📚 Likelihood of Global Governance: A Maximum Likelihood Approach | 全球治理的似然性：最大似然估计法
In the realm of statistical inference, the term "likelihood" takes on a precise mathematical meaning that is central to parameter estimation. This article explores the concept of maximum likelihood estimation (MLE) through an unconventional lens: assessing the probability that a country has achieved effective global governance. By framing governance success as a Bernoulli trial, we can bring to life the fundamental principles of the likelihood function, log-likelihood, and estimator properties. Whether you are preparing for Edexcel A-Level Mathematics or simply curious about how statisticians measure confidence in grand societal outcomes, this journey will illuminate the power of MLE.
在统计推断领域，“似然”一词具有精确的数学含义，是参数估计的核心。本文通过一个非常规的视角来探讨最大似然估计（MLE）：评估一个国家实现有效全球治理的概率。将治理成功构建为伯努利试验，我们可以生动地展现似然函数、对数似然以及估计量性质的基本原理。无论你是在为 Edexcel 数学 A-Level 做准备，还是仅仅好奇统计学家如何衡量宏大社会结果的置信度，这段旅程都将揭示 MLE 的强大力量。
1. What Is Likelihood? | 什么是似然？
In everyday language, likelihood is synonymous with probability, but in statistics they are distinct concepts. Probability refers to the chance of observing data given a known parameter. Likelihood, on the other hand, measures how well a particular parameter value explains the already-observed data. It is a function of the parameter with the data held fixed.
在日常用语中，似然与概率同义，但在统计学中它们是不同的概念。概率是指在已知参数下观测到数据的可能性。而似然则是衡量某个特定参数值对已观测数据的解释程度有多好。它是参数的函数，而数据保持固定。
inomial probability. The likelihood function L(p) is proportional to p⁷(1-p)³, and we ask: which value of p makes this result most plausible?
举一个简单的例子，如果我们抛一枚硬币 10 次得到 7 次正面，则在 p = 0.5 的条件下看到 7 次正面的概率是一个二项概率。似然函数 L(p) 正比于 p⁷(1-p)³，而我们追问的是：哪个 p 值使得这个结果最为合理？
2. Likelihood vs Probability | 似然与概率的区别
Probability is forward‑looking: given a parameter θ, we calculate P(Data | θ). Likelihood is backward‑looking: having observed Data, we evaluate L(θ | Data) = P(Data | kelihood is not a probability distribution over the parameter; it does not integrate to 1.
概率是前向的：给定参数 θ，我们计算 P(数据 | θ)。似然是后向的：在观测到数据后，我们评估 L(θ | 数据) = P(数据 | θ)，但将其解释为 θ 的函数。关键点在于，似然不是参数上的概率分布；它的积分不为 1。
This distinction is crucial when we build models for global governance. We might define θ as the true, unknown proportion of countries that satisfy governance criteria. Our observed data consists of a sample of countries, and we wish to estimate θ using maximum likelihood.
当我们为全球治理建立模型时，这一区别至关重要。我们可以将 θ 定义为满足治理标准的国家的真实而未知的比例。观测数据由一组国家样本构成，我们希望通过最大似然来估计 θ。
3. Formulating a Global Governance Model | 建立全球治理模型
pose a political science institute ranks 100 countries and declares each country as having "effective global governance" means failure. The probability of success, p, is assumed constant across countries and independent.
假设某个政治科学研究机构对 100 个国家进行排名，并宣布每个国家是否具备“有效的全球治理”（成功）或未达到（失败）。国家 i 的结果是一个二元变量 Xᵢ，其中 Xᵢ = 1 表示成功，Xᵢ = 0 表示失败。假设成功的概率 p 在不同国家间是常数且相互独立。
This is a classic Bernoulli trials setup. Although in reality governance outcomes may be correlated, the independence assumption simplifies the mathematics and is often used as a starting point for MLE applications.
