115 年 國立臺灣大學國際企業學系碩士班乙組《統計學(K)》
第 1 題
- A renowned investment bank in Frankfurt has designed a unique assessment exercise for its graduate
recruitment program. The exercise, called the "Strategic Selection Protocol," is conducted as follows:
Four identical envelopes are presented to the candidate. One envelope contains an invitation to join the
prestigious Mergers & Acquisitions team with a guaranteed first-year bonus of €150,000. The other three
envelopes each contain an assignment to the Operations department with a standard bonus structure. The
candidate has no initial information about which envelope contains the M&A invitation. The assessment
proceeds in stages:
Stage 1: The candidate selects one envelope but does not open it.
Stage 2: The managing director, who knows the contents of all envelopes, deliberately opens two of the
three remaining envelopes, always revealing two Operations assignments. If the candidate's initial selection
contains the M&A invitation, the director randomly chooses which two of the three Operations envelopes to
open. If the candidate's initial selection contains an Operations assignment, the director opens two of the
remaining Operations envelopes (leaving the M&A envelope and one Operations envelope unopened).
Stage 3: The candidate is offered the opportunity to either retain their original selection or switch to the
other unopened envelope.
Mr. Weber, a candidate from Munich, initially selects Envelope A. The managing director then opens
Envelopes C and D, both revealing Operations assignments. Mr. Weber must now decide whether to keep
Envelope A or switch to Envelope B.
If Mr. Weber switches to Envelope B, what is the probability that he receives the M&A invitation?
(A) 1/4
(B) 1/2
(C) 2/3
(D) 3/4
(E) 1/3
登入後即可作答並保存紀錄。
核心觀念
本題考查「蒙提霍爾問題(Monty Hall problem)」與條件機率。關鍵在於:
- Weber 最初選中 M&A 邀請的機率為 。
- 最初選中 Operations 的機率為 。
- 管理董事知道各信封內容,且一定會打開兩個 Operations 信封,因此開啟資訊不是隨機產生,而是帶有篩選作用。
- 當原先選擇不是 M&A 時,未被打開的另一個信封必定是 M&A。
因此,改選另一個未開啟信封的成功機率為原先選錯的機率,即 。
解題方法一:直接依最初選擇分類
最初選擇 Envelope A,可能有兩種情況:
-
Envelope A 是 M&A 邀請
其機率為
此時 C、D、B 都是 Operations。管理董事打開 C、D 後,B 仍是 Operations。
若 Weber 改選 B,必定失敗。
-
Envelope A 是 Operations
其機率為
此時 M&A 邀請必在 B、C、D 其中一個。管理董事必須打開兩個 Operations 信封,因此在 C、D 已經被打開且內容均為 Operations 的情況下,另一個未開啟信封 B 必定是 M&A。
若 Weber 改選 B,必定成功。
所以改選 B 的成功機率為
解題方法二:用條件機率驗證
令事件:
- :A 含有 M&A 邀請
- :B 含有 M&A 邀請
- :董事打開 C、D,且兩者均為 Operations
各種情況如下:
| M&A 所在信封 | 先驗機率 | 發生 的機率 | 聯合機率 |
|---|---|---|---|
| A | |||
| B | |||
| C | |||
| D |
第 2 題
- A multinational consulting firm based in Mumbai has designed an innovative case interview format to
assess candidates' persistence and analytical capabilities. The interview room contains a complex decision-
simulation system with four pathways, each represented by a door of identical appearance.
The pathways are engineered as follows:
Door A leads to immediate success (job offer) after a 15-minute problem-solving exercise.
Door B leads to immediate success (job offer) after a 25-minute problem-solving exercise.
Door C leads to a 40-minute exercise that returns the candidate to the starting room with no indication of
having returned.
Door D leads to a 60-minute exercise that returns the candidate to the starting room with no indication of
having returned.
Upon each return to the starting room, the doors are randomized in appearance. The candidate has no
memory of which doors were previously attempted and selects each door with equal probability 1/4,
independently of all previous selections.
Ms. Sharma enters the interview room at exactly 10:00 AM.
What is the expected total time (in minutes) for Ms. Sharma to receive a job offer?
(A) 35 minutes
(B) 70 minutes
(C) 90 minutes
(D) 120 minutes
(E) 140 minutes
登入後即可作答並保存紀錄。
本題考驗的是期望值的計算,特別是當過程包含隨機選擇、重複嘗試且有成功與失敗路徑時。由於每次選擇都是獨立且隨機的,且沒有記憶,這是一個典型的馬可夫鏈問題,但我們可以直接透過期望值的遞迴關係來求解。
首先,我們定義 為從起始房間獲得工作機會的預期總時間。
房間內有四扇門:A, B, C, D。
每次進入房間,有 1/4 的機率選擇任一扇門。
門 A:成功,耗時 15 分鐘。
門 B:成功,耗時 25 分鐘。
門 C:失敗,耗時 40 分鐘,回到起始房間。
門 D:失敗,耗時 60 分鐘,回到起始房間。
我們需要計算獲得工作機會的總預期時間。
設 為從起始點到達成功狀態的預期時間。
在起始點,有四種選擇,每種選擇的機率都是 1/4。
- 選擇門 A:耗時 15 分鐘,直接獲得工作機會。
- 選擇門 B:耗時 25 分鐘,直接獲得工作機會。
- 選擇門 C:耗時 40 分鐘,回到起始點。由於沒有記憶,回到起始點後,預期還需 分鐘才能成功。總時間為 。
- 選擇門 D:耗時 60 分鐘,回到起始點。回到起始點後,預期還需 分鐘才能成功。總時間為 。
根據期望值的定義,我們可以寫出關於 的方程式:
將上述時間代入:
現在,我們解這個方程式來求 :
將 移到等號左邊:
因此,Ms. Sharma 獲得工作機會的預期總時間是 70 分鐘。
我們來驗證一下這個結果是否合理。
如果只有門 C 和 D,平均耗時是 (40+60)/2 = 50 分鐘,且每次都回到原點。
如果只有門 A 和 B,平均耗時是 (15+25)/2 = 20 分鐘。
這裡有成功的門和失敗的門。失敗的門會增加時間。
第 3 題
- A leading intellectual property law firm headquartered in Boston specializes in patent litigation for
technology companies. The firm's managing partners are developing a revenue forecasting model for the
upcoming fiscal year based on their historical case data.
The firm's analytics team has established the following stochastic model:
Case Acquisition: The number of new patent litigation cases assigned to a randomly selected senior partner
during a fiscal year follows a Poisson distribution with parameter cases.
Case Success Rate: Each case, independently of all other cases, has a probability of resulting in a
"favorable verdict"-defined as either a court ruling in favor of the client or a settlement exceeding the
client's initial demands.
The firm's CFO is preparing financial projections for the partnership meeting. He explains to a junior
associate: "Let N denote the number of cases assigned to a partner, and let X denote the number of favorable
verdicts achieved by that partner. We need to understand both the expected number of favorable verdicts and
the variability around that expectation."
The junior associate responds: "I calculated E[X] = p = 2.1. For the variance, since X|N follows a Binomial
distribution, I used the law of total variance. My calculation shows Var(X) = p(1-p) = 1.365."
The CFO replies: "Your expected value is correct, but check your variance calculation again. You're missing
something important."
Which of the following correctly states E[X] and Var(X)?
(A) E[X] = 2.10 and Var(X) = 1.365
(B) E[X] = 2.10 and Var(X) = 2.10
(C) E[X] = 2.10 and Var(X) = 2.835
(D) E[X] = 3.90 and Var(X) = 2.10
(E) E[X] = 2.10 and Var(X) = 1.735
登入後即可作答並保存紀錄。
本題考驗的是複合隨機變數的期望值與變異數的計算,特別是當一個隨機變數的分布依賴於另一個隨機變數時。此處,有利判決的數量 的分布取決於案件總數 。
首先,我們知道:
- 案件總數 服從參數為 的 Poisson 分布。即 。
- 在給定案件總數 的情況下,有利判決的數量 服從二項分布,其成功機率為 。即 。
計算期望值 E[X]:
我們使用全期望值法則 (Law of Total Expectation):。
對於二項分布 ,其條件期望值為 。
所以,。
由於 是隨機變數, 是常數,我們可以將 提出期望值運算:
。
我們知道 ,其期望值為 。
因此,。
這與 CFO 的說法一致:「Your expected value is correct」。
計算變異數 Var(X):
我們使用全變異數法則 (Law of Total Variance):。
首先,計算 :
對於二項分布 ,其條件變異數為 。
所以,。
由於 是常數,我們可以提出:
。
我們已經知道 。
第 4 題
- A large health insurance company headquartered in Amsterdam launched a voluntary "Digital Health
Engagement Program" (DHEP) designed to encourage policyholders to adopt healthier lifestyles through a
smartphone application. The program offers personalized fitness recommendations, nutrition tracking, and
virtual health coaching.
