南方醫(yī)科大學(xué)公共衛(wèi)生學(xué)院生物統(tǒng)計(jì)學(xué)系(510515) 王 巖 錢(qián)若遠(yuǎn) 陳平雁

1.3 多樣本的均數(shù)比較

1.3.1 差異性檢驗(yàn)

1.3.1.4 協(xié)方差分析(ANCOVA)

(1-43)

計(jì)算樣本量時(shí)，先設(shè)定樣本量初始值，然后迭代樣本量直到所得的檢驗(yàn)效能滿足設(shè)定值為止，此時(shí)的樣本量即研究所需的樣本量[2]。

【例1-27】某研究欲比較三種不同教學(xué)干預(yù)方法(有聲思維教學(xué)法、指導(dǎo)性閱讀思維教學(xué)法、指導(dǎo)性閱讀教學(xué)法)對(duì)小學(xué)四年級(jí)學(xué)生閱讀理解能力的影響。采用平衡設(shè)計(jì)，在干預(yù)前后分別評(píng)估受試者的錯(cuò)誤檢測(cè)任務(wù)(error detection task，EDT)得分。以干預(yù)前EDT得分和干預(yù)后理解監(jiān)控問(wèn)卷(comprehension monitoring questionnaire)得分為2個(gè)協(xié)變量，干預(yù)后EDT得分為因變量進(jìn)行協(xié)方差分析。根據(jù)既往研究數(shù)據(jù)，三組均值分別為8.2220，9.8148，6.1904，標(biāo)準(zhǔn)差為2.3788，ρ2為0.4434，設(shè)定檢驗(yàn)水準(zhǔn)為0.05，試以檢驗(yàn)效能分別為0.8和0.9進(jìn)行樣本量估計(jì)。

nQuery Advanced實(shí)現(xiàn):設(shè)定檢驗(yàn)水準(zhǔn)α=0.05；檢驗(yàn)效能分別取80%和90%。依據(jù)上述基礎(chǔ)數(shù)據(jù)可知，G=3，c=2，ρ2=0.4344，σ=2.3788，r1=r2=r3=1，μ={μ1,μ2,μ3}={8.2220,9.8148,6.1904}計(jì)算可得V=2.20，在nQueryAdvanced 主菜單選擇：

Design：⊙Fixed Term

Goal：⊙Means

No.of Groups：⊙>Two

Analysis Method：⊙Test

方法框中選擇Analysis of Covariance(ANCOVA)。

在彈出的樣本量計(jì)算窗口將各參數(shù)鍵入，如圖1-65所示，檢驗(yàn)效能取80%的結(jié)果N=18，即總樣本量為18例，每組6例。檢驗(yàn)效能取90%的結(jié)果N=22，即總樣本量為22例，三組例數(shù)分別為8例、7例和7例(因總例數(shù)不是組數(shù)的整倍數(shù)，故各組例數(shù)不絕對(duì)相等)。

圖1-65 nQuery Advanced 關(guān)于例1-27樣本量估計(jì)的參數(shù)設(shè)置與計(jì)算結(jié)果

SAS 9.4軟件實(shí)現(xiàn)：

%let u={8.2220 9.8148 6.1904}; /*各組均值*/

%let r={1 1 1}; /*各組樣本量與第一組(參照組)樣本量的比值*/

proc IML;

start MGT4_1(a,G,sd,c,R2,power);

error=0;

if ( a>=1 | a<=0 ) then do; error=1; print "Error" "Test significance Level must be in 0-1"; end;

if ( sd<=0 ) then do; error=1; print "Error" "Common standard deviation must be >=0"; end;

if (G<=0 | ceil(G)^=G) then do; error=1; print "Error" "The Number of groups must be positive integer "; end;

if (c<=0 | ceil(c)^=c) then do; error=1; print "Error" "The Number of coveriates must be positive integer "; end;

if(power>=100 | power<1) then do; error=1; print "error" "Power(%) must be in 1-100"; end;

if(R2>1 | R2<0) then do; error=1; print "error" "R_squared value must be in 0-1";end;

if(error=1) then stop;

if(error=0) then do;

n1=ceil(c/G+1); n=n1#&r.;sum_N=n[1,+]-0.01;

do until(pw>=power/100);

sum_n=ceil(sum_n+0.01); n1=sum_N/&r.[1,+]; n=n1#&r.;

u_mean=(n#&u.)[,+]/sum_N;

v=(&r.#(&u.-u_mean)##2)[,+]/&r.[,+];

n_mean=sum_N/G;

sigma_e2=(1-R2)*sd##2;

lamda=n_mean*G*v/sigma_e2;

df1=G-1;df2=sum_N-G-c;

f=finv(1-a,df1,df2);

pw=1-probf(f,df1,df2,lamda); pw1=100*pw;

end;

end;

print a[label="Test Significance level"]

