Estadística Mèdica

Block 2 Clinical Trials

1. Sample Size & Randomization

 Block 2 RCTs – 1.-Sample size

“Reference” PRIMARY

ENDPOINT
treatment
(OUTCOME)

Randomized We compare
TOTAL PATIENTS Sample
allocation treatment effect

(Population)

“Active,
PRIMARY
experimental”
ENDPOINT
treatment
(OUTCOME)
Estadística Médica, Grau UB-UPC
 Block 2 RCTs – 1.-Sample size

PRIMARY
Control treatment
ENDPOINT

(OUTCOME)

Randomized We compare
TOTAL PATIENTS Sample
allocation treatment effect

(Population) Number of individuals?? (sample size)

New PRIMARY

treatment ENDPOINT

(OUTCOME)
Estadística Médica, Grau UB-UPC
 Block 2 RCTs – 1.-Sample size

Sample size to compare 2 means (independent samples)

H0 : mA - mB = 0

H1 : mA - mB = D or mB - mA = D (two-sided)

Let be: alpha risk, one or two-sided (usually a=0'05, two-sided),

some desired power (usually 1-b=0'8),

some given standard deviation (s), and

some expected effect size D to be tested

2 2 2
V(YA - YB) = V(YA) + V(YB) = s /n1 + s /n2 = 2s /n (given n1 = n2)

D = Z1-a/2 sÖ(2/n) - Zb sÖ(2/n)

And so, each sample size should be: n = [ 2 s² (z1-a/2 - zb)²] / D² (Bilateral)
Estadística Médica, Grau UB-UPC 4
(Unilateral (one-side test): use z1-a)
Block 2 RCTs – 1.-Sample size

Sample size to compare 2 means (independent samples)

Example:

Sample size to proof a gender height difference of D =10 cm?

Be σ=8 cm and usual risk values (a=0'05 ; b = 0'2), two-side test.

n = [ 2 ·8² (1'96 +0'84)²] / 10² =10’05 » 11 units per sample (round up)

 N = 11x2 = 22.

Note: round up n (sample size per patient), not N (N must always be pair).

Exercises:

1) What is needed to assess sample size? (What values do we need?)

2) What does D mean? Desired or expected effect size?

3) Can an observed (YA - YB) effect smaller to D still being statistically significant?

Estadística Médica, Grau UB-UPC

5
 Block 2 RCTs – 1.-Sample size

Sample size to compare 2 means (independent

samples)

Altman Nomogram

Overall size N

Group size n

N = 2·n

In the gender example,

standardized difference

10/8=1.25

Estadística Médica, Grau UB-UPC

6
 Block 2 RCTs – 1.-Sample size

Exercise:

Assume that =50 and you need to contrast H0: =0 versus H1: =10

1. What is your power with a sample size of n=10?

2. Draw a graph for the power (Y-axis) for

samples from n=1 to n=100 (X-axis).

Interpret.

3. Draw a graph for the power (Y-axis) of a sample of n=10

for H1: =1 to 80 (X-axis).

Interpret.

Estadística Médica, Grau UB-UPC

7
 Block 2 RCTs – 1.-Sample size

2.- La potencia según el tamaño ilustra cómo crece al aumentar el tamaño (n), ya que disminuye la oscilación aleatoria (/n)

Estadística Médica, Grau UB-UPC

8
 Block 2 RCTs – 1.-Sample size

3.- Al mirar la potencia en función de  (magnitud de H1) se ve el incremento de potencia que implica querer probar valores de  más alejados de

H0

Estadística Médica, Grau UB-UPC

9
 Block 2 RCTs – 1.-Sample size

OPTIMAL DESIGN

Let: overall available cases N, K treatments and 1 Placebo, number of cases in treatment k=n and in placebo=nP , so N = nP +

K·n

We want to optimize the K comparisons against P

Find the number of cases n to be allocated to each study arm

V(YA - YB) is

For K=1  n = nP = N/2

Estadística Médica, Grau UB-UPC

10
Block 2 RCTs – 1.-Sample size

Power and Sample size is related to 

Consider a baseline, before treatment, determination Z of the outcome variable Y:

Being:

or patient “idiosyncrasy” (among patients differences that remain constant on the two measured periods) with variance

or any measurement error and any patient non-stability either on the baseline (’ij) or on the outcome (ij) measurement period

So the outcome standard deviation (s) can be decomposed into:

V( = +

[Sometimes called ‘between’ and ‘within’.]

Estadística Médica, Grau UB-UPC

11
Block 2 RCTs – 1.-Sample size

Power: correlation between Z and Y

If we assume that ’ij, ij and i are mutually independent:

Exercise: proof it

And so, the correlation coefficient represents the proportion of variance among individuals to the total variance.

The correlation coefficient (amount of total variance shared by both determinations) can be interpreted as the amount of patient ‘idiosyncrasy’

i

If we further assume that the patient status doesn’t change, that is, that ij is only the result of the measurement error, the correlation

coefficient ρ corresponds to the reliability coefficient.

Please, note that this modelling implies that ρ should be positive.

Estadística Médica, Grau UB-UPC

12
Block 2 RCTs – 1.-Sample size

Power: strategies (1) to diminish  by controlling 

The change or outcome-baseline difference is defined as: Dij = Yij - Zij

and its variance is:

Usually, we assume .

