矛盾”不断增加。 “统计显着性”大致定义为该p值的倒数：p值越低，统计显着性越高。 也就是说，如果我们从较大的样本中计算出相同的精确值D，则D在统计上会更有意义。 如果我们从较小的样本中计算出相同的准确值D，则D在统计上的重要性会降低。
The relevant statistical hypothesis test:
相关的统计假设检验：
Does answer: “How much evidence do we have for the existence of a non-null C?” Put differently, “Do we have enough evidence to detect a non-null C?”
是否回答：“我们有多少证据证明存在非null C？” 换句话说，“我们是否有足够的证据检测到非空C？”
Does not answer: “How much evidence do we have that C is scientifically or clinically significant, important, meaningful, or useful?” Again, remember that statistical significance is completely different from scientific/clinical significance. In particular, C can never be statistically significant because true quantities are fixed values that generally have nothing to do with the size of the samples used to estimate them.
没有回答：“我们有多少证据证明C具有科学或临床意义，重要性，意义或用途？” 再次提醒您， 统计意义与科学/临床意义完全不同 。 特别地， C 永远不会在统计上有意义，因为真实数量是固定值，通常与用于估计它们的样本大小无关。
- Does not answer: “Does D approximate C?” This is a property of the estimator (e.g., sample average, MLE estimator) and study design. 不回答：“ D大约等于C吗？” 这是估计量(例如样本平均值，MLE估计量)和研究设计的属性。
Here’s some loose intuition for that last bullet point.
这是最后一点的一些直觉。
- Apply your estimator to your entire target population (where nothing varies). This simply calculates some fixed (i.e., non-random) quantity T. 将您的估算器应用于整个目标人群(不变)。 这只是计算一些固定(即非随机)数量T。
- Good estimators are statistically consistent for T (i.e., increasingly unbiased for T with larger sample sizes). 好的估计量对于T在统计上是一致的(即，对于具有更大样本量的T，越来越无偏)。
But your estimator will always estimate something! In statistical machine learning, your clustering or prediction algorithm will always find a pattern if you tell it that one exists — even when there is none. (Thanks to Professor Daniela Witten for clearly demonstrating this, along with some brilliant solutions, in a recent talk.)
但是您的估算师将始终估算某些东西！ 在统计机器学习中，如果您告诉集群或预测算法存在某个模式，即使它根本不存在，它也总是会找到一种模式。 (感谢Daniela Witten教授在最近的一次演讲中清楚地展示了这一点以及一些出色的解决方案。)
- Ensuring T=C depends on how well you design and execute your study: how you formulate your scientific hypotheses, collect data, and conduct all analyses. 确保T = C取决于您设计和执行研究的水平：您如何制定科学的假设，如何收集数据以及进行所有分析。
So the common interpretation is wrong for two reasons:
因此，常见的解释是错误的，原因有两个：
- It confuses D (a known sample quantity) with C (an unknown population quantity). 它将D(已知样本数量)与C(未知总体数量)混淆。
- It mistakenly claims that your sample provides evidence that C is scientifically or clinically significant, important, meaningful, or useful. 它错误地声称您的样本提供了证据，证明C在科学或临床上是重要，重要，有意义或有用的。
*注意：错误词组1 (* Note: Wrong Phrase 1)
释义 (Interpretations)
Another Correct Interpretation: “There was significant statistical evidence that a true, unknown mean change (i.e., decrease/increase) in the outcome exists that is not null (e.g., not zero).”
另一个正确的解释是： “有大量统计证据表明，结果中存在真实的，未知的均值变化(即减少/增加)，但不为零(例如，不为零)。”
This is a subtle but important difference from the earlier interpretation, and can implicate different estimator variances. For example, in a pre-post or otherwise paired design, an estimator that averages first and then takes the difference will be more variable than one that subtracts first and then takes the average if outcome pairs are associated with each other.
与先前的解释相比，这是一个微妙但重要的区别，并且可能暗示不同的估计量方差。 例如，在事前或成对的设计中，如果结果对彼此关联，则先求平均值然后求差的估算器比先减去然后求平均值的估算器更具可变性。
The actual correct interpretation will depend on the study design and statistical hypotheses.
实际的正确解释将取决于研究设计和统计假设。
关于作者 (About the Author)
Dr. Daza is a biostatistician and health data scientist who develops causal inference methods for personal (n-of-1) digital health. | 🇺🇸🇵🇭 ericjdaza.com @ericjdaza linkedin.com/in/ericjdaza | statsof1.org @statsof1 @fsbiostats
Daza博士是一位生物统计学家和健康数据科学家，他开发了用于个人(n-1)数字健康的因果推理方法。 | 🇺🇸🇵🇭 ericjdaza.com @ericjdaza linkedin.com/in/ericjdaza | statsof1.org @ statsof1 @fsbiostats
翻译自: https://towardsdatascience.com/significant-you-really-mean-detectable-b3e8819e3491
重大oj
4068
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