G[label="Number of groups"]

V[label="Variance of means"]

sd[label="Common standard deviation"]

c[label="Number of coveriates"]

R2[label="R_squared value between the reponse and the covariates"]

Pw1[label="Power(%)"]

sum_N[label="Total sample size"];

finish MGT4_1;

run MGT4_1(0.05,3,2.3788,2,0.4434,80);

run MGT4_1(0.05,3,2.3788,2,0.4434,90);

quit;

SAS 9.4運(yùn)行結(jié)果：

圖1-66a SAS 9.4關(guān)于例1-27樣本量估計(jì)的參數(shù)設(shè)置與計(jì)算結(jié)果(power=80%)

圖1-66b SAS 9.4關(guān)于例1-27樣本量估計(jì)的參數(shù)設(shè)置與計(jì)算結(jié)果(power=90%)

1.3.1.5 多變量方差分析(MANOVA)

方法：多變量方差分析針對(duì)分析中有多個(gè)反應(yīng)變量的情況，目前有多種常用統(tǒng)計(jì)量可用于分析，nQuery給出了其中三種供用戶選擇，分別是Wilks’似然比統(tǒng)計(jì)量、Pillai-Bartlett Trace統(tǒng)計(jì)量、Hotelling-Lawley Trace統(tǒng)計(jì)量。Muller and Barton (1989)[3]，以及 Muller，LaVange，Ramey(1992)[4]等給出了多變量方差分析的樣本量和檢驗(yàn)效能的估計(jì)方法，各主效應(yīng)及其交互效應(yīng)的檢驗(yàn)效能估計(jì)是建立在各自的自由度及非中心參數(shù)的F分布上。其檢驗(yàn)效能的計(jì)算公式為：

1-β=1-ProbF(F1-α,df1,df2,df1,df2,λ)

(1-44)

式中，df1，df2代表F分布的自由度，λ為非中心參數(shù)，根據(jù)不同的統(tǒng)計(jì)量，其各參數(shù)計(jì)算方式如下：

首先，我們用p代表因變量個(gè)數(shù)，q代表所研究的分組變量的水平數(shù)，X代表設(shè)計(jì)矩陣,r代表設(shè)計(jì)矩陣的秩，M代表均值矩陣，Σ代表協(xié)方差矩陣，C代表對(duì)比矩陣，不同效應(yīng)的檢驗(yàn)可以構(gòu)建不同對(duì)比矩陣C，N代表總樣本量。Wilks’似然比統(tǒng)計(jì)量、Pillai-Bartlett Trace統(tǒng)計(jì)量、Hotelling-Lawley Trace統(tǒng)計(jì)量可基于如下矩陣構(gòu)建 ：

H=(CM)′[C(X′X)-1C′]-1(CM)

(1-45)

E=Σ(N-r)

(1-46)

T=H+E

(1-47)

(1)Wilks’ Lambda

利用矩陣(1-46), (1-47)可得Wilks’似然比統(tǒng)計(jì)量：

W=|ET-1|

(1-48)

該檢驗(yàn)統(tǒng)計(jì)量轉(zhuǎn)化為近似的F統(tǒng)計(jì)量為：

(1-49)

其中，

df1=ap，

df2=g[(N-r)-(p-a+1)/2]-(ap-2)/2，

(2) Pillai-Bartlett Trace

利用矩陣(1-45)，(1-47)可得Pillai-Bartlett Trace統(tǒng)計(jì)量：

PBT=tr(HT-1)

(1-50)

該檢驗(yàn)統(tǒng)計(jì)量轉(zhuǎn)化為近似的F統(tǒng)計(jì)量為：

(1-51)