The company's actuarial division conducted a two-year retrospective study to evaluate the program's impact
on healthcare costs. At the program's launch, all 120,000 eligible policyholders (ages 30-55, no pre-existing
chronic conditions) were invited to participate. Enrollment was entirely voluntary.
28,000 policyholders enrolled in and actively used the DHEP.
92,000 policyholders did not enroll in the program.
The actuarial team measured each policyholder's "Annual Claims Ratio" (ACR)-the ratio of actual medical
claims to expected claims based on demographic factors:
Group
DHEP Participants
Non-Participants
Mean ACR
0.72
0.94
The Chief Actuarial Officer initially reports to the executive board: "The DHEP program reduces the claims
ratio by 0.22, representing a 23% reduction in healthcare costs. This demonstrates exceptional ROI for our
wellness investment."
However, further analysis of pre-program characteristics reveals that participants had significantly lower
baseline BMI (24.2 vs. 28.7), higher gym membership rates (62% vs. 18%), lower smoking rates (8% vs.
24%), and higher household income (78,000).
Which of the following BEST explains why the observed 0.22 reduction in claims ratio cannot be
interpreted as the causal effect of the DHEP program?
(A) Measurement error in the claims ratio creates systematic bias that inflates the apparent program effect.
(B) Confounding exists because policyholders who chose to enroll were already healthier and more health-
conscious, and would have had lower claims regardless of program participation.
(C) The Hawthorne effect causes participants to temporarily change behavior due to being observed, not due
to the program itself.
(D) Survivorship bias occurs because policyholders who dropped out of the program are excluded from the
analysis.
(E) Ecological fallacy occurs because group-level statistics cannot be applied to individual policyholders.
登入後即可作答並保存紀錄。
核心觀念
本題考查「觀察性研究中的混淆因子(confounding)」與「相關不等於因果」。
若要判斷 DHEP 是否造成 ACR 降低,理想上應比較同一批人在兩種狀態下的結果:
但同一位保戶不可能同時參加與不參加,因此實務上必須比較兩組不同保戶。此時,若兩組在參加計畫以前就存在重要差異,組間 ACR 差異便不一定來自 DHEP。
本題中,參與者與未參與者在計畫開始前即有下列差異:
- 參與者平均 BMI 較低;
- 健身房會員比例較高;
- 吸菸率較低;
- 家戶所得較高。
這些變數同時會影響「是否選擇參加 DHEP」與「醫療理賠成本」,因此屬於混淆因子。
解題方法
先計算兩組的觀察差異:
表示參與者的平均 ACR 比未參與者低 。
若以未參與者作為比較基準,觀察上的相對降低比例為:
因此約為 ,題目所稱的 是由此計算而來。
但是,這個 只是兩組的「觀察差異」,不是必然的「因果效果」。原因在於參與 DHEP 並非隨機分派,而是由保戶自行決定。較健康、較重視健康且所得較高的人,本來就更傾向參加計畫,也本來就可能有較低的醫療理賠。
可用下列概念表示:
本題缺乏隨機分派,因此無法單憑 與 證明 DHEP 導致 ACR 降低。正確解釋是:健康特徵造成的組間差異,混入了 DHEP 的觀察效果。
選項分析
(A) Measurement error in the claims ratio creates systematic bias that inflates the apparent program effect.
錯誤。測量誤差是指 ACR 的實際值與記錄值之間存在誤差,例如理賠資料登錄錯誤或計算方式不一致。題目沒有提供 ACR 存在測量錯誤的資訊,反而明確指出兩組在計畫開始前的健康與社經特徵不同。因此本題的主要問題是選擇參與造成的混淆,而非測量誤差。
(B) Confounding exists because policyholders who chose to enroll were already healthier and more health-conscious, and would have had lower claims regardless of program participation.
正確。參與者在計畫開始前已具有較低 BMI、較高健身房會員率、較低吸菸率與較高家戶所得,顯示他們原本就比未參與者更健康、也更重視健康。
第 5 題
- A prestigious business school in Chicago offers an optional "Executive Mentorship Program" (EMP)
connecting current MBA students with alumni executives. The career services office conducted a study to
evaluate the program's impact on post-graduation salary outcomes.
Observed Outcomes (Base Salary, in thousands USD):
Group
EMP Participants (180 students)
Non-Participants (420 students)
Mean Salary
142,000
Natural Experiment: Due to a scheduling system malfunction during the registration period, a random subset
of 75 students who attempted to register for EMP were unable to complete enrollment before the deadline.
This technical failure was completely random and unrelated to student characteristics.
Outcome Data for System-Affected Students (None received mentorship):
Registration Attempt
Tried to Register (25 students)
Did NOT Try to Register (50 students)
Mean Salary
138,000
Using the decomposition: Observed Difference = True Causal Effect (ATT) + Selection Bias
Using the natural experiment data, what is the magnitude of the selection bias in the original observational
comparison?
(A) 14,000
(C) 26,000
(E) $30,000
登入後即可作答並保存紀錄。
本題考驗的是如何利用「自然實驗」來估計「選擇性偏差」(Selection Bias),並進而得出真實的因果效應。題目給了一個觀察性研究的結果,以及一個隨機事件(系統故障)所產生的「準實驗」數據。
我們有以下資訊:
-
原始觀察性研究:
- EMP 參與者 (180人):平均薪資 $168,000
- 非參與者 (420人):平均薪資 $142,000
- 觀察性研究的薪資差異 (Observed Difference) = 142,000 = $26,000。
-
自然實驗 (系統故障):
- 75 名學生嘗試註冊 EMP,但因系統故障未能完成。
- 這 75 人是隨機抽樣的,與其特徵無關。
- 這 75 人中,有 25 人「嘗試註冊」(Tried to Register),但未成功。這群人可以視為「想參加 EMP 但未成功」的群體。
- 有 50 人「未嘗試註冊」(Did NOT Try to Register)。這群人可以視為「根本不想參加 EMP」的群體。
- 「嘗試註冊但未成功」者的平均薪資 = $156,000。
- 「未嘗試註冊」者的平均薪資 = $138,000。
-
分解公式:
Observed Difference = True Causal Effect (ATT) + Selection Bias
目標: 計算原始觀察性比較中的 Selection Bias。
思路:
- 原始觀察性研究的差異 () 包含了真實的因果效應 (ATT) 和選擇性偏差。
- 自然實驗的部分,由於「嘗試註冊但未成功」的 25 人,他們的薪資是 ,而「未嘗試註冊」的 50 人薪資是 $138,000
第 6 題
- A national retail chain headquartered in Toronto is evaluating the effectiveness of a new. employee training
program designed to improve customer service quality. The program, called "Service Excellence Initiative"
(SEI), consists of intensive workshops on customer engagement, conflict resolution, and upselling
techniques.
The company's HR Analytics team conducted a six-month observational study across 200 retail locations.
The primary outcome measure is the "Customer Satisfaction Index" (CSI)-a composite score from 0 to 100
based on customer feedback surveys.
Observed Adoption and Outcomes:
Group
SEI Participants (500 employees)
Non-Participants (1,500 employees)
Mean CSI
84.5
71.5
Instrumental Variable: During the study period, 80 of the 200 retail locations received a corporate mandate
requiring store managers to allocate dedicated training hours for staff development. This mandate was issued
based on regional distribution center logistics schedules established two years prior completely
independent of individual employee or store characteristics.
Data Summary by IT Upgrade Status:
WITH Mandate
WITHOUT Mandate
SEI participation rate
45%
15%
Mean CSI
78.9
72.9
The LATE formula is: LATE = (Difference in outcomes by instrument) / (Difference in participation rates by
instrument)
Based on the instrumental variable analysis, what is the estimated true causal effect (LATE) of SEI
participation on Customer Satisfaction Index?
(A) 6.0 points
(B) 13.0 points
(C) 15.0 points
(D) 20.0 points
(E) 24.0 points
登入後即可作答並保存紀錄。
本題考驗的是工具變數法 (Instrumental Variable, IV) 的應用。工具變數法的目的是在存在內生性(例如選擇性偏差)的情況下,估計處理效應的因果關係。
核心概念:
工具變數 必須滿足三個條件:
- 相關性 (Relevance): 必須與內生變數 (在此是 SEI 參與率) 相關。
- 外生性 (Exogeneity): 對應變數 (在此是 CSI) 的影響,只能透過 傳遞,不能直接影響 。
- 獨特性 (Exclusion Restriction): 不能直接影響 ,除非透過 。
題目中,工具變數 (Instrument) 是「公司是否發布了強制培訓時數的指令 (Mandate)」。
內生變數 (Endogenous Variable) 是「員工是否參與 SEI 計畫」(SEI participation rate)。
應變數 (Outcome Variable) 是「客戶滿意指數 (CSI)」。
數據整理:
-
總體觀察數據:
- SEI 參與者:500 人,Mean CSI = 84.5
- 非參與者:1500 人,Mean CSI = 71.5
- 觀察性差異 (Observed Difference in CSI) = 。這個差異包含了真實的因果效應和選擇性偏差。
-
工具變數數據 (依據是否有 Mandate 分組):
- 有 Mandate 的地點 (WITH IT Upgrade):
- SEI 參與率 = 45%
- Mean CSI = 78.9
- 沒有 Mandate 的地點 (WITHOUT IT Upgrade):
- SEI 參與率 = 15%
- Mean CSI = 72.9
- 有 Mandate 的地點 (WITH IT Upgrade):
第 7 題
- A leading pharmaceutical corporation headquartered in Basel, Switzerland, is conducting a critical Phase
III clinical trial for a novel oncology drug designated as XR-7721. The drug's efficacy is measured by the
"Tumor Response Index" (TRI), a continuous metric ranging from 0 to 1, where higher values indicate
greater tumor shrinkage. Regulatory agencies require that the drug demonstrate a median TRI of at least 0.5
and a mean TRI of at least 0.45 to receive approval for market authorization.
