# SWSamp for simulation-based sample size calculations in a Stepped Wedge Trial (and more)

## Introduction

Sample size calculations are one of the fundamental components in the design of experimental studies and are mandatory in virtually all settings involving randomised trials, eg in medical research. Sample size calculations aim at determining the smallest sample that is necessary to observe, under a given design (eg distributional assumptions and expected characteristics of the intervention(s) being assessed), in order to correctly determine the “signal” (eg the “treatment effect”) as statistically significant and thus not due to chance.

Usually, sample size calculations are based on analytic formulae, often relying on some simplifying assumption, such as (asymptotic) Normality of the outcome. Once the significance level (type I error or false positive rate, typically $$\alpha=0.05$$) and the type II error (false negative rate, typically $$\beta=0.2$$) are set, a formula is obtained linking the “power” $$(1-\beta)$$ to the expected effect, a measure of variability in the relevant population and the number of observations, $$n$$. Solving the equation for $$n$$ provides the required sample size. Notice that, while slightly abusing the common notation, in order to avoid confusion we prefer to indicate $$\alpha$$ as “$$sl$$” and $$(1-\beta)$$ as “$$power$$”, in the rest of this tutorial. This allows us to use Greek letters to indicate parameters in the models we consider.

The classic example of a sample size calculation based on an analytic formula is given by the case of an outcome $$y\sim\mbox{Normal}(\mu+\theta t,\sigma^2)$$, where $$\mu$$ is a baseline level, $$t=0,1$$ is a “treatment” indicator, $$\theta$$ is the alleged treatment effect and $$\sigma^2$$ is the (pooled) population variance, assumed common across the two groups of treatment. Under these assumptions, the power is determined as $power = \Phi\left(\sqrt{\frac{n\theta^2}{4\sigma^2}}-z_{sl/2}\right),$ where $$\Phi$$ is the cumulative standard Normal distribution and $$z_{sl/2}$$ is its $$(1- sl/2)-$$th quantile. This can be solved for $$n$$ to determine the optimal sample size $n = \frac{2\left(z_{\rm{\textit{power}}} + z_{1-\rm{\textit{sl}}/2}\right)^2}{\theta^2}.$

While extremely helpful in principle, sample size calculations based on closed-form analytic solutions may suffer from some drawbacks. First, as a general point often the input parameters (eg the expected treatment effect or the relevant population variance) are plugged in the formula as point estimates. This does not account for any uncertainty in their “true” value, which may lead to biased estimates of the required sample size. In addition, there may be complex designs for which closed-form solutions do not exist or if they do, they rely on asymptotic results or approximations, which may render their applicability less general.

One such case is perhaps represented by the Stepped Wedge (SW) design, a variant of cluster randomised trials (CRTs) where all clusters receive the intervention in a randomised order — the main difference between a SW trial (SWT) and a CRT is that in the former design all groups eventually receive the intervention, while in a standard CRT the groups are randomly assigned to either treatment groups (eg intervention or control).

The panel below shows a very basic example of as SWT, conducted with $$I=6$$ clusters and for $$J=3$$ active time points at which the units are randomised to switch permanently to the treatment arm. In this case, all clusters would be measured at baseline, where all are in the control arm - this could be considered as a preliminary step before the intervention is implemented at all. Then, clusters 3 and 6 are randomised to switch to the active treatment at time point 1 and continue with this ever since. At time point 2, clusters 1 and 4 are also randomised to treatment and at time point 3 the remaining clusters 2 and 5 join the treatment arm.

   I = number of clusters: 6
J = number of time points: 3 (not including baseline)
H = number of units randomised at each time: 2

Baseline  Time1  Time2  Time3
Cluster 1      0        0      1      1
Cluster 2      0        0      0      1
Cluster 3      0        1      1      1
Cluster 4      0        0      1      1
Cluster 5      0        0      0      1
Cluster 6      0        1      1      1

0 = control (no intervention, Treatment as Usual, ...)
1 = active intervention

Information about the randomisation scheme is included in a design matrix $$\boldsymbol{X}$$, in this case $\boldsymbol{X} = \left( \begin{array}{cccc} 0 & 1 & 1 & 1 \\ 0 & 0 & 0 & 1 \\ 0 & 1 & 1 & 1 \\ 0 & 0 & 1 & 1 \\ 0 & 0 & 0 & 1 \\ 0 & 1 & 1 & 1 \end{array} \right).$

There are several potential complications with the desing of SWTs, as systematically reviewed and analysed for example by Hargreaves et al. (2015), Beard et al. (2015) and Copas et al. (2015). These, however, are beyond the purpose of the current work and thus we do not explore them further here, unless they are directly relevant.