其中，

df1=ap，

df2=s[(N-r)-p+s]，

(3)Hotelling-Lawley Trace

利用矩陣(1-45), (1-46)可得Hotelling-Lawley Trace統(tǒng)計(jì)量：

HLT=tr(HE-1)

(1-52)

該檢驗(yàn)統(tǒng)計(jì)量轉(zhuǎn)化為近似的F統(tǒng)計(jì)量為：

(1-53)

其中，

df1=ap，

df2=s[(N-r)-p-1]+2，

對(duì)于上述各統(tǒng)計(jì)量，其非中心參數(shù)λ為：

λ=F·df1

(1-54)

計(jì)算樣本量時(shí)，先設(shè)定樣本量初始值，然后選擇相應(yīng)統(tǒng)計(jì)方法并迭代樣本量直到所得的檢驗(yàn)效能滿足設(shè)定值為止，此時(shí)的樣本量即研究所需的樣本量[2]。

【例1-28】某研究欲比較不同教學(xué)方法在培養(yǎng)中學(xué)生良好學(xué)習(xí)行為中的作用，教學(xué)方法分為傳統(tǒng)教學(xué)組、視聽(tīng)教學(xué)組和對(duì)照組三個(gè)組，因變量包括4個(gè)衡量學(xué)習(xí)行為的指標(biāo)：學(xué)習(xí)環(huán)境、學(xué)習(xí)習(xí)慣、筆記能力和總結(jié)能力。根據(jù)既往研究，均值矩陣如矩陣M所示，協(xié)方差矩陣如矩陣Σ所示，設(shè)定檢驗(yàn)水準(zhǔn)為0.05，檢驗(yàn)效能為80%，試采用Wilks’ Lambda方法據(jù)此參數(shù)進(jìn)行樣本量估計(jì)。

nQuery Advanced實(shí)現(xiàn):設(shè)定檢驗(yàn)水準(zhǔn)為α= 0.05；檢驗(yàn)效能取80%。依據(jù)上述基礎(chǔ)數(shù)據(jù)可知，p=4，試驗(yàn)包含1個(gè)影響因素，其水平數(shù)為3，均值矩陣如矩陣M所示，協(xié)方差矩陣Σ如矩陣所示。在nQueryAdvanced 主菜單選擇：

Design：⊙Fixed Term

Goal：⊙Means

No.of Groups：⊙>Two

Analysis Method：⊙Test

方法框中選擇Multivariate Analysis of Variance(MANOVA)。

在彈出的樣本量計(jì)算窗口將各參數(shù)鍵入，如圖1-67a所示，均值矩陣和協(xié)方差矩陣如圖1-67b和圖1-67c所示，結(jié)果n和N分別為37和111，即每組樣本量為37例，總樣本量為111例。

圖1-67a nQuery Advanced 關(guān)于例1-28樣本量估計(jì)的參數(shù)設(shè)置與計(jì)算結(jié)果

圖1-67b nQuery Advanced 關(guān)于例1-28樣本量估計(jì)的參數(shù)設(shè)置與計(jì)算結(jié)果(均值矩陣)

圖1-67c nQueryAdvanced 關(guān)于例1-28樣本量估計(jì)的參數(shù)設(shè)置與計(jì)算結(jié)果(協(xié)方差矩陣)

SAS 9.4軟件實(shí)現(xiàn)：

%let f_level={3}; /*各分組變量的水平數(shù)*/

%let m= { 8.75 7.79 8.14,8.10 7.71 7.19,17.83 17.32 16.67,18.90 19.18 17.69};

/*因變量在所有分組變量各個(gè)水平組合下的均值*/

prociml;

start MGT3(alpha,power,p,f,rho,sd,sigma_def,method);

f_level=&f_level.;

m=t(&m.);

/*parameter check*/

error=0;

if ( alpha>=1 | alpha<=0 ) then do; error=1; print "Error" "Test significance Level must be in 0-1"; end;

/*參數(shù)f代表分組變量個(gè)數(shù)，不超過(guò)3個(gè)*/

if (f^=1 & f^=2 & f^=3) then do; error=1; print "Error""The number of factors must be in 1,2,3";end;