The corporation's biostatistics division has developed a probability model based on preliminary Phase II
data. They propose that the TRI for a randomly selected patient follows a continuous distribution with
probability density function:
for , and 0 otherwise
where and are constants. Extensive pharmacokinetic modeling suggests that the parameter equals 2 for
the target patient population.
The biostatistics team presents the following assertions to the executive committee ahead of the regulatory
submission:
(I) For the density function to be valid when , the normalizing constant must equal 3/5.
(II) Under this model, the probability that a randomly selected patient exhibits a TRI below 0.5 is 11/40.
(III) The expected value of the TRI under this model is 3/5, which exceeds the regulatory threshold of 0.45.
(IV) The cumulative distribution function evaluated at can be expressed as .
(V) If the regulatory agency instead required that for approval, the drug would fail to
meet this criterion under the proposed model.
Which of the above assertions are CORRECT?
(A) I, II, and IV only
(B) I, III, and V only
(C) I, III, IV, and V only
(D) II, III, and IV only
(E) I, II, III, and V only
登入後即可作答並保存紀錄。
本題要求我們驗證五個關於給定機率密度函數 (PDF) 的敘述是否正確。PDF 為 ,其中 ,且定義域為 。
首先,我們將 代入 PDF:,對於 。
驗證敘述 (I):正規化常數 c
一個有效的 PDF 必須滿足 。
對於此問題,積分範圍是 。
。
敘述 (I) 是 正確 的。
驗證敘述 (II):P(TRI < 0.5)
TRI < 0.5 表示 。我們要計算 。
使用 和 :。
。
將 7/20 轉換成分數:。
敘述 (II) 說機率是 11/40。
。
。
所以,。
敘述 (II) 是 錯誤 的。
驗證敘述 (III):期望值 E[X]
期望值 。
。
。
監管機構要求的平均 TRI 是 0.45。
。
敘述 (III) 是 正確 的。
驗證敘述 (IV):累積分布函數 F(0.5)
累積分布函數 (CDF) 。
第 8 題
- A multinational management consulting firm based in London operates a performance-based
compensation system for its associates. Each associate's quarterly bonus is determined by a "Client Impact
Score" (CIS), a discrete random variable that reflects the measurable value delivered to clients during the
quarter. The firm's compensation committee has analyzed historical performance data and established that
the CIS for a randomly selected associate follows the probability distribution shown below:
CIS (X)
P(X=x)
1
0.10
2
p
3
0.30
4
q
5
0.15
The committee's actuarial analysis indicates that E[X] = 3.05.
The firm employs a tiered bonus structure across its three regional offices. Let X denote the CIS for a
randomly selected associate. The quarterly bonus (in thousands of GBP) is calculated as follows:
London office:
Singapore office:
New York office:
The Chief Financial Officer is evaluating whether to standardize the bonus structure globally. After solving
for the unknown parameters p and q, the CFO asks an analyst to compare the three compensation schemes.
Which of the following statements is TRUE?
(A) The New York office has the highest expected bonus, and the London office has the highest variance in
bonus payments.
(B) The Singapore office has the highest expected bonus, and the New York office has the highest variance
in bonus payments.
(C) All three offices have the same expected bonus, but the New York office has the highest variance in
bonus payments.
(D) The London office has the highest expected bonus, and the Singapore office has the lowest variance in
bonus payments.
(E) The Singapore office has the highest expected bonus, and the London office has the lowest variance in
bonus payments.
登入後即可作答並保存紀錄。
核心觀念
本題考查:
- 機率分配的總和為 。
- 離散型隨機變數的期望值:
- 線性轉換的期望值與變異數:
其中,常數項 會影響期望值,但不影響變異數;變異數只會受到係數 的平方影響。
解題方法
一、求出 與
由機率總和為 :
因此:
又已知 :
整理得:
聯立
將第一式乘以 :
相減可得:
因此:
二、求 的變異數
先計算 :
因此:
三、比較三個辦公室的期望獎金
London
Singapore
New York
所以期望獎金由高至低為:
Singapore 辦公室的期望獎金最高。
四、比較三個辦公室的變異數
利用:
第 9 題
- A prestigious private equity firm in Hong Kong has developed an unconventional final-round interview
process designed to assess candidates' decision-making capabilities under uncertainty. The process, known
internally as the "Strategic Pivot Challenge," operates as follows:
Three identical briefcases are presented to the candidate. One briefcase contains a binding offer letter for a
senior associate position with a signing bonus of HK$2,000,000. The other two briefcases each contain a
polite rejection letter. The candidate has no initial information regarding which briefcase contains the offer.
The interview proceeds in stages:
Stage 1: The candidate selects one briefcase but does not open it.
Stage 2: The managing director, who knows the contents of all briefcases, deliberately opens one of the two
remaining briefcases, always revealing a rejection letter. Crucially, if the candidate's initial selection contains
the offer, the director randomly chooses which of the two rejection-containing briefcases to open with equal
probability. If the candidate's initial selection contains a rejection, the director opens the only other briefcase
that contains a rejection.
Stage 3: The candidate is offered the opportunity to either retain their original selection or switch to the
other unopened briefcase.
A candidate named Ms. Chen initially selects Briefcase A. The managing director then opens Briefcase B,
revealing a rejection letter.
However, this year the firm introduces a modification to assess probabilistic reasoning at a deeper level.
Before Ms. Chen decides whether to switch, she is informed that a junior analyst-who was unaware of the
briefcase contents entered the room during Stage 2 and observed only that Briefcase B was opened and
contained a rejection letter. The junior analyst did not observe Ms. Chen's initial selection.
From the perspective of the junior analyst, what is the probability that Briefcase C contains the offer letter?
(A) 1/3
(B) 1/2
(C) 2/3
(D) 3/4
(E) The probability cannot be determined without knowing Ms. Chen's initial selection.
登入後即可作答並保存紀錄。
核心觀念
本題考查:
- 條件機率與貝氏定理;
- 資訊集合不同,所得到的條件機率也不同;
- Monty Hall 問題中「主持人知道答案,且開啟方式具有規則」所造成的機率更新。
關鍵在於:陳小姐知道自己的初始選擇是 A,因此可直接計算 ;但 junior analyst 不知道陳小姐最初選了哪一個箱子,因此必須把初始選擇一併視為未知事件來計算。
假設陳小姐初始選擇 A、B、C 的機率相同,且 offer 位於三個箱子的機率也相同,皆為 。
解題方法:以貝氏定理分情況計算
令:
- 分別表示陳小姐初始選擇 A、B、C;
- 分別表示 offer 位於 A、B、C;
- 表示「主管開啟 B,且 B 內是拒絕信」。
要求的是:
由於 junior analyst 不知道初始選擇,必須考慮三種初始選擇。
情況一:初始選擇 A
若主管開啟 B 且 B 是拒絕信:
- 若 ,B、C 都是拒絕信,主管以 機率開 B;
- 若 ,B 是唯一可開啟的拒絕信,主管必定開 B;
- 若 ,B 內是 offer,不可能發生事件 。
因此:
其中「offer 在 C 且事件 發生」的聯合機率為:
情況二:初始選擇 B
主管只能開啟 A 或 C,不會開啟陳小姐已選的 B,因此:
情況三:初始選擇 C
若主管開啟 B:
- 若 ,B 是唯一的拒絕信,主管必定開 B;
- 若 ,A、B 都是拒絕信,主管以 機率開 B;
- 若 ,B 內是 offer,不可能發生 。
因此:
其中「offer 在 C 且事件 發生」的聯合機率為:
計算事件 的總機率
由全機率定理:
代入各項:
第 10 題
- A global management consulting firm has designed an elaborate assessment center for its final-round
recruitment process at its headquarters in Boston. The centerpiece of this assessment is an architectural maze
constructed in the basement level of the building, intended to evaluate candidates' resilience, decision-
making under uncertainty, and psychological composure.
The maze operates as follows: A candidate enters and immediately encounters a junction with three visually
indistinguishable corridors-labeled internally by the facility engineers as Corridor X, Corridor Y, and
Corridor Z, though candidates cannot differentiate among them. The corridors are engineered such that:
• Corridor X leads directly to the exit after exactly 20 minutes of walking.
• Corridor Y leads directly to the exit after exactly 35 minutes of walking.