In a two-arms (X=1, treated; and X=0, control) randomized clinical trial, the treatment effect is usually estimated by:

Depending upon the b value, we obtain:

the “final” estimator =

the “change score” =

or the “ANCOVA” =

being the least mean squares estimator with residual variance:

(+

Estadística Médica, Grau UB-UPC

13
 Block 2 RCTs – 1.-Sample size

Variance of each effect estimator

ˆ(0) : Final outcome ˆ(1) : Change from baseline ˆ (  ) : ANCOVA

q(  y2   z2  2  y z )
q y2 q(1   2 ) y2
2q( 1   ) y2 *

q 1 / n0  1 / n1 *
Provided σY=σZ

It implies that is the most efficient estimator for any value of ρ and that is more efficient than for ρ>0.5, but less efficient in the opposite situation.

Variance of each outcome

V( = + 𝑉 ( 𝐷𝑖𝑗 )=𝑉 ( 𝑌 𝑖𝑗 − 𝑍 𝑖𝑗 ) =𝜎 2𝜀 +𝜎 2𝜀𝑉 ( 𝐷𝑖𝑗 )=𝑉 ( 𝑌 𝑖𝑗 − 𝑍 𝑖𝑗 𝛽 ) =(1 − 𝜌 2)𝜎 2𝑌

′ ❑ ❑

2
(Usually, we assume .) ,where σ Y = V (Y ij ¿
❑

Estadística Médica, Grau UB-UPC

14
Block 2 RCTs – 1.-Sample size

Estadística Médica, Grau UB-UPC

15
Block 2 RCTs – 1.-Sample size

TOPIC: Analyzing “Change” Measures in RCT’s

Rationale: Many authors and pharmaceutical clinical trialists make the mistake of

analyzing change from baseline instead of making the raw follow-up measurements

the primary outcomes, covariate-adjusted for baseline. To compute change scores

requires many assumptions to hold (for more detail, see Frank’s post on this:

https://www.fharrell.com/post/errmed/#change 89). It is generally better to analyze

the “follow up” measurement as the outcome with a covariate adjustment for the

baseline value, as this seems to better match the question of interest: for two patients

Estadística Médica, Grau UB-UPC

16
with the same pre-trial value of the study outcome, one given treatment A and the
Block 2 RCTs – 1.-Sample size

Power: strategies (2) to diminish  by controlling 

If we obtain K independent repeated measures of the outcome Y,

and we average them:

Then, the variance of m(Y) is:

/K

Furthermore:

Estadística Médica, Grau UB-UPC

17
 Block 2 RCTs – 1.-Sample size

Exercise:

Systolic blood pressure (SBP) is an usual outcome variable in cardiovascular CTs.

In one particular population, with some specific measurement procedure, SDP has ² = (15mm/Hg)² and ² = (5mm/Hg)².

Assume trial objective  =5mm/Hg and usual risk values (=0'05 two-sided y =0'2 one-sided).

Calculate outcome variance and sample size in a 2 arms CT with:

1) A single SBP determination day 8 after treatment.

2) The average of 7 repeated SBP determination from days (D) 8 to 14.

3) The change score from day 0 to day 8.

4) The ANCOVA from day 0 to day 8.

5) The change score for two averages of 7 days.

6) Another alternative could be to study both treatments in the same patients, so that each patient provide two outcomes, one for each

treatment: YT and YP both averaged over 7 days.

7) Does it make sense to study the difference from baseline Z in the last design?

Estadística Médica, Grau UB-UPC

18
 Block 2 RCTs – 1.-Sample size

1) V(Y) = VE(Y)+VI(Y) = + = 225+25 = 250

n= [2² (Z/2 +Z)²]/² = [ 2· 250 (1.96 +0.84)²]/5² ≈ 156.8  157 each arm

2) ; 225 + 25/7 ≈ 228'57

n= [ 2· 228.57 (1.96 +0.84)²]/5² ≈ 143.4  144 each arm

3) V(Y8-Z0 ) = 2 = 2·25 = 50

n= [ 2 ·50 (1.96 +0.84)²] / 5² ≈ 31.4  32 each arm

2 2
4) =225/250=0.9  V = (1-  ) ²Y = (1-0.9 )·250 ≈ 47.5

n= [ 2 ·47.5 (1.96 +0.84)²] / 5² ≈ 29.8  30 each arm

5) V[m7(Y)-m7(Z)] = 2 /7 = 2·25/7 ≈ 7.14

n= [ 2 ·7.14 (1.96 +0.84)²] / 5² ≈ 4.48  5 each arm

6) V[m7(YT)-m7(YP)] = 2 /7 = 2·25/7 ≈ 7.14 . But now, just 1 (paired) sample:

n= [ 2 ·7.14 (1.96 +0.84)²] / 5² ≈ 4.48  5 patients in the only arm

7) No benefit: is already cleared, but 1 additional baseline ² is added.

Estadística Médica, Grau UB-UPC
19
Block 2 RCTs – 1.-Sample size

Estadística Médica, Grau UB-UPC

20
Block 2 RCTs – 1.-Sample size

Estadística Médica, Grau UB-UPC

21
Block 2 RCTs – 1.-Sample size

Estadística Médica, Grau UB-UPC

22