/*參數(shù)power代表各個(gè)效應(yīng)的檢驗(yàn)效能向量；當(dāng)研究包含A和B兩個(gè)分組變量時(shí)，可依次定義A、B及其交互效應(yīng)AB的檢驗(yàn)效能；當(dāng)研究包含A、B、C三個(gè)分組變量時(shí)，可依次定義A、B、C、AB、AC、BC、ABC的檢驗(yàn)效能*/

if f=1 & nrow(power)^=1 & ncol(power)^=1 then do;error=1;print "Error""Please input power of factor A";end;

if f=2 & nrow(power)^=3 & ncol(power)^=3 then do;error=1;print "Error""Please input power of factor A,factor B,factor AB (input "." if missing)";end;

if f=3 & nrow(power)^=7 & ncol(power)^=7 then do; error=1; print "Error""Please input power of factor A,factor B,factor C,factor AB,factor AC,factor BC,factor ABC (input "." if missing)"; end;

if (p<=0 | ceil(p)^=p) then do; error=1; print "Error" "The number of response variables must be a positive integer "; end;

if error=0 then do; q=1;do i=1 to f; q= q * &f_level.[i]; end;end;else stop;

if nrow(f_level)^=f then do;error=1; print "Error""The row number of factor levels should be equal to the number of factors";end;

if ncol(m)^=p then do; error=1; print "Error" "The row number of mean matrix should be equal to the number of response variables"; end;

if nrow(m)^=q then do; error=1; print "Error" "The column number of mean matrix should be equal to the product of factor levels"; end;

/*協(xié)方差矩陣可以由參數(shù)rho和sd生成，也可以由參數(shù)sigma_def直接定義*/

if (rho^=.& (rho>=1 | rho<=0)) then do; error=1;print "error" "Correlation must be in 0-1";end;

if (sd^=.& ( sd<=0 )) then do;error=1; print "Error" "Standard deviation at each level must be >=0"; end;

if (sigma_def^=.& (ncol(sigma_def)^=p | nrow(sigma_def)^=p)) then do;

error=1; print "Error" "The row and column numbers of covariance matrix should be equal to the number of response variables"; end;

/*參數(shù)method代表所選用的檢驗(yàn)統(tǒng)計(jì)量，其中1=Wilks' Lambda統(tǒng)計(jì)量，2=Pillai-Bartlett Trace統(tǒng)計(jì)量，3=Hotelling-Lawley Trace統(tǒng)計(jì)量。*/

if (method^=1 & method^=2 & method^=3) then do; error=1; print "Error""method must be in 1,2,3";end;

if (rho^=.& sd^=.) then do;

covariance=rho*sd**2;sigma=j(p,p,covariance);

do i=1 to p; sigma[i,i]=sd**2; end;

end;

else sigma=sigma_def;

if(error=1) then stop;

if(error=0)then do;

/*C matrix*/

max_level=&f_level.[<>,];

origin_c=j(max_level-1,max_level,0);

origin_j=j(max_level-1,max_level,.);

do i=2 to max_level;

e1=-(i-1)/sqrt(i*(i-1)); e2=1/sqrt(i*(i-1)); e3=j(1,i-1,e2); e4=1/sqrt(i);

e5=j(1,i,e4);

index=max_level+1-i;

origin_c[index,index]=e1;

origin_c[index,index+1:max_level]=e3;

origin_j[index,index:max_level]=e5;

end;

f1=&f_level.[1]; index1=max_level-f1+1;

C1=origin_c[index1:max_level-1,index1:max_level];

J1=origin_j[index1,index1:max_level];C_A=C1; power_a=power[1];

if f>1 then do;

f2=&f_level.[2]; index2=max_level-f2+1; C2=origin_c[index2:max_level-1,index2:max_level];

J2=origin_j[index2,index2:max_level];

C_A=C1@J2;C_B=J1@C2;C_AB=C1@C2;

power_b=power[2];power_ab=power[3];

end;

if f>2 then do;

f3=&f_level.[3];

index3=max_level-f3+1;

C3=origin_c[index3:max_level-1,index3:max_level];

J3=origin_j[index3,index3:max_level];

C_A=C1@J2@J3;C_B=J1@C2@J3;C_C=J1@J2@C3;

C_AB=C1@C2@J3;C_AC=C1@J2@C3;C_BC=J1@C2@C3;

C_ABC=C1@C2@C3;

power_c=power[3];

power_ab=power[4];power_ac=power[5];power_bc=power[6];

power_abc=power[7];

end;