这是一个经典的伯努利试验设定。尽管现实中治理结果可能相关，但独立性假设简化了数学，并且常被用作 MLE 应用的起点。
4. The Bernoulli Trial Framework | 伯努利试验框架
Each country i has the probability mass function P(Xᵢ = x) = pˣ (1-p)¹⁻ˣ for x ∈ or n independent countries, the joint probability of observing a specific sequence x₁, x₂, …, xₙ is the product of the individual probabilities:
每个国家 i 的概率质量函数为 P(Xᵢ = x) = pˣ (1-p)¹⁻ˣ，其中 x ∈ {0,1}。对于 n 个独立国家，观测到特定序列 x₁, x₂, …, xₙ 的联合概率是单个概率的乘积：
P(X₁=x₁, …, Xₙ=xₙ | p) = ∏ᵢ₌₁ⁿ pˣⁱ (1-p)¹⁻ˣⁱ = p∑xᵢ (1-p)n - ∑xᵢ
successes in the sample, often denoted by k. This expression is the foundation of the likelihood function.
这里 ∑xᵢ 表示样本中成功的总数，通常记作 k。这个表达式就是似然函数的基础。
5. Constructing the Likelihood Function | 构建似然函数
Given the observed data (k successes out of n), the likelihood function is defined as:
给定观测数据（n 个样本中出现 k 次成功），似然函数定义为：
L(p) = pᵏ (1-p)n-k
Note that we may omit the binomial coefficient because it does not depend on p and does not affect the maximisation. L(p) quantifies the relative plausibility of different values of p for the governance success probability.
注意我们可以省略二项式系数，因为它不依赖于 p，不影响最大化。L(p) 量化了不同 p 值对于治理成功概率的相对合理程度。
.7⁶⁸ × 0.3³². By computing and comparing these values, we can find the p that maximises L.
例如，若 n = 100，k = 68，则 L(0.6) = 0.6⁶⁸ × 0.4³²，而 L(0.7) = 0.7⁶⁸ × 0.3³²。通过计算并比较这些值，我们可以找到能使 L 取最大的 p。
6. Log-Likelihood for Simplicity | 对数似然简化运算
lly tiny for even moderate n, leading to computational underflow. Statisticians therefore work with the natural logarithm of the likelihood, denoted ℓ logarithm is a monotonic transformation, the value of p that maximises L(p) also maximises ℓ(p).
L(p) 中的乘积即使对于中等大小的 n 也会在数值上变得极小，导致计算下溢。因此统计学家使用似然的自然对数，记为 ℓ(p) = ln L(p)。由于对数是单调变换，使 L(p) 最大的 p 值同样使 ℓ(p) 最大。
For our Bernoulli model, the log-likelihood function is:
对于我们的伯努利模型，对数似然函数为：
ℓ(p) = k·ln(p) + (n-k)·ln(1-p)
This expression is much easier to differentiate and yields a simple equation for the maximum likelihood estimator.
这个表达式更容易求导，并能产生一个关于最大似然估计量的简单方程。
7. Deriving the Maximum Likelihood Estimator | 推导最大似然估计量
To find the maximum, we differentiate ℓ(p) with respect to p and set the derivative to zero:
为求最大值，我们对 ℓ(p) 关于 p 求导并令导数为零：
dℓ/dp = k/p - (n-k)/(1-p) = 0
ies to k - kp = np - kp, and hence k = np. Therefore, the maximum likelihood estimator is:
整理得到 k/p = (n-k)/(1-p)。交叉相乘得 k(1-p) = (n-k)p，化简为 k - kp = np - kp，从而 k = np。因此，最大似然估计量为：
p̂ = k/n
This result is intuitive: the best estimate for the probability of global governance success is simply the sample proportion of successes.
这个结果是直观的：对全球治理成功概率的最佳估计，就是样本中的成功比例。
8. Worked Example with Fictional Data | 使用虚构数据的实例
effective global governance. Thus, n = 50, k = 29. The MLE of the success probability is p̂ = 29/50 = 0.58.
设想一个智库随机评估了 50 个国家，发现其中 29 个表现出有效的全球治理。于是 n = 50，k = 29。成功概率的 MLE 为 p̂ = 29/50 = 0.58。
We can verify this using the log-likelihood function. The log-likelihood at p = 0.58 is ℓ = 29·ln(0.58) + 21·ln(0.42). If we test a rival value, say p = 0.5, ℓ(0.5) = 29·ln(0.5) + 21·ln(0.5) = 50·ln(0.5) ≈ -34.66, while ℓ(0.58) ≈ -33.91, which is higher, confirming that p̂ indeed maximises ℓ.