## Analytic sample size calculations for a SWT

Analytic formulae for sample size calculations exist for specific SWTs. For example, Hussey and Hughes (2007) considered a SW design with cross-sectional data assuming $$I$$ clusters, $$J$$ crossover points and $$K$$ individuals sampled per cluster at each time point. In the most basic formulation, the observed continuous response is then modelled as $$Y_{ijk}=\mu_{ij}+e_{ijk}$$, where $\begin{eqnarray} \mu_{ij} = \mu+ \alpha_i + \beta_j + X_{ij}\theta \tag{1} \end{eqnarray}$ is the cluster- and time-specific mean, while $$e_{ijk}\sim \mbox{Normal}(0,\sigma^2_e)$$ represent independent individual-level error terms (within-cluster variability). Here, $$\mu$$ is the overall intercept, $$\alpha_i\sim\mbox{Normal}(0,\sigma^2_\alpha)$$ are a set of cluster-specific random effects, $$\beta_j$$ are fixed effects for time $$j$$, $$X_{ij}$$ is an intervention indicator taking on the value 1 if cluster $$i$$ is allocated to the active intervention at time $$j$$ and 0 otherwise, and $$\theta$$ is the intervention effect. This model implies that the response $$Y_{ijk}$$ is normally distributed with mean $$\mu_{ij}$$ and total variance $$\sigma^2_y=\sigma^2_\alpha+\sigma^2_e$$, while the cluster-level variance is $$\frac{\sigma^2_\alpha+\sigma^2_e}{K}\left[ 1+(K-1)\rho \right]$$, where $$\rho=\frac{\sigma^2_\alpha}{\sigma^2_\alpha+\sigma^2_e}$$ is the (ICC), a measure of the proportion of the total variance due to variation between the clusters.

For the basic model considered by Hussey and Hughes (HH), power calculations can be obtained by using the following relationship $power=\Phi\left(\frac{\theta}{\sqrt{V(\theta)}}-z_{sl/2}\right),$ where $$V(\theta)$$ is the variance of the estimator of the intervention effect $$\theta$$, defined as a relatively simple function of the design matrix $$\boldsymbol{X}$$. This formula can be used to determine the optimal sample size $$n$$.

### Simulation-based sample size calculations

The use of a simulation-based approach to determine the optimal sample size for a study is not a new concept, nor is it specific to the design of SWTs (Gelman and Hill 2006; Burton et al. 2006; Landau and Stahl 2013). In a nutshell, the idea is to consider a model to represent the data generating process (DGP), which describes how the researchers envisage the way in which the trial’s data will eventually be observed. This should be the model that is used to analyse the data, after the study has been conducted. Using the assumed DGP, data can be simulated a large number of times and the resulting “virtual trials” can be analysed using the proposed analysis model.

Some of the parameters may be varied across the simulations: for example, it is interesting to investigate the results obtained by varying the total number of observations. The optimal sample size is set to the minimum number of subjects for which the proportion of simulated trials that correctly deem the intervention as significant at the set $$sl-$$level is greater than or equal to the required power.

The main advantage of using simulation-based approaches to determine the sample size is that, in principle, any DGP can be assumed, no matter how complex. Of course, trials associated with more complicated designs will also require longer computational time to produce a sufficient number of runs to fully quantify the operating characteristics, eg in terms of the relationship between power and sample size. This is essential to estimate the required sample size properly.

In the case of a SWT, regardless of the modelling assumptions for the outcomes or the form assumed for the cluster- and time-specific mean, the simulation procedure can be schematically described as follows.

• i. Select a total sample size $$n$$ (eg total number of individuals measured) and a suitable combination of the number of clusters $$I$$ and time points $$J$$.
• ii. Provide an estimate of the main parameters. These can be derived from the relevant literature or expert opinion. We recommend thorough sensitivity analyses to investigate the impact of these assumptions on the final results, in terms of optimal sample size.
• iii. Simulate a dataset of size $$n$$ from the assumed model.
• iv. Analyse the resulting dataset and record whether the intervention effect is detected as statistically significant.

Steps iii. and iv. are repeated for a large number $$S$$ of times for each of the selected values of $$n$$ and the proportion of times in which the analysis correctly detects as significant the assumed intervention effects is used as the estimated power. The lowest value of $$n$$ in correspondence of which the estimated power is not less than the pre-specified threshold (usually, 0.8 or 0.9) is selected as the optimal sample size. A Monte Carlo estimate of the error around the estimated power can be easily computed and used as a guideline to determine the optimal number of simulations to be used. In many situations, a value of $$S$$ in the order of 1000s will suffice.

