Using power/sampsi command in stata 13

[Monzur Alam]

Dear Statalisters,


I am trying to solve the following problem using power analysis in Stata and was hoping to get some advice.

“You are considering undertaking a pilot study of health in a population of interest, and you need to determine sample size. One indicator is systolic blood pressure (BP). Using mean and standard deviation measurements from “x.dta”, determine how large a sample you would need to be able to detect a 10 point differential between two groups (equal size) in your study, with group A having systolic BP at the level of the “x.dta” sample and Group B having BP 10 points higher. (Assume the SD of both distributions matches the x.dta overall value), at a statistical power of 0.8. Mean of BP of x.dta is 130.8817, sd is 23.33625”

In response to the above question, I wrote down 2 Stata commands.

sampsi 0 10, sd1(23.33265) sd2(23.33265) power(.8) onesided
power twomeans 130.8817, sd(23.33265) diff(10) power(0.8) onesided

I was wondering whether anyone could tell me:

1. Whether the above commands are correct
2. Whether it makes sense to use “onesided” or “twosided”
3. Why do the two commands give different values for sample size? Which one should I be using?

Thank you.

Monzur

[Carlo Lazzaro]

Manzur:
although your query looks suspiciously similar to a statistical assignment (and this kind of questions are out of the scope of ths forum), you laudably gave it a try before posting instead of addressing it simply to the list. hence, I think you problem is worth receiveing an answer:
- I think your codes are correct. However, please note that if you're using Stata 13.1 (otherwise, as per FAQ, you should declare it) -sampsi- is an out-of-date command, replaced by -pss-
However, I have quite the same results with both the commands (two-sided):

power twomeans 130.8817 130.9817, sd(23.3363)

Performing iteration ...

Estimated sample sizes for a two-sample means test
t test assuming sd1 = sd2 = sd
Ho: m2 = m1 versus Ha: m2 != m1

Study parameters:

alpha = 0.0500
power = 0.8000
delta = 0.1000
m1 = 130.8817
m2 = 130.9817
sd = 23.3363

Estimated sample sizes:

N = 1709746
N per group = 854873
----------------------------------------

sampsi 130.8817 130.9817, sd1(23.33625) sd2(23.33625) power(.80)

Estimated sample size for two-sample comparison of means

Test Ho: m1 = m2, where m1 is the mean in population 1
and m2 is the mean in population 2
Assumptions:

alpha = 0.0500 (two-sided)
power = 0.8000
m1 = 130.882
m2 = 130.982
sd1 = 23.3363
sd2 = 23.3363
n2/n1 = 1.00

Estimated required sample sizes:

n1 = 854870
n2 = 854870

I suppose that the slight difference in sample size is due to the iterative approach of -pss- (that I would use).

- I would go two-sided, unless you are confident that one-side inference is peculiar to your problems. If I recall it correctly, an interesting case of one-side hypothesis testing, as afar as the systolic BP is concerned, is the following one. When the mean systolic BP of air traffic controllers (ATCs) is tested against the mean reference value of the general population, a one-side t-test is usually performed, because, due to the previous reserches, ATCs have, on average, a higher mean systolic BP than general population.

[Svend Juul]

Well, maybe we shouldn't tell you. But Carlo works with a difference of 0.10 points; in the question the assumption is 10 points, giving quite a different sample size.

[Carlo Lazzaro]

Svend:
thanks for pointing out my unfortunate misreading.
Revised codes and results are reported below:

power twomeans 130.8817 140.8817, sd(23.3363)

Performing iteration ...

Estimated sample sizes for a two-sample means test
t test assuming sd1 = sd2 = sd
Ho: m2 = m1  versus  Ha: m2 != m1

Study parameters:

        alpha =    0.0500
        power =    0.8000
        delta =   10.0000
           m1 =  130.8817
           m2 =  140.8817
           sd =   23.3363

Estimated sample sizes:

            N =       174
  N per group =        87

sampsi 130.8817 140.8817, sd1(23.33625) sd2(23.33625) power(.80)

Estimated sample size for two-sample comparison of means

Test Ho: m1 = m2, where m1 is the mean in population 1
                    and m2 is the mean in population 2
Assumptions:

         alpha =   0.0500  (two-sided)
         power =   0.8000
            m1 =  130.882
            m2 =  140.882
           sd1 =  23.3363
           sd2 =  23.3363
         n2/n1 =     1.00

Estimated required sample sizes:

            n1 =       86
            n2 =       86

Other previous comments still holds.

[Monzur Alam]

Carlo:

Thank you very much for your help regarding the Stata commands, and for making me aware of the Stata forum rules!

Thanks, Svend for pointing out the differences in the numbers used.


Regards,

Monzur