• Corridor Z is a deceptive loop that returns the candidate to the original junction after exactly 110
minutes of walking, with no indication that they have returned to their starting point.
The junction room is designed with rotating architecture such that, upon returning via Corridor Z, the three
corridors appear in a completely randomized configuration. The candidate retains no memory of which.
corridor was previously attempted, as all corridors are identical in appearance. At each visit to the junction,
the candidate selects one of the three corridors with equal probability 1/3, independently of all previous
selections.
Mr. Rodriguez, an MBA candidate from São Paulo, enters the maze at exactly 9:00 AM.
What is the expected total time (in minutes) for Mr. Rodriguez to exit the maze?
(A) 55 minutes
(B) 65 minutes
(C) 80 minutes
(D) 82.5 minutes
(E) 110 minutes
登入後即可作答並保存紀錄。
本題考驗的是期望值的計算,特別是當存在一個可能無限循環的路徑時。這是一個典型的馬可夫鏈問題,但我們可以使用期望值的遞迴關係來求解。
我們需要計算從進入迷宮(起始點)到成功離開迷宮的預期總時間。
設 為從起始點到達出口的預期總時間。
在起始點,有三條走廊:X, Y, Z。每條走廊被選擇的機率是 1/3。
- 選擇走廊 X: 耗時 20 分鐘,直接到達出口。
- 選擇走廊 Y: 耗時 35 分鐘,直接到達出口。
- 選擇走廊 Z: 耗時 110 分鐘,回到起始點。由於候選人沒有記憶且走廊會隨機化,回到起始點後,預期還需要 分鐘才能成功。因此,總時間為 。
根據期望值的定義,我們可以寫出關於 的方程式:
將上述時間代入:
現在,我們解這個方程式來求 :
第 11 題
- A premier venture capital firm based in Menlo Park, California, has developed a sophisticated
quantitative framework for modeling its investment pipeline. The firm's analytics team, composed of former
quantitative researchers from leading hedge funds and technology companies, has been tasked with
forecasting the expected number of successful exits for the upcoming fiscal year to guide capital reserve
planning and limited partner communications.
The firm's historical data, accumulated over two decades of investment activity across multiple fund
vintages, supports the following stochastic model:
Deal Flow Model: The number of qualified investment opportunities that reach the final due diligence stage
for a randomly selected partner during a fiscal year follows a Poisson distribution with parameter
deals.
Conversion Rate: Each deal that reaches final due diligence, independently of all other deals, has a
probability of ultimately becoming a "successful exit" defined as an IPO or acquisition that
returns at least 3x the invested capital within seven years.
The firm's Managing Director is preparing for the annual Limited Partners Advisory Committee meeting.
She convenes the analytics team to address the following question:
"Consider a randomly selected partner from our firm. Let N denote the number of deals reaching final due
diligence for this partner during the fiscal year, and let X denote the number of these deals that ultimately
become successful exits. Our LP from the Yale Endowment-herself a former statistician-will certainly ask
about the variance of X, not just its expected value."
A junior analyst, recently graduated from a top MBA program, quickly calculates: "The expected number of
successful exits per partner is simply . For the variance, I applied the law of total variance and
obtained Var(X) = E[Var(X|N)] = p(1-p) = 0.72."
The Chief Risk Officer shakes his head. "You've forgotten a term. The law of total variance has two
components."
Which of the following correctly states E[X] and Var(X)?
(A) E[X] = 0.90 and Var(X) = 0.72
(B) E[X] = 0.90 and Var(X) = 0.90
(C) E[X] = 0.90 and Var(X) = 1.08
(D) E[X] = 4.50 and Var(X) = 0.90
(E) E[X] = 3.60 and Var(X) = 0.72
登入後即可作答並保存紀錄。
本題與第三題類似,考驗的是複合隨機變數的期望值與變異數計算。這裡,案件總數 服從 Poisson 分布,而成功次數 在給定 的情況下服從二項分布。
已知資訊:
- 案件總數 服從 Poisson 分布,參數為 。即 。
- 在給定案件總數 的情況下,成功次數 服從二項分布,成功機率為 。即 。
計算期望值 E[X]:
使用全期望值法則:。
對於 ,有 。
所以,。
對於 ,有 。
因此,。
這與 junior analyst 的計算一致:「The expected number of successful exits per partner is simply 」。
計算變異數 Var(X):
使用全變異數法則:。
- 計算 :
對於 ,有 。
所以,。
。
, 。
。
第 12 題
- A major retail banking corporation headquartered in Singapore is evaluating the effectiveness of a new
financial literacy program designed to increase customers' adoption of long-term investment products. The
program, branded "Wealth Wise Academy," consists of a series of online modules covering topics such as
compound interest, portfolio diversification, risk-adjusted returns, and retirement planning.
The bank's Customer Analytics division conducted a study over a twelve-month period. At the beginning of
the study, the program was made available to all 48,000 retail banking customers in the target demographic
(ages 28-45, annual income above SGD 80,000). Customers could voluntarily choose whether to enroll. By
the end of the study period:
• 12,000 customers enrolled in and completed the Wealth Wise Academy program.
• 36,000 customers did not enroll in the program.
The analytics team measured each customer's "Investment Portfolio Growth Rate" (IPGR)-the percentage
increase in total invested assets during the twelve-month period:
Group
Program Participants
Non-Participants
Mean IPGR
18.7%
11.2%
The Chief Marketing Officer initially reports: "The Wealth Wise Academy program caused a 7.5 percentage
point increase in customers' investment portfolio growth."
The Chief Risk Officer reviews pre-program characteristics and finds that participants had significantly
higher prior-year IPGR (14.2% vs. 8.6%), higher financial literacy scores (72.4 vs. 51.8), higher proportion
with graduate degrees (68% vs. 34%), and more prior investment accounts (3.8 vs. 1.9).
Which of the following BEST explains why the observed 7.5 percentage point difference cannot be
interpreted as the causal effect of the Wealth Wise Academy program?
(A) The sample sizes are unequal between the two groups, which biases the comparison of means.
(B) Selection bias arises because customers who chose to enroll were already healthier and more health-
conscious, and would have had lower claims regardless of program participation.
(C) The study lacks statistical power because the standard deviations are too large relative to the observed
difference.
(D) The comparison is invalid because the IPGR is influenced by macroeconomic factors that affected all
customers equally.
(E) Regression to the mean explains the entire observed difference.
登入後即可作答並保存紀錄。
本題考驗的是在觀察性研究中,如何區分相關性與因果性,以及可能存在的選擇性偏差 (selection bias)。題目中描述了一個自願性計畫 (Wealth Wise Academy),參與者與未參與者在計畫開始前就存在顯著差異。
觀察到的結果是:計畫參與者平均投資組合增長率 (IPGR) 為 18.7%,未參與者為 11.2%。
這表示參與者比未參與者高了 個百分點。
總體而言,這似乎顯示該計畫提高了客戶的投資組合增長率。
然而,題目進一步指出:
參與者在計畫開始前就具有更優越的特徵:
- 先前一年的 IPGR 更高 (14.2% vs 8.6%)
- 財務知識分數更高 (72.4 vs 51.8)
- 擁有大學學歷的比例更高 (68% vs 34%)
- 擁有更多先前投資帳戶 (3.8 vs 1.9)
這些差異是「在計畫開始前」就存在的。這些特徵本身就可能與較高的投資組合增長率有關。例如,本來就對投資更感興趣、更有知識、更有經驗的客戶,更有可能參與此計畫,並且他們本來就可能獲得更高的投資增長率,無論是否參與計畫。
這就引出了「選擇性偏差」(selection bias)。由於參與者是自願選擇的,他們可能與未參與者在未觀察到的(或已觀察到但未控制的)特徵上存在系統性差異,而這些特徵同時也影響了應變數 (IPGR)。
分析選項:
(A) The sample sizes are unequal between the two groups, which biases the comparison of means.
樣本數不均 (12,000 vs 36,000) 本身不會直接導致比較有偏差,除非樣本數不均是因為某種偏差造成的。在統計推斷中,樣本數不均是常見的,並且有方法可以處理(例如 t 檢定)。這個選項不是最根本的解釋。
(B) Selection bias arises because customers who chose to enroll were already healthier and more health-
conscious, and would have had lower claims regardless of program participation.
這個選項的措辭與問題 4 的選項 (B) 非常相似,但這裡的例子是關於「健康」而非「投資」。
正確的說法應該是:參與者因為本身就更積極、更有知識、更有經驗於投資,所以他們原本就有較高的 IPGR,無論是否參與計畫。這正是選擇性偏差的經典例子。
"customers who chose to enroll were already [more investment-savvy/active/knowledgeable], and would have had higher [IPGR] regardless of program participation."
這個選項雖然提到了「健康」和「較低的 Claims Ratio」,這與本題的投資背景不符,但其核心概念「選擇性偏差」是正確的,且描述了這種偏差發生的原因(參與者本身就具備有利於結果的特徵)。
第 13 題
- A leading technology conglomerate based in Shenzhen, China, implemented a voluntary "Executive
Leadership Acceleration Program" (ELAP) for its mid-level managers. The primary outcome measure is the
"Performance Rating Improvement" (PRI)-the change in annual performance rating (on a 100-point scale)
from the year before program availability to two years after.