/*power*/

%macro pw(c,pw_exp);

n_orig=j(1,q,1);n=j(1,q,1);

if &pw_exp.^=.then do;

if method=1 then do;

test="Wilks' Lambda";

do until (pw>=&pw_exp.);

n=n+n_orig;xx=t(n)#I(q);E=sigma * (n[<>,+]-q);

theta=&C.* m; H=t(theta)*inv((&C.*inv(XX)*t(&C.)))*theta;

T=H+E;df1=a*p;W=det(E*inv(T));fmm=a**2+p**2-5;

if fmm>0 then g=sqrt((a**2*p**2-4)/(a**2+p**2-5)); else g=1;

eta=1-W**inv(g);

df2=g*((n[<>,+]-q)-(p-a+1)/2)-(a*p-2)/2;

if df2>0 then do;

F_statistics=(eta/df1)/((1-eta)/df2);

f_c=finv(1-alpha,df1,df2);lambda=df1*F_statistics;

pw=(1-probf(f_c,df1,df2,lambda))*100; sum_n=n[<>,+];

end;

else pw=0;

end;

end;

if method=2 then do;

test="Pillai-Bartlett Trace";

do until (pw>=&pw_exp.);

n=n+n_orig;xx=t(n)#I(q);E=sigma * (n[<>,+]-q);

theta=&C.* m; H=t(theta)*inv((&C.*inv(XX)*t(&C.)))*theta;

T=H+E;df1=a*p;

ht=h*inv(t);PBT=trace(ht);s=min(a,p);

eta=pbt/s;df2=s*((n[<>,+]-q)-p+s);

if df2>0 then do;

F_statistics=(eta/df1)/((1-eta)/df2);

f_c=finv(1-alpha,df1,df2);lambda=df1*F_statistics;

pw=(1-probf(f_c,df1,df2,lambda))*100; sum_n=n[<>,+];

end;

else pw=0;

end;

end;

if method=3 then do;

test="Hotelling-Lawley Trace";

do until (pw>=&pw_exp.);

n=n+n_orig;xx=t(n)#I(q);E=sigma * (n[<>,+]-q);

theta=&C.* m; H=t(theta)*inv((&C.*inv(XX)*t(&C.)))*theta;

T=H+E;df1=a*p;

he=h*inv(E);hlt=trace(he);

s=min(a,p);eta=(hlt/s)/(1+hlt/s);

df2=s*((n[<>,+]-q)-p-1)+2;

if df2>0 then do;

F_statistics=(eta/df1)/((1-eta)/df2);

f_c=finv(1-alpha,df1,df2);lambda=df1*F_statistics;

pw=(1-probf(f_c,df1,df2,lambda))*100; sum_n=n[<>,+];

end;

else pw=0;

end;

end;

end;

else do; pw=.; sum_n=0; end;

%mend;

/*compute total sample size*/

a=&f_level.[1]-1; %pw(C_A,power_a);sum_n_a=sum_n;Smp_size=sum_n_a;

if f>1 then do;

a=&f_level.[2]-1;%pw(C_B,power_b);sum_n_b=sum_n;

a=(&f_level.[1]-1)*(&f_level.[2]-1);%pw(C_AB,power_ab);

sum_n_ab=sum_n;

Smp_size=j(3,1,.);

Smp_size[1]=sum_n_a;

Smp_size[2]=sum_n_b;

Smp_size[3]=sum_n_ab;

end;

if f>2 then do;

a=&f_level.[3]-1;%pw(C_C,power_c);pw_c=pw;sum_n_c=sum_n;

a=(&f_level.[1]-1)*(&f_level.[3]-1);

%pw(C_AC,power_ac);sum_n_ac=sum_n;

a=(&f_level.[2]-1)*(&f_level.[3]-1);

%pw(C_BC,power_bc);sum_n_bc=sum_n;

a=(&f_level.[1]-1)*(&f_level.[2]-1)*(&f_level.[3]-1);

%pw(C_ABC,power_abc);sum_n_abc=sum_n;

Smp_size=j(7,1,.);

Smp_size[1]=sum_n_a;

Smp_size[2]=sum_n_b;

Smp_size[3]=sum_n;

Smp_size[4]=sum_n_ab;

