Sample Size Power Analysis
PwrSampleSize() allows to calculate the a
priori statistical power with different sample sizes. We can modify
any of the parameters, such as the time points of the experiment, the
growth rate of the different groups, or the variance in the model (given
by the standard deviation of the random effects and residuals).
However, usually the researcher will be interested in analyzing the
a priori power for an experiment, to check, for example, which
sample size should be used in order to reach an appropriate statistical
power. We can do this using the lmmModel_estimates()
function, to retrieve the model estimates from the experiment, and then
play with the different parameters:
First, lets extract the information from our experiment:
> # Vector with the time points
> days <- unique(grwth_data$Time)
>
> # Model estimates
> estimates <- lmmModel_estimates(lmm_ex_var)
Now, we can evaluate the a priori power for this experiment,
varying the sample size per group. For example, let’s evaluate how the
sample size varies from 1 to 10 subjects per group:
> PwrSampleSize(npg = 1:10,
+ time = days,
+ grwrControl = round(estimates$Control,3),
+ grwrA = round(estimates$DrugA,3),
+ grwrB = round(estimates$DrugB, 3),
+ grwrComb = round(estimates$Combination, 3),
+ sd_ranef = round(estimates$sd_ranef, 3),
+ sgma = round(estimates$sd_resid, 3),
+ method = "Bliss")
#> N Power
#> 1 1 0.2479771
#> 2 2 0.4491883
#> 3 3 0.6155323
#> 4 4 0.7414961
#> 5 5 0.8313399
#> 6 6 0.8926587
#> 7 7 0.9331019
#> 8 8 0.9590513
#> 9 9 0.9753250
#> 10 10 0.9853353
The plot on the left represents the hypothetical data that we have
provided, with the regression lines for each treatment group according
to grwrControl, grwrA, grwrB, and
grwrComb values. The values assigned to
sd_ranef and sgma are also shown.
The plot on the left shows the values of the power calculation
depending on the values assigned to npg. We can observed
that, for this experiment, a sample size of 5 would be enough to reach
0.8 statistical power.
Time Power Analysis
The number of measurement points per subject is another experimental
factor that can influence the experimental design and statistical power.
This depends on the duration of tumor growth follow-up and the frequency
of the measurements during that period. PwrTime() function
allows for the a priori power calculation of a a hypothetical
two-drugs combination study of synergy depending on the time of
follow-up or the frequency of measurements.
For example, the researcher may be interested in studying the
statistical power of the experiment depending on the time of follow-up,
to decide the appropriate end point. If we set the argument
type = "max" in PwrTime() function, we can
evaluate how the power varies at different end times. For this, we need
to provide the function with a list of vectors with the times at which
the tumor measurements are performed. As we are interested in addressing
how the power varies for different study duration, the measurement
should be taken at the same intervals.
Let’s see an example. Using the estimates from the model we have
fitted previously, we can explore how the power varies if the study ends
in a range from 9 to 30 days, with measurements performed every 3
days.
The first step is to define the list of vectors tumor measurements
time points:
> max_time <- list(seq(0,9,3), seq(0,12,3), seq(0,15,3),
+ seq(0,18,3), seq(0,21,3), seq(0,24,3),
+ seq(0,27,3), seq(0,30,3))
Another argument for the function is npg, which
indicates the sample size per group. As we are interested in calculate
the a priori power for different study duration in our
experiment and model, we will use the mean sample size per group in the
experiment:
> # We can calculate the average sample size dividing the number of subjects
> # by the number of groups, in this case, 4 groups
> (npg <- round(length(unique(grwth_data$subject))/4,0))
#> [1] 8
Now, the a priori power for the Bliss independence model at
different end points of the study is:
> PwrTime(npg = npg,
+ time = max_time,
+ type = "max",
+ grwrControl = round(estimates$Control,3),
+ grwrA = round(estimates$DrugA,3),
+ grwrB = round(estimates$DrugB, 3),
+ grwrComb = round(estimates$Combination, 3),
+ sd_ranef = round(estimates$sd_ranef, 3),
+ sgma = round(estimates$sd_resid, 3),
+ method = "Bliss")
#> Time Power
#> 1 9 0.3619997
#> 2 12 0.5820609
#> 3 15 0.7506471
#> 4 18 0.8506838
#> 5 21 0.9043890
#> 6 24 0.9332181
#> 7 27 0.9494116
#> 8 30 0.9590513
We can observe that a study with a follow-up time of 18 days would be
enough to achieve a statistical power > 0.8, if all the other
variables do not vary.
Now, we may also wonder how frequently the measurements should be
taken. Is it necessary to take measurements every 3 days, or can we
reduce the measurements to every 6 days?
We can do this by setting the type = "freq" argument in
PwrTime() function. In this case, we need to provide a list
of vectors with the tumor measurements with the same maximum time of
follow-up, but changing the intervals at which the measurements have
been taken. Let’s first define this list. We will evaluate how many
measurements are needed in an experiment with a follow-up period of 18
days.
> freq_time <- list(seq(0,18,1), seq(0,18,3), seq(0,18,6), seq(0,18,9),seq(0,18,18))
Now, let’s analyze the a priori power:
> PwrTime(npg = npg,
+ time = freq_time,
+ type = "freq",
+ grwrControl = round(estimates$Control,3),
+ grwrA = round(estimates$DrugA,3),
+ grwrB = round(estimates$DrugB, 3),
+ grwrComb = round(estimates$Combination, 3),
+ sd_ranef = round(estimates$sd_ranef, 3),
+ sgma = round(estimates$sd_resid, 3),
+ method = "Bliss")
#> Time Power
#> 1 19 0.9416015
#> 2 7 0.8506838
#> 3 4 0.7512980
#> 4 3 0.6911254
#> 5 2 0.6123858
The plot at the right shows the power for different number of
(evenly) distributed measurements. 2 measurements correspond to an
experiment in which the tumor measurements were taken just at the
initial time point (day 0), and at the final time point (day 18), while
19 measurements correspond to an experiment in which the measurements
were taken every single day, including day 0.
The results show that with 4 measurements, which corresponds to take
measurements every 6 days, the statistical power is close to 0.8.
Variability Power Analysis
A key determinant of statistical power is the effect size of the drug
combination effect, given by the magnitude of the estimated coefficient
for the growth rate after combination treatment, relative to that
observed in the monotherapy and control groups. Treatments with larger
effect sizes are easier to detect, even with smaller sample sizes,
resulting in higher statistical power. Another critical factor is the
variance components of the model, including residual variance
(within-subject variance) and random effects variance (between-subject
variance). High estimated coefficients for these variances reduce
precision, thereby diminishing statistical power.
APrioriPwr() function allows for total customization of
an hypothetical drug combination study and allows the user to define
several experimental parameters, such as the sample size, time of
measurements, or drug effect, for the power evaluation of synergy for
Bliss and HSA reference models.
Usually, we will define a experiment with the estimated parameters
from the input data and fitted model, and then play with different
values of the effect size of the drug combination effect, and/or values
of residual and random effects variance.
From our model, the estimated parameters are:
> estimates
#> Control DrugA DrugB Combination sd_ranef sd_resid
#> estimate 0.07888867 0.07496401 0.06308966 0.03535055 0.0386823 0.2148223
Then, we can for example, evaluate how the statistical power varies
if the drug combination group estimate ranges from -0.03 to 0.06, or if
the standard deviation of the random effects ranges from 0.01 to 0.1 and
the standard deviation of the residuals ranges from 0.01 to 1, using the
Bliss independence model:
> APrioriPwr(npg = npg, # Sample size per group, calculated above
+ time = days, # Time points of measurements, calculated above
+ # Model estimates:
+ grwrControl = round(estimates$Control,3),
+ grwrA = round(estimates$DrugA,3),
+ grwrB = round(estimates$DrugB, 3),
+ grwrComb = round(estimates$Combination, 3),
+ sd_ranef = round(estimates$sd_ranef, 3),
+ sgma = round(estimates$sd_resid, 3),
+ sd_eval = seq(0.01, 0.1, 0.01),
+ sgma_eval = seq(0.01, 1, 0.01),
+ grwrComb_eval = seq(-0.05, 0.1, 0.001)
+ )
#> numDF denDF F-value nc Power
#> 1 1 316 13.77187 13.77187 0.9590513
The red dot and the vertical line in the two plots on the right
indicate the power result corresponding to the values assigned to
sd_ranef, sgma, and grwrComb,
respectively.
The results demonstrate that the power increases as the residual and
random effects variances decrease. Furthermore, the plot on the right
reveals a U-shaped power profile as a function of the combination group
growth rate. This indicates that with lower growth rates (i.e., higher
effect size), the power is increased for detecting synergy, but it
decreases as growth rates approach the expected additive effect under
the Bliss independence model until it reaches its minimum. As growth
rates exceed the expected additive effect, the power starts to increase
again, reflecting the model’s ability to detect negative deviations from
additivity (indicating antagonism).
All these functions constitute an interesting toolbox for the
researcher. One may perform a first pilot experiment, and then optimize
the study design based on the data from that experiment.
It seems that there are 4 subjects with a higher influence in the group
coefficients, being subject 8 the one with the most remarkable
influence.
Again, subject 8 seems to be having a higher influence in the model.
Since we performed 1000 simulations, the minimum p-value is 0.001. If we
would like to obtain more precise p-values, we should run the analysis
increasing the number of simulations.