我们可以用对数似然函数验证这一点。在 p = 0.58 时，对数似然 ℓ = 29·ln(0.58) + 21·ln(0.42)。如果我们测试一个竞争值，比如 p = 0.5，则 ℓ(0.5) = 29·ln(0.5) + 21·ln(0.5) = 50·ln(0.5) ≈ -34.66，而 ℓ(0.58) ≈ -33.91，后者更大，证实 p̂ 确实使 ℓ 最大化。
The estimated probability 0.58 is the value that makes the observed data "most likely" under the Bernoulli model.
估计概率 0.58 是在伯努利模型下使观测数据“最可能”出现的值。
9. Properties of MLE: Unbiasedness and Consistency | MLE的性质：无偏性与一致性
The maximum likelihood estimator for a Bernoulli proportion possesses desirable large‑sample properties. It is consistent, meaning that as the sample size n increases, p̂ converges in probability to the true parameter p.
伯努利比例的最大似然估计量具有良好的大样本性质。它是一致的，即随着样本容量 n 增加，p̂ 会依概率收敛到真实参数 p。
It is also unbiased: the expectation E[p̂] = E[k/n] = (1/n)·E[k] = (1/n)·np = p. Therefore, on average, p̂ correctly estimates the global governance success rate.
它也是无偏的：期望 E[p̂] = E[k/n] = (1/n)·E[k] = (1/n)·np = p。因此，平均而言，p̂ 能正确估计全球治理的成功率。
These properties justify the use of the sample proportion as a reliable estimator when the independence and identical distribution assumptions hold reasonably well.
当独立同分布假设基本成立时，这些性质证明了使用样本比例作为可靠估计量的合理性。
10. Limitations of the Simple Bernoulli Model | 简单伯努利模型的局限
Our governance example assumes a constant success probability p for every country. In reality, governance outcomes are influenced by economic, cultural, and political factors, leading to heterogeneity. A more advanced model might incorporate covariates through logistic regression, where the likelihood principle remains central.
我们的治理示例假设每个国家的成功概率 p 为常数。现实中，治理结果受到经济、文化和政治因素的影响，导致异质性。更先进的模型可以通过 logistic 回归纳入协变量，而其中似然原理仍居于核心地位。
Additionally, the independence assumption may be violated if countries within the same region share common shocks. In such cases, clustered or hierarchical models with correlated likelihoods are employed.
此外，如果同一区域内的国家受到共同冲击的影响，独立性假设可能会被违反。这时会使用带有相关似然的聚类或分层模型。
11. Extending the Concept: Likelihood Ratio Tests | 概念延伸：似然比检验
nce success is as likely as a coin toss) versus H₁: p ≠ 0.5. The likelihood ratio test statistic is:
一旦我们有了 MLE，就可以比较关于全球治理的嵌套假设。例如，H₀: p = 0.5（治理成功概率相当于抛硬币）与 H₁: p ≠ 0.5。似然比检验统计量为：
Λ = L(p₀) / L(p̂)
Under certain regularity conditions, -2 ln Λ follows a chi-squared distribution, allowing us to assess the evidence against the null hypothesis.
在一定的正则条件下，-2 ln Λ 服从卡方分布，使得我们可以评估反对零假设的证据强度。
12. Conclusion: Inference Under Uncertainty | 结论：不确定性下的推断
The maximum likelihood framework transforms a seemingly abstract question—"How likely is global governance?"— e. By defining a probability model, collecting data, and constructing the likelihood function, we obtain an estimator p̂ rties. This methodology underpins countless applications in science, economics, and engineering. As you delve deeper into A-Level Statistics, remember that every likelihood calculation is a step towards quantifying the unknown and supporting evidence-based decisions in an uncertain world.
最大似然框架将一个看似抽象的问题——“全球治理的似然性有多高？”——转化为一种切实可操作的统计演练。通过定义概率模型、收集数据并构建似然函数，我们得到了具有可证明最优性的估计量 p̂ = k/n。这一方法论支撑着科学、经济学和工程领域的无数应用。当你在 A-Level 统计学中深入探索时，请记住，每一次似然计算都是朝着量化未知并在不确定的世界中支持循证决策迈出的一步。
Published by TutorHao | Mathematics Revision Series | aleveler.com
更多咨询请联系16621398022（同微信）
或截图后用「扫一扫 → 相册」选择本图