Smp_size[5]=sum_n_ac;

Smp_size[6]=sum_n_bc;

Smp_size[7]=sum_n_abc;

end;

Total_n=Smp_size[<>,];grp_size=Total_n/q;

/*compute power for total sample size*/

%macro pw2(c,pw_exp);

if &pw_exp.^=.then do;

n=j(1,q,grp_size);xx=t(n)#I(q);E=sigma * (n[<>,+]-q);

theta=&C.* m; H=t(theta)*inv((&C.*inv(XX)*t(&C.)))*theta;

T=H+E;df1=a*p;

if method=1 then do;

W=det(E*inv(T));fmm=a**2+p**2-5;

if fmm>0 then g=sqrt((a**2*p**2-4)/(a**2+p**2-5)); else g=1;

eta=1-W**inv(g);

df2=g*((n[<>,+]-q)-(p-a+1)/2)-(a*p-2)/2;

F_statistics=(eta/df1)/((1-eta)/df2);

end;

if method=2 then do;

ht=h*inv(t);PBT=trace(ht);s=min(a,p);

eta=pbt/s;df2=s*((n[<>,+]-q)-p+s);

F_statistics=(eta/df1)/((1-eta)/df2);

end;

if method=3 then do;

he=h*inv(E);hlt=trace(he);s=min(a,p);

eta=(hlt/s)/(1+hlt/s);df2=s*((n[<>,+]-q)-p-1)+2;

F_statistics=(eta/df1)/((1-eta)/df2);

end;

f_c=finv(1-alpha,df1,df2);lambda=df1*F_statistics;

pw=(1-probf(f_c,df1,df2,lambda))*100;

end;

else pw=.;

%mend;

Level= &f_Level.;

factor="Factor A";

a=&f_level.[1]-1; %pw2(C_A,power_a); pw_a=pw;power=pw_a;Alpha1=alpha;

if f>1 then do;

factor={"Factor A","FactorB","Factor AB"};

a=&f_level.[2]-1; %pw2(C_B,power_b); pw_b=pw;

a=(&f_level.[1]-1)*(&f_level.[2]-1); %pw2(C_AB,power_ab); pw_ab=pw;

power=j(3,1,.);power[1]=pw_a; power[2]=pw_b;power[3]=pw_ab;

Alpha1=j(3,1,alpha);

end;

if f>2 then do;

factor={"Factor A","FactorB","FactorC","FactorAB","FactorAC","FactorBC","Factor ABC"};

a=&f_level.[3]-1; %pw2(C_C,power_c); pw_c=pw;

a=(&f_level.[1]-1)*(&f_level.[3]-1); %pw2(C_AC,power_ac); pw_ac=pw;

a=(&f_level.[2]-1)*(&f_level.[3]-1); %pw2(C_BC,power_bc); pw_bc=pw;

a=(&f_level.[1]-1)*(&f_level.[2]-1)*(&f_level.[3]-1);

%pw2(C_ABC,power_abc);pw_abc=pw;

power=j(7,1,.);power[1]=pw_a; power[2]=pw_b;power[3]=pw_c;

power[4]=pw_ab;power[5]=pw_ac;power[6]=pw_bc;power[7]=pw_abc;

Alpha1=j(7,1,alpha);

end;

Mean_Matrix=&m.;

end;

print test[Label="Test"]

p[Label="Number of Response Variables"]

sd[Label="Common Standard Deviation"]

rho[Label="Correlation"]

grp_size[Label="Group Size"]

Total_n[Label="Total Sample Size"];

Print Factor [label="Factor"]

Level [Label="Level"]

Alpha1 [Label="Alpha"]

Power [Label="Power(%)"];

Print Mean_Matrix;

Print sigma[label="Covariance Matrix"];

finish MGT3;

%let sigma1={3.641 1.274 2.641 4.555,1.274 2.623 1.947 2.722,2.641 1.947 9.548 7.001,4.555 2.722 7.001 15.914};

%let power={80};

run MGT3(0.05,&power.,4,1,.,.,&sigma1.,1);

quit;

SAS 9.4運(yùn)行結(jié)果：

圖1-68 SAS 9.4關(guān)于例1-28樣本量估計(jì)的參數(shù)設(shè)置與計(jì)算結(jié)果