Sensitivity to the choice of the fundamental parameters can be checked by selecting different values and repeating the procedure. For example, it is possible to assess the impact of varying the cluster size. An alternative version of this algorithm may involve the adoption of a fully Bayesian approach (Spiegelhalter, Abrams, and Myles 2004, and @Cunananetal:2016); this amounts to modelling the uncertainty in the basic parameters using suitable probability distributions. For example, one could assume that, based on currently available evidence, the between-cluster standard deviation is likely to lie in a range between two extreme values $$a$$ and $$b$$. This may be translated, for example, into a prior Uniform distribution defined in $$(a,b)$$. The sample size calculations would then account for the extra uncertainty in the actual value of this parameter. The benefits of this strategy are of course higher if genuine information is available to the researchers.

Baio et al. (2015) introduce a general framework for sample size calculations in a SWT using a simulation-based approach. In this paper, we explore the technical details of that framework and introduce the R package SWSamp.

## The R package SWSamp

SWSamp is designed to allow a wide range of simulation-based sample size calculations, specifically (but not exclusively!) for a SWT. In its current version, SWSamp consists of 5 main functions: the first one (which is currently in fact specified by three different commands) performs the analytic sample size calculations using the method of HH. This can be used as a quick alternative for more standard designs or as a first-order approximation in the case of complicated designs (eg including multiple layers of correlation). The second function replicates the sample size calculations based on the (correct form of) the “design effect” specified by Woertman et al. (2013). This too can be used very effectively in relatively standard cases, but is less efficient in cases of more complex designs. The core functions of SWSamp are make.swt, which can be used to simulate data as obtained by a reasonably wide range of possible SWTs and sim.power, which actually performs the simulation-based computation of the required sample size. As we will show in the following, sim.power can be used for other DGPs (ie not specifically for a SWT), which increases the applicability of SWSamp.

In order to speed up the computation process (which can be intensive, under particularly complex designs), SWSamp is designed to take full advantage of R multi-core processing. For this reason, SWSamp “imports” the NAMESPACE of the R packages foreach, doParallel, iterators and parallel (technically, this means that these packages need to be installed on the user’s local machine, for SWSamp to work). In addition to these, SWSamp formally “depends” on the R package lme4, which can be used to fit linear and generalized linear mixed-effects models. This is relevant when analysing a SWT (which invariably includes some form of clustering in the DGP). Other modelling strategies could be used instead, for example Generalised Estimating Equations (GEE; Zeger and Liang 1986) or a full Bayesian model, e.g. based on Integrated Nested Laplace Approximation (INLA; Rue, Martino, and Chopin 2009). These extensions are under development or testing in the current version of SWSamp.

Detailed instructions for installation can be found here.

## A simple example for a SWT: analytic sample size calculations

As a first simple example, consider a SWT based on a continuous outcome for which the experimenter is assuming the following setting:

• $$J=5$$ time points at which measurements are taken, in addition to a baseline time — this amounts to a total of $$J+1=6$$ measurements;
• An average cluster size of $$K=20$$ individuals;
• A (relatively large!) cluster-level ICC $$\rho=0.5$$, perhaps estimated using previous data from similar studies;
• A baseline average value for the outcome $$\mu=0.3$$;
• A treatment effect $$\theta=-0.3875$$, implying a reduction in the outcome level for the treated units;
• A standard deviation $$\sigma_e=1.55$$;
• A cross-sectional design, in which measurements are taken at discrete time points. Individuals are measured only once and at the next time point it is assumed that the sample is made by independent units (the case of a Intensive Care Unit, ICU, where patients are assumed to spend a small amount of time; if the distance between consecutive time points is reasonably large, it can be assumed that the sample of patients at time $$j$$ is different than that observed at time $$j+1$$);
• A pre-specified significance level of 0.05 and a target power of at least 0.8.

### Analytic sample size calculations based on HH: Normally distributed outcome

Because of the specific design considered here, it is possible to directy use HH’s exact formula to compute the power, for any given value of the number of clusters $$I$$. For instance, we could use the SWSamp function HH.normal in the following way.

library(SWSamp)
HH.normal(mu=0.3,b.trt=-0.3875,sigma=1.55,I=14,J=5,K=20,rho=.5)
$power [1] 0.8112651$sigma.y
[1] 2.192031