Observed Outcomes:
Group
ELAP Participants (400 managers)
Non-Participants (1,600 managers)
Mean PRI
+12.6 points
+5.4 points
Natural Experiment: Due to an administrative error during enrollment, a randomly selected subset of 300
managers experienced a system glitch that prevented them from accessing the enrollment portal. This glitch
was completely random and unrelated to any manager characteristics.
Outcome Data for Glitch-Affected Managers (None received treatment):
Intention
Intended to Enroll (60 managers)
Did NOT Intend to Enroll (240 managers)
Mean PRI
+10.8 points
+4.2 points
The decomposition framework is: Observed Difference = True Causal Effect (ATT) + Selection Bias.
Using the natural experiment data to estimate the counterfactual, what is the magnitude of the selection bias
in the original observational comparison?
(A) 1.8 points
(B) 4.2 points
(C) 5.4 points
(D) 6.6 points
(E) 7.2 points
登入後即可作答並保存紀錄。
本題考驗的是如何利用「自然實驗」來估計「選擇性偏差」(Selection Bias),並進而得出真實的因果效應。題目給了一個觀察性研究的結果,以及一個隨機事件(系統故障)所產生的「準實驗」數據。
我們有以下資訊:
-
原始觀察性研究:
- ELAP 參與者 (400人):平均 PRI = +12.6 點
- 非參與者 (1600人):平均 PRI = +5.4 點
- 觀察性研究的 PRI 差異 (Observed Difference) = 點。
-
自然實驗 (系統故障):
- 300 名經隨機選擇的經理因系統故障未能完成註冊。
- 這 300 人是隨機抽樣的,與其特徵無關。
- 這 300 人中,有 60 人「有意願註冊」(Intended to Enroll),但未能成功。這群人可以視為「想參加 ELAP 但未成功」的群體。
- 有 240 人「無意願註冊」(Did NOT Intend to Enroll)。這群人可以視為「根本不想參加 ELAP」的群體。
- 「有意願註冊但未成功」者的平均 PRI = +10.8 點。
- 「無意願註冊」者的平均 PRI = +4.2 點。
-
分解公式:
Observed Difference = True Causal Effect (ATT) + Selection Bias
目標: 計算原始觀察性比較中的 Selection Bias。
思路:
- 原始觀察性研究的差異 (7.2 點) 包含了真實的因果效應 (ATT) 和選擇性偏差。
- 自然實驗的部分,由於「有意願註冊但未成功」的 60 人,他們的 PRI 是 +10.8,而「無意願註冊」的 240 人 PRI 是 +4.2。這兩組人都是「未參加 ELAP」的,但他們在「是否有意願註冊」上有差異。
- 「有意願註冊」這個行為本身,可能與經理的能力、潛力、對計畫的興趣等有關,而這些特徵又可能與他們的 PRI 有關。
第 14 題
- A multinational pharmaceutical company headquartered in Basel, Switzerland, is evaluating the
effectiveness of a novel AI-powered diagnostic support tool designed to assist oncologists in treatment
planning. The tool, branded "OncoAssist AI," analyzes patient medical records, genomic data, and clinical
imaging to generate personalized treatment recommendations.
The company's Health Economics and Outcomes Research (HEOR) division conducted a twelve-month
observational study across 45 cancer treatment centers. The primary outcome measure is the "Treatment
Optimization Score" (TOS)-a composite metric ranging from 0 to 100.
Observed Adoption and Outcomes:
Group
OncoAssist AI Adopters (360)
Non-Adopters (540)
Mean TOS
78.4
62.8
Instrumental Variable: During the study period, 18 of the 45 treatment centers experienced a hospital-wide
IT infrastructure upgrade mandated by regional health authorities. This upgrade made the installation of
OncoAssist AI substantially easier. The timing was determined by regulatory schedules established three
years prior-completely independent of oncologist characteristics.
Data Summary by IT Upgrade Status:
Adoption rate
Mean TOS
WITH IT Upgrade
WITHOUT IT Upgrade
54%
28%
72.6
66.2
The LATE formula is: LATE = (Difference in outcomes by instrument) / (Difference in participation rates by
instrument)
Based on the instrumental variable analysis, what is the estimated true causal effect (LATE) of OncoAssist
AI adoption on Treatment Optimization Score?
(A) 6.4 points
(B) 15.6 points
(C) 22.3 points
(D) 24.6 points
(E) 28.4 points
登入後即可作答並保存紀錄。
本題考驗的是工具變數法 (Instrumental Variable, IV) 的應用。題目提供了觀察性研究的結果,以及一個工具變數(IT 基礎設施升級)如何影響處理的採用率和結果。
核心概念:
工具變數法用於估計在存在內生性(如選擇性偏差)時的因果效應。
- 工具變數 (Instrument): IT 基礎設施升級 (Mandate)。
- 內生變數 (Endogenous Variable): OncoAssist AI 的採用率 (Adoption rate)。
- 應變數 (Outcome Variable): Treatment Optimization Score (TOS)。
數據整理:
-
總體觀察數據:
- AI 採用者:360 人,Mean TOS = 78.4
- 未採用者:540 人,Mean TOS = 62.8
- 觀察性差異 (Observed Difference in TOS) = 點。這個差異包含了真實的因果效應和選擇性偏差。
-
工具變數數據 (依據是否有 IT 升級分組):
- 有 IT 升級的地點 (WITH IT Upgrade):
- 採用率 (Adoption rate) = 54%
- Mean TOS = 72.6
- 沒有 IT 升級的地點 (WITHOUT IT Upgrade):
- 採用率 (Adoption rate) = 28%
- Mean TOS = 66.2
- 有 IT 升級的地點 (WITH IT Upgrade):
工具變數法的計算 (LATE - Local Average Treatment Effect):
LATE 公式為:
LATE = (在工具變數存在與不存在下的結果差異) / (在工具變數存在與不存在下的處理率差異)
第 15 題
- A direct-to-consumer e-commerce company headquartered in Los Angeles specializes in premium
skincare products. The company's Chief Marketing Officer has developed a sophisticated framework for
allocating marketing budgets across experimental advertising campaigns based on the principle of
maximizing the expected logarithmic growth rate of customer acquisition capital.
The marketing analytics team has identified an unconventional opportunity: a celebrity influencer with 15
million followers has offered an exclusive 24-hour promotional partnership. Based on historical data from
similar influencer collaborations, the company's data scientists have analyzed engagement patterns, follower
demographics, and conversion funnels to estimate the campaign's potential outcomes.
The analytics team presents their findings:
• Probability of campaign success: 75% (defined as achieving viral engagement that generates
substantial new customer acquisition)
• Probability of campaign failure: 25% (defined as minimal engagement resulting in negligible
customer acquisition)
The financial structure of the campaign operates as follows:
• If successful: The company's customer acquisition capital grows by 200% of the amount invested in
the campaign (i.e., for every 2.00 in customer lifetime value,
resulting in 2 million available for experimental
campaigns this quarter. Let f denote the fraction of this budget allocated to the influencer campaign, where 0
<f≤1.
The CMO's objective is to choose f to maximize the expected logarithmic utility of terminal marketing
capital:
max E[In(W1)]
where Wo represents the initial marketing budget and W. represents the marketing capital after the campaign
outcome is realized.
The CMO explains to the marketing team: "If we invest fraction f of our budget and the campaign succeeds,
our terminal capital becomes Wo(1+2f)-we keep our original capital plus gain 200% on the invested
portion. If the campaign fails, our terminal capital becomes Wo(1-f)-we lose the entire invested amount.
Our goal is to find the optimal f* that maximizes our expected log-capital, which corresponds to maximizing
our long-term compounded growth across many such campaign decisions."
A junior brand manager asks: "With a 75% success probability and 200% returns, shouldn't we invest
aggressively perhaps 75% or more of our budget?"
The CMO responds: "That intuition ignores the asymmetry between gains and losses in a compounding
environment. The mathematics will reveal a more nuanced answer."
What is the optimal fraction f* of the marketing budget to allocate to this influencer campaign?
(A) 0.250
(B) 0.375
(C) 0.500
(D) 0.625
(E) 0.750
登入後即可作答並保存紀錄。
本題考驗的是在有風險情況下,如何最大化期望對數效用 (expected logarithmic utility),這與最大化長期複利增長率 (Kelly Criterion 的概念相關)。
設定:
- 初始資金 (本題中是 W_0f$)。
- 投資於影響者活動的比例為 ()。
- 投資金額為 。
- 成功機率 。
- 失敗機率 。
結果:
-
如果成功:終端資本 。
(CMO 的解釋是 "Wo(1+2f)", 這與 相符。客戶終身價值增加 200% on invested portion,所以 變成 。總資本是未投資部分 加上投資後的總回報 。這有誤。
CMO 的說法是 "gains 200% on the invested portion",即 變成 。
那麼總資本 。
這裡 CMO 的解釋 "we keep our original capital plus gain 200% on the invested portion" 似乎指的是總資本是 。
如果 ,則 。這意味著投入 獲得 的淨收益,總回報是 。
根據題目描述 "grows by 200% of the amount invested",即 變為 。
所以,成功的終端資本 。
這個公式是正確的。) -
如果失敗:終端資本 。
(CMO 的解釋是 "lose the entire amount invested",所以 消失,剩下 。這個公式也是正確的。)
目標: 最大化期望對數效用 。
。
由於 是常數,最大化 等同於最大化:
。
我們需要對 對 求導,並令導數為 0 來找到極值。
求導數 :
令導數為 0:
$\frac{1.5}{1+2f} = \frac{0.25}{1-f}
第 16 題
- The following is an excerpt from an internal memorandum circulated among the senior leadership team
of Meridian Therapeutics, a mid-sized biopharmaceutical company headquartered in Cambridge,
Massachusetts.
In the fourth quarter of last fiscal year, Meridian Therapeutics initiated a pivotal Phase III clinical trial
designated AURORA-7 to evaluate the efficacy of our lead compound, MRD-4821, for the treatment of
treatment-resistant major depressive disorder (TR-MDD). The trial enrolled 840 patients across 62 clinical
sites in North America and Western Europe, randomizing participants in a 1:1 ratio to receive either MRD-
4821 or placebo over a 12-week treatment period. The primary endpoint was defined as the mean change
from baseline in the Montgomery-Åsberg Depression Rating Scale (MADRS) score at Week 12.
The Data Safety Monitoring Board (DSMB) convened last week to review the unblinded interim analysis
results. Dr. Helena Vasquez, the DSMB chair and a renowned biostatistician from Johns Hopkins University,
presented the following summary to our executive committee:
"The interim analysis was conducted when approximately 60% of enrolled patients had completed the Week
12 assessment. At this juncture, patients receiving MRD-4821 demonstrated a mean MADRS reduction of
18.4 points compared to 14.2 points in the placebo arm, yielding a treatment difference of 4.2 points. The
two-sided p-value associated with this comparison was 0.037. Per the pre-specified interim analysis plan,
which employed an O'Brien-Fleming alpha-spending function, the boundary for declaring efficacy at this
interim look was set at p < 0.012."
Following Dr. Vasquez's presentation, several members of our executive committee offered their
interpretations of these results:
• Chief Executive Officer, Marcus Chen: "While I understand we haven't crossed the formal efficacy
boundary, the p-value of 0.037 would be considered statistically significant under conventional
standards. The DSMB's conservative threshold is merely a procedural technicality. I recommend we
begin preparing our regulatory submission package immediately, as these results clearly demonstrate
that MRD-4821 works."
• Chief Medical Officer, Dr. Priya Sharma: "I must respectfully disagree with Marcus's
characterization. The alpha-spending approach exists precisely to preserve the overall Type I error
rate across multiple analyses. If we had pre-specified only a single final analysis, a p-value of 0.037
would indeed cross the conventional 0.05 threshold. However, by conducting this interim look with
the possibility of early stopping, we have effectively performed multiple hypothesis tests on
accumulating data. The O'Brien-Fleming boundary of 0.012 at this interim analysis is not a
'technicality' it is the mathematically correct threshold required to maintain a 5% overall false
positive rate. Our current results, while encouraging, do not constitute definitive evidence of
efficacy."
(Continued on next page)
(Continued from previous page)
• Chief Financial Officer, Jonathan Mills: "From a capital markets perspective, I'm concerned about
the investment community's reaction if these results were to become public prematurely. Our stock
price has already appreciated 40% over the past six months in anticipation of positive AURORA-7
data. If we announce that the trial 'failed to meet its interim endpoint,' shareholders may interpret this
as a negative signal, even though the trial is continuing to its final analysis. I suggest we refrain from
any public disclosure until the final results are available."
• Chief Commercial Officer, Diana Rodriguez: "I've been analyzing the competitive landscape, and I
believe timing is critical. Two of our competitors are expected to report Phase III results for their
own TR-MDD candidates within the next nine months. If we wait for our final analysis-which
won't be available for another eight months-we risk losing first-mover advantage in physician
awareness and formulary positioning. The interim p-value of 0.037 suggests a very high probability
that the final analysis will be positive. I recommend we accelerate our commercial launch
preparations based on these interim findings."
• Vice President of Biostatistics, Dr. Samuel Okonkwo: "I want to address a point of confusion that
seems to be emerging. Diana mentioned that the interim p-value 'suggests a very high probability that
the final analysis will be positive.' This interpretation reflects a common misconception. A p-value of
0.037 means that, assuming the null hypothesis is true that is, assuming MRD-4821 has no effect
whatsoever-there would be a 3.7% probability of observing a treatment difference at least as large
as what we found. It does not tell us the probability that MRD-4821 actually works, nor does it
directly predict the outcome of the final analysis. The final analysis could yield a smaller treatment
effect due to regression to the mean, differential dropout patterns, or site-by-treatment interactions
that manifest differently as enrollment matures."
The DSMB has recommended that the trial continue to its pre-planned final analysis without modification.
Based on the passage, which of the following statements represents the most accurate interpretation of the
statistical evidence and reasoning presented?
(A) The CEO's position is fundamentally sound because a p-value of 0.037 provides strong evidence that
MRD-4821 is effective; the O'Brien-Fleming boundary is an arbitrary administrative threshold that does not
affect the scientific interpretation of the data.
(B) The CFO's concern about investor reaction is the most statistically sophisticated perspective presented,
as it recognizes that the market's interpretation of hypothesis testing outcomes may differ from the technical
statistical conclusions.
(C) The CMO's reasoning is correct: the interim efficacy boundary of p < 0.012 is not merely procedural but
is the appropriate threshold for this analysis, because conducting multiple analyses on accumulating data
inflates the probability of a false positive finding unless stricter significance levels are applied at each look.
(D) The CCO's inference is valid because a p-value below 0.05 at the interim analysis mathematically
guarantees at least a 95% probability that the final analysis will also yield a statistically significant result.
(E) Dr. Okonkwo's critique of the CCO's reasoning is flawed because p-values do, by definition, represent
the probability that the experimental treatment is effective given the observed data.
登入後即可作答並保存紀錄。
核心觀念
本題考查臨床試驗中期分析的多重檢定、 值解釋,以及 O’Brien–Fleming alpha-spending function 的意義。
- 值的定義
在虛無假設 為真的前提下, 值是觀察到目前結果或更極端結果的機率:
本題的 表示:若 MRD-4821 實際上沒有治療效果,觀察到目前或更極端治療差異的機率為 。它不是「藥物有效的機率」。
- 中期分析與整體型一錯誤
若只在試驗結束時進行一次檢定,顯著水準通常設定為 。但本題先進行中期分析,之後還會進行最終分析,相當於針對累積資料進行多次檢定。
多次檢定會增加整體型一錯誤率:
因此,研究設計須將總體顯著水準 分配到不同分析時點,這就是 alpha spending 的概念。
- O’Brien–Fleming 邊界
O’Brien–Fleming 方法通常在早期分析採用非常嚴格的拒絕界線,在最終分析時再放寬界線。其理由是:早期資料量較少、資訊量較低,若僅因早期偶然出現顯著結果就停止試驗,容易增加錯誤宣稱有效的機率。
本題中期分析的預先規定界線為:
實際結果為:
因為
所以尚未達到中期宣告療效的標準。這不代表藥物已被證明無效,而是表示目前證據不足以在中期提前宣告成功,試驗應依原計畫繼續至最終分析。
解題方法
本題不需要由平均數 、 及治療差異 重新計算檢定統計量,因為題目已直接給出中期分析的 值。判斷流程如下:
-
確認目前是預先規劃的中期分析,而非唯一一次最終分析。
-
使用研究計畫預先指定的中期顯著性界線 。
-
比較實際 值與中期界線:
-
得出結論:中期結果具有鼓舞性,但未達提前宣告療效的門檻。
-
進一步檢查各主管的統計解釋,判斷誰正確理解多重檢定與 值。
選項分析
(A) 錯誤。
CEO 將 與一般單次檢定的 比較,忽略本研究已事先規劃中期分析與最終分析。若只進行一次最終檢定, 確實小於 ;但在序列性試驗中,必須依中期分析所設定的 O’Brien–Fleming 界線判斷。
第 17 題
- Dr. Helena Morrison, a behavioral economist at the London School of Economics, is conducting a study
on decision-making under uncertainty. She has designed an experimental maze for her research subjects. The
maze has a single starting chamber with exactly three doors, labeled A, B, and C. Each door leads to a
different tunnel, and the subjects have no information about where each tunnel leads. Due to the maze's
construction, once a subject enters a tunnel, they cannot turn back and must complete that path.
The three tunnels have the following properties, unknown to the subjects: Door A leads through a winding
passage that takes exactly 2 hours to traverse, after which the subject emerges at the exit and successfully
completes the maze. Door B leads through a deceptive corridor that takes 3 hours to traverse, but ultimately
returns the subject to the original starting chamber. Door C leads through a short loop that takes 1 hour to
traverse, also returning the subject to the starting chamber.
Each time a subject arrives at (or returns to) the starting chamber, they choose among the three doors with
equal probability of 1/3 each, as subjects cannot remember or distinguish which doors they have previously
tried. Dr. Morrison wishes to calculate the expected time for a subject to exit the maze.
What is the expected number of hours for a subject to exit the maze?
(A) 4 hours
(B) 6 hours
(C) 8 hours
(D) 10 hours
登入後即可作答並保存紀錄。
本題考驗的是期望值的計算,特別是當存在一個可能無限循環的路徑時。這與第 10 題類似,但有不同的時間和路徑設定。
設 為從起始點到達出口的預期總時間。
在起始點,有三扇門:A, B, C。每扇門被選擇的機率是 1/3。
- 選擇門 A: 耗時 2 小時,直接到達出口。
- 選擇門 B: 耗時 3 小時,回到起始點。回到起始點後,預期還需 小時才能成功。總時間為 。
- 選擇門 C: 耗時 1 小時,回到起始點。回到起始點後,預期還需 小時才能成功。總時間為 。
根據期望值的定義,我們可以寫出關於 的方程式:
將上述時間代入:
現在,我們解這個方程式來求 :
第 18 題
- Global Entertainment Network (GEN) is reviving a classic television game show format for the 2025
season. The show's signature segment works as follows: A contestant is presented with three identical doors
on stage. Behind one door is a brand-new Tesla Model S valued at 15. The contestant has no prior information
about which door hides the car.
The game proceeds in two stages. In Stage 1, the contestant selects one of the three doors but does not open
it. In Stage 2, the host-who knows exactly what is behind each door deliberately opens one of the two
remaining doors to reveal a houseplant. Crucially, the host never opens the door concealing the car, and if
the contestant's initial choice was a houseplant, the host has only one valid door to open; if the contestant
initially chose the car, the host randomly selects one of the two houseplant doors to open. After the host
reveals a houseplant, the contestant is offered a choice: stay with the original selection or switch to the other
unopened door.
The show's statistical consultant has been asked to determine the probability of winning the car under each
strategy, assuming the contestant's initial choice is random.
What is the probability of winning the car if the contestant always switches, compared to always staying?
(A) Switching: 1/2; Staying: 1/2 (no advantage either way)
(B) Switching: 2/3; Staying: 1/3 (switching doubles the odds)
(C) Switching: 1/3; Staying: 2/3 (staying is better)
(D) Switching: 3/4; Staying: 1/4 (switching triples the odds)
登入後即可作答並保存紀錄。
核心觀念
本題考的是經典的「蒙提霍爾問題」(Monty Hall Problem),核心在於:
- 初始選中的車門藏車機率為 。
- 初始未選中的兩扇門合計藏車機率為 。
- 主持人知道答案,且一定會打開一扇藏植物的門,因此他的行動不具有隨機淘汰效果,而是保留了原先未選區域的 機率。
因此,主持人打開植物門後:
- 留在原門:勝率仍為 。
- 改選另一扇未開門:勝率為 。
解題方法
設參賽者一開始選擇 1 號門。
情況一:1 號門後是 Tesla
其機率為:
此時另外兩扇門都是植物門,主持人隨機打開其中一扇。參賽者若:
- 留在 1 號門:獲勝。
- 改選剩下的未開門:選到植物,失敗。
情況二:1 號門後是植物
其機率為:
此時另外兩扇門中,一扇是 Tesla,另一扇是植物。主持人只能打開那扇植物門,因此剩下未開的門必定是 Tesla。參賽者若:
- 留在 1 號門:失敗。
- 改選剩下的未開門:獲勝。
整理如下:
| 初始選擇 | 機率 | 留下 | 改選 |
|---|---|---|---|
| 選到 Tesla | 勝 | 敗 | |
| 選到植物 | 敗 | 勝 |
所以:
一直留下
只有初始選到 Tesla 才會獲勝:
一直改選
只要初始選到植物,改選就會換到 Tesla:
故:
選項分析
第 19 題
- The Taipei Metro operates the Wenhu Line with trains departing from Taipei Zoo station according to a
Poisson process. On average, trains depart at a rate of trains per hour during peak hours, meaning the
inter-arrival times between successive trains are independent and exponentially distributed with mean 10
minutes.
A commuter arrives at the platform at a uniformly random time during the peak hour (i.e., the commuter's
arrival time is independent of the train schedule). Intuitively, one might expect the average waiting time for
the next train to be half the average inter-arrival time, i.e., 5 minutes. However, a transportation analyst
claims that the expected waiting time is actually longer due to a phenomenon known as the "inspection
paradox" or "waiting time paradox."
The paradox arises because a randomly arriving passenger is more likely to arrive during a longer-than-
average inter-arrival interval (since longer intervals occupy more time and thus have a higher probability of
being "hit" by a random arrival). For a Poisson process, however, the memoryless property of the
exponential distribution implies that the waiting time from any arbitrary arrival point has the same
distribution as the original inter-arrival time.
Given that trains arrive according to a Poisson process with rate 6 per hour, what is the expected waiting
time for a randomly arriving passenger?
(A) 5 minutes
(B) 10 minutes
(C) 15 minutes
(D) 20 minutes
登入後即可作答並保存紀錄。
本題考驗的是「檢驗悖論」或「等待時間悖論」,以及 Poisson 過程和指數分布的記憶無關性。
核心概念:
- Poisson 過程: 事件(火車抵達)發生的時間點遵循 Poisson 過程。
- 指數分布: 相鄰事件之間的時間間隔 (inter-arrival times) 服從指數分布。
- 記憶無關性 (Memorylessness): 指數分布具有記憶無關性,這意味著在任意時刻,下一個事件發生所需的時間的分布,與從上一個事件發生後經過的時間無關。對於一個服從指數分布的隨機變數 (例如,火車間隔時間),對於任意 ,有 。
- 檢驗悖論 (Inspection Paradox): 當我們在一個隨機時間點觀察一個過程時,我們更有可能落在較長的間隔中,這會導致觀察到的平均間隔時間比實際的平均間隔時間更長。
題目資訊:
- 火車抵達率 列/小時。
- 平均車間隔時間 (mean inter-arrival time) = 小時 = 10 分鐘。
- 一個通勤者在一個隨機的時間點到達月台,與火車時刻表無關。
悖論的解釋:
直觀上,如果平均間隔是 10 分鐘,那麼隨機到達的乘客預期等待時間應該是 10 分鐘的一半,即 5 分鐘。
然而,檢驗悖論指出,隨機到達的乘客更有可能在「較長的」火車間隔中到達。為什麼?想像一個時間軸,上面標示著火車的抵達時間點。這將時間軸分割成一系列的火車間隔。一個隨機到達的乘客,有更高的機率落在一個較長的區間內,因為較長的區間「佔據」了更多的時間。
數學推導:
設 為火車的間隔時間,服從指數分布,其均值為 分鐘。
對於一個隨機到達的乘客,其等待時間 的期望值 。
第 20 題
- A major technology company has implemented an Al-powered security system to detect unauthorized
access attempts on its corporate network. The system was developed by analyzing historical data and has the
following validated performance characteristics: When a genuine intrusion attempt occurs, the system
correctly identifies it as an intrusion with probability 0.99 (true positive rate / sensitivity). When normal,
authorized activity occurs, the system incorrectly flags it as an intrusion with probability 0.02 (false positive
rate).
The company's cybersecurity team estimates that, on any given day, approximately 0.1% of all network
activities represent genuine intrusion attempts, while 99.9% are normal authorized operations. On a
particular Monday morning, the Al system flags a specific activity as a potential intrusion, triggering an alert
to the security operations center.
The Chief Information Security Officer (CISO) asks the analytics team to calculate the probability that this
flagged activity is actually a genuine intrusion, given that the system has raised an alert. This requires
applying Bayes' theorem to update the prior probability of intrusion (0.1%) based on the evidence of a
positive alert.
What is the probability that the flagged activity is a genuine intrusion, given that the Al system raised an
alert?
(A) Approximately 1%
(B) Approximately 5%
(C) Approximately 33%
(D) Approximately 99%
登入後即可作答並保存紀錄。
本題考驗的是貝氏定理 (Bayes' Theorem) 的應用,用於計算後驗機率。我們需要根據系統的性能指標(敏感度和假陽性率)以及網路活動的先驗機率,來計算在系統發出警報的情況下,該活動確實是真實入侵的機率。
設定:
- 令 為事件「活動是真實入侵」。
- 令 為事件「活動是正常操作」。
- 令 為事件「系統發出警報 (Flagged activity)」。
已知資訊:
-
先驗機率 (Prior probabilities):
- (真實入侵的機率)
- (正常操作的機率)
- 注意:。
-
系統性能指標 (Likelihoods):
- 敏感度 (Sensitivity) / 真陽性率 (True Positive Rate, TPR): 在真實入侵的情況下,系統發出警報的機率。
(當是入侵時,99% 的機率被正確標記為入侵)。 - 假陽性率 (False Positive Rate, FPR): 在正常操作的情況下,系統錯誤地發出警報的機率。
(當是正常時,2% 的機率被錯誤標記為入侵)。
- 敏感度 (Sensitivity) / 真陽性率 (True Positive Rate, TPR): 在真實入侵的情況下,系統發出警報的機率。
目標:
計算後驗機率 ,即在系統發出警報的情況下,該活動是真實入侵的機率。
第 21 題10 分
The length of time between calls received by the switchboard in a large legal firm is an exponential random variable. The average length of time between calls is 20 seconds. If a call has just been received, what is the probability that no calls are received in the next 30 seconds?
登入後即可作答並保存紀錄。
核心觀念
相鄰來電的時間間隔 服從指數分配。若平均間隔為 秒,則指數分配的平均數與率參數滿足
因此率參數為 ,單位是每秒。指數分配的存活機率(等待時間超過 秒的機率)為
解題方法
題目要問「接到一通電話後,接下來 秒沒有新電話」的機率,也就是下一個來電的等待時間超過 秒:
第 22 題10 分
Suppose a surfboard company has found its yearly profit equation to be
where is the number (in thousands) of type-A surfboards produced per year, is the number (in thousands) of type-B surfboards produced per year, and is profit (in thousands of dollars). How many of each type of board should be produced per year to realize a maximum profit? What is the maximum profit?
登入後即可作答並保存紀錄。
核心觀念
這題考二變數二次函數的極大值。若利潤函數的 Hessian 矩陣為負定矩陣,函數即為嚴格凹函數;其唯一駐點就是全域最大值點。
題目給定:
其中 分別以「千片」為單位, 以「千美元」為單位。
解題方法
先令利潤函數對兩個產量變數的偏導數為零,求出駐點:
由第二式得 ,即 。代入第一式:
接著確認此駐點確為最大值點。Hessian 矩陣為:
Question I (20 points; 10 points each)
In this question, we introduce autocorrelation in the disturbances but retain all the other assumptions of the classical model. The term autocorrelation may be defined as “correlation between members of series of observations ordered in time as in time series data.” To explain the basic ideas involved, we consider the two-variable regression model
Note that we are using the subscript to emphasize that we are dealing with time series data. Assume that the disturbance terms follow the AR(1) scheme, which is known as a first-order autoregressive scheme:
where is the coefficient of autocovariance and is the stochastic disturbance term such that it satisfies the standard ordinary least squares (OLS) assumptions, namely,
It can be shown that the properties of the disturbance term are
第 23 題10 分
Show that, under the AR(1) scheme, the disturbance term exhibits autocorrelation by deriving .
登入後即可作答並保存紀錄。
核心觀念
題目要證明 AR(1) 誤差項 與前一期誤差項 有非零乘積期望,也就是兩期誤差項存在自我相關。題目圖中給定
且 、;各期創新項 彼此不相關。
解題方法
由圖中的 AR(1) 關係式,把 代入 。關鍵是 與過去的 不相關: 由當期之前的創新項構成,而 與這些過去創新項不相關。
因此,
又因 ,所以 。依題目給定的變異數,
第 24 題10 分
Suppose that is known. Show that the original model can be transformed into the following form:
where . In particular, find the in terms of and .
登入後即可作答並保存紀錄。
核心觀念
本題考的是 AR(1) 誤差的「準差分」轉換。由誤差結構
可知,將原迴歸式減去其落後一期式的 倍後,誤差項會化為 。
解題方法
依原卷圖,模型為 ,誤差遵循 ,題目定義 ,並要求改寫為含 的形式。對原模型取落後一期,再乘上 :
將此式乘以 ,再從原模型相減:
Question II (20 points; 5 points each)
In a study prepared for the U.S. Department of Labor, researchers were interested in examining the labor-force participation of various categories of labor as a function of several explanatory variables. In all their regressions, the dependent variable is a dummy, taking a value of 1 if a person is in the labor force, 0 if she is not. Table II reports one of the estimated regressions, which is obtained using ordinary least squares (OLS).
Turning to interpretations of the findings, we see that each slope coefficient gives the rate of change in the conditional probability of the event occurring for a given unit change in the value of the explanatory variable. For instance, the coefficient of attached to the variable “age 65 and over” means, holding all other factors constant, the probability of participation in the labor force by women in this age group is smaller by about 27 percent (as compared with the base category of women aged 22 to 54). Now consider the interaction term marital status and age. Table II shows that the probability of participation of women who were never married and are 65 or over is smaller by about 20 percent as compared with the base category. Following this procedure, readers can easily interpret the rest of the coefficients given in Table II.
Table II. Labor-Force Participation
Regression of women, age 22 and over, living in largest 96 standard metropolitan statistical areas (SMSA) (dependent variable: in or out of labor force during 1966)
| Explanatory Variable | Coefficient | Ratio |
|---|---|---|
| Constant | 0.24 | 15.4 |
| Marital status (dummy) | ||
| Married | — | — |
| Never married | 0.29 | 22.0 |
| Age (dummy) | ||
| 22–54 | — | — |
| 55–64 | ||
| 65 and over | ||
| Years of schooling | 0.02 | 5.8 |
| Interaction (marital status and age) | ||
| Never married, 55–64 | ||
| Never married, 65 and over |
No. of observations = 25,153
Note: — indicates the base or omitted category.
第 25 題5 分
Based on the regression results, interpret the effect of education on women’s labor-force participation and determine whether the effect is statistically significant.
登入後即可作答並保存紀錄。
核心觀念
這題考線性機率模型的迴歸係數解讀與 檢定。依原卷表格,依變數是是否參與勞動力市場的虛擬變數;受教育年數的估計係數為 , 值為 ,樣本數為 。
在線性機率模型中,斜率係數表示解釋變數增加一單位時,條件參與機率的變化量。因此, 表示機率增加 ,也就是增加 2 個百分點。
解題方法
控制婚姻狀態、年齡組別及婚姻狀態與年齡的交互作用後,教育年數每增加一年,女性參與勞動力的估計機率增加 。模型沒有列出教育年數與其他變數的交互作用,因此此係數表示各組共用的教育效果。
檢定教育效果是否顯著,設定:
第 26 題5 分
We would like to obtain the estimates of the conditional probabilities of labor-force participation of the various categories. In particular, compute the probability for women who were never married, aged 55 to 64, with 16 years of schooling.
登入後即可作答並保存紀錄。
核心觀念
本題考線性機率模型(以 OLS 估計的虛擬變數迴歸)。因變數為是否參與勞動市場的虛擬變數,迴歸式的預測值可解讀為給定條件下的參與機率。計算時須納入截距、各虛擬變數係數、交互作用係數,以及教育年數的係數乘上年數。
解題方法
圖中表格給出截距 、未婚係數 、55–64 歲係數 、教育年數係數 ,以及「未婚且 55–64 歲」交互作用係數 。已婚、22–54 歲是基準組;題目指定教育年數為 16 年。
第 27 題5 分
By how many percentage points is the labor-force participation probability higher or lower for women who were never married and aged 55–64 compared with those who were never married and aged 65 or over?
登入後即可作答並保存紀錄。
核心觀念
含交互作用的迴歸模型中,某一類別的影響要把該類別的主效果與適用的交互作用係數相加。比較兩組預測機率時,其他相同條件的項目會相互抵銷。
解題方法
圖中表二以「已婚、22–54 歲」為婚姻狀態與年齡的基準組;未婚係數為 ,55–64 歲係數為 ,65 歲以上係數為 ,兩組未婚交互作用係數分別為 與 。
兩組都是未婚女性,因此未婚主效果、常數項,以及相同的受教育年數效果都會抵銷。各年齡組相對於基準組的年齡相關效果為:
第 28 題5 分
How would you test the hypothesis that the effect of married status on labor-force participation probability does not depend on age? State the hypotheses and the appropriate test.
登入後即可作答並保存紀錄。
核心觀念
題目考驗如何用迴歸中的交互作用項,檢定婚姻狀態對勞動參與機率的影響是否會隨年齡改變。若交互作用係數全為零,代表婚姻狀態的效果不因年齡組別而異;因此這是對多個係數的聯合檢定。
解題方法
圖中迴歸以「已婚」及「22–54 歲」為基準組,兩個交互作用項分別是「未婚、55–64 歲」與「未婚、65 歲以上」,其 值為 、。因此應同時檢定這兩個交互作用係數是否為零。
令 、 分別為上述兩個交互作用係數,假設為:
使用部分 檢定:比較含有兩個交互作用項的完整模型,與刪除這兩項的受限模型。完整模型共有 7 個參數:截距、婚姻狀態 1 個係數、年齡 2 個係數、教育年數 1 個係數,以及交互作用 2 個係數。樣本數為 ,故分母自由度為 。檢定統計量為: