2. 臨床醫學統計方法攻略
重複測量統計方法
Repeat
measurement
ANOVA (RM ANOVA)
Mixed model
Random intercept and
trend model
Pattern mixture model
Mixed model for
longitudinal binary data
GEE model
存活分析
Kaplain-Meier
method and Cox
model
Time-dependent cox
model
Parametric survival
analysis
Exponential/Weibull
regression model
Interval censor data
analysis
3. " There are three kinds of lies: lies,
damned lies, and statistics"
Benjamin Disraeli (1804-1881)
3
13. Mauchly„s test
Sphericity Test - Mauchly, 1940
小樣本數則power低/大樣本則敏感
易受極端值或非常態影響
參考但非唯一準則
如果rejected,且真的很明顯
Adjusted univariate F-tests
Greenhouse-Geisser (1959)
Huynh-Feldt (1976)
考慮使用Multivariate repeated measures analysis
(MANOVA)
allows for general V (yi)
but doesn‟t allow any missing data across time
22. Little's test for missing value
completely at random
Test of Missing Completely at Random for Multivariate Data with Missing
Values
Roderick J. A. Little
Journal of the American Statistical Association
Vol. 83, No. 404 (Dec., 1988), pp. 1198-1202
22
29. Peto Finding
Richard Peto的發現
臨床上的研究並不能完全按照費雪的”隨機”要求
過去只要退出或者出現遺漏值,就是刪除不予以分析
若癌症病人因安慰劑沒反應退出,照傳統做法就是不
列入分析,我們將會得到?
對安慰劑持續有反應的病人留下來
對新療法有反應的人也留下來
結論會是??
新療法跟安慰劑一樣有用
ITT population分析的限制
只能用來找出不同
若找不出差異,並不代表兩種療法的效果一樣
29
30. Mixed model的總結
30
Mixed model
Random covariance model
Random intercept model
Random coefficient model
Pattern Mixture model to solve missing at
random
GLMM model (Generalized linear mixed model)
31. Pattern-mixture models for
missing data
31
根據遺漏值的pattern,創造一個變數來描述這個
pattern
注意事項:這個變數是用來描述個體間的遺漏狀態
例如一個觀察值測量三次,則最簡單的分法做一個D
變項,D=0表示至少有一次遺漏,D=1表示
completer
33. 33
認識連結函數 (Link Function)
Y link function g,
Y distribution and
Link function
Ex:Binomial
distribution and
logit link function
0
1
( ( ))
p
h h
h
g E Y x
0.0
0.2
0.4
0.6
0.8
1.0
34. 34
GEE的工作矩陣 (Working matric)
• 如果一個人有三筆重複測量資料
且非獨立,亦即有3個相關係數要
估計
• 如果有100名個案,則會出現300
個相關係數要估計
• 如果是測量4次呢?
6*100=600
• parameter>n
• 因此需要一個工作矩陣去假設這
些重複測量的關係
1
12 13
21 23
31 32
1
1
1
0
0 n
B
B
a a
a a
a a
Cluster is a unit, to solve
within unit correlation in
a cluster
41. Time-to-event examples
Oncology
Overall survival (OS)
Time to progression (TTP)
CV disease
Time to hospitalization
Time to nonfatal stroke
Time to MI
Time to death
Time to abnormal 1mm ST segment depression
COPD (Chronic Obstructive Pulmonary Disease )
Time to exacerbation
Time to the Adverse event
41
54. 母數存活分析
(Parametric survival analysis)
何謂母數存活分析
Linear regression, logistic regression, and Poisson
regression are examples of parametric models that
are commonly used in the health sciences.
Distributions commonly used for parametric
survival models:
Weibull
Exponential
Log-logistic
Lognormal
Generalized gamma
55. 母數存活分析 與cox-model有何不同?
The Cox proportional hazards model is not a fully
parametric model.
if the regression parameters (the betas) are known,
the distribution of the outcome remains unknown.
The baseline survival (or hazard) function is not
specified in a Cox model.
Cox model widely popular:
No reliance on assumed distribution
Baseline not necessary for estimation of hazard
ratio
58. Accelerated Failure Time (AFT)
AFP是測量因子與存活時間的關係
HR是測量因子與Hazard的關係
AFP舉例
T1:survival time for smoker
T2:survival time for non-smoker
T2/T1=γ
AFP(γ)是強調時間的關係
如果γ =2
SMOKER存活率0.5時是5個月
則NON-SMOKER存活率0.5時____個月