$sigma.e [1] 1.55$sigma.a
[1] 1.55

$setting$setting$n.clusters [1] 14$setting$n.time.points [1] 5$setting$avg.cluster.size [1] 20$setting$design.matrix Baseline Time 1 Time 2 Time 3 Time 4 Time 5 6 0 1 1 1 1 1 14 0 1 1 1 1 1 8 0 0 1 1 1 1 10 0 0 1 1 1 1 9 0 0 1 1 1 1 2 0 0 0 1 1 1 11 0 0 0 1 1 1 1 0 0 0 1 1 1 4 0 0 0 0 1 1 12 0 0 0 0 1 1 3 0 0 0 0 1 1 13 0 0 0 0 0 1 5 0 0 0 0 0 1 7 0 0 0 0 0 1 The last element returned as output is a design matrix with a suggested randomisation list - in this case, clusters 10 and 4 switch to the intervention at the first active time, clusters 13, 12 and 8 switch at the second active time point and so on. The function HH.normal has three additional inputs: sig.level is the significance level and by default is set at 0.05. If this is acceptable for the user, then there is no need to specify it explicitly. Similarly, the argument which.var is a text string taking values on either 'within' (the default) or 'total'. If nothing is specified (eg, in this case), then SWSamp will assume that the value given in the argument sigma is for $$\sigma_e$$, the residual standard deviation and will compute the total and cluster-specific standard deviations $$\sigma_y$$ and $$\sigma_\alpha$$, respectively, using the fact that by definition $$\sigma_\alpha =\sqrt{\rho\sigma_e^2/(1-\rho)}$$ and $$\sigma_y=\sqrt{\sigma_\alpha^2+\sigma_e^2}$$. On the other hand, if the user sets which.var='total', then the value of 1.55 will be associated with the total standard deviation $$\sigma_y$$; HH.normal then computes and $$\sigma_\alpha=\sqrt{\sigma^2_y\rho}$$ and $$\sigma_e=\sqrt{\sigma^2_y-\sigma^2_\alpha}$$. In the case described above, since no value is specified for the argument which.var, SWSamp estimates $$\sigma_e=1.55$$, the value set for the input sigma and because $$\rho$$ is set to 0.5 then $$\sigma_e=\sigma_\alpha$$. Finally, the third argument is X, which indicates the (re-ordered, so as to become triangular) SWT design matrix and takes the default value NULL. When nothing is specified, then SWSamp will automatically compute a “balanced” design matrix, trying to divide as evenly as possible the $$I$$ clusters across the $$(J+1)$$ time points. In this case, the ratio $$I/J$$ is not an integer and thus the allocation is done by having 2 clusters switching at the first active time point and then having 3 clusters switching at each of the remaining five active time points. This means that out of the $$I(J+1)=84$$ cells of the design matrix, 48% are “1”s, indicating instances where a cluster is allocated to the intervention arm. While in this case the effect of this choice is likely to be minimal in terms of the resulting power, there may be cases where the automatic allocation favours the intervention arm (ie produces a design matrix in which the number of “1”s is much greater than the number of “0”s). For example, suppose the user specified a design matrix using the following code X <- matrix(0,14,6) X[1:4,2:6] <- X[1:8,3:6] <- X[1:10,4:6] <- 1; X[1:12,5:6] <- 1; X[1:14,6] <- 1 X [,1] [,2] [,3] [,4] [,5] [,6] [1,] 0 1 1 1 1 1 [2,] 0 1 1 1 1 1 [3,] 0 1 1 1 1 1 [4,] 0 1 1 1 1 1 [5,] 0 0 1 1 1 1 [6,] 0 0 1 1 1 1 [7,] 0 0 1 1 1 1 [8,] 0 0 1 1 1 1 [9,] 0 0 0 1 1 1 [10,] 0 0 0 1 1 1 [11,] 0 0 0 0 1 1 [12,] 0 0 0 0 1 1 [13,] 0 0 0 0 0 1 [14,] 0 0 0 0 0 1 This would imply a much higher proportion of the cells in $$\boldsymbol{X}$$ that are “1”s (57%). If the power was computed using this specification, the results would be slightly different than the one in which the design matrix is automatically determined to have an allocation that is as even as possible. auto <- HH.normal(mu=0.3,b.trt=-0.3875,sigma=1.55,I=14,J=5,K=20,rho=.5) user <- HH.normal(mu=0.3,b.trt=-0.3875,sigma=1.55,I=14,J=5,K=20,rho=.5,X=X) user$power
[1] 0.8027561

If on the other hand, the user had specified yet another design matrix, this time defined as

X2 <- matrix(0,14,6)
X2[1:2,2:6] <- X2[1:4,3:6] <- X2[1:6,4:6] <- X2[1:8,5:6] <- X2[1:14,6] <- 1

(this implies that only 40% of the cells are “1”s) and then computed the analytic power, the results would be slightly different again.

user2 <- HH.normal(mu=0.3,b.trt=-0.3875,sigma=1.55,I=14,J=5,K=20,rho=.5,X=X2)
user2$power [1] 0.7971512 As a general rule, the application of a “balanced” design will tend to increase the power by balancing the number of measurements in either arm of the trial. ### Analytic sample size calculations based on HH: non-Normally distributed outcome HH’s computations can be brought to bear even in cases where the outcome is not directly Normally distributed, by considering asymptotic arguments. For instance, we could consider a SWT where the outcome is the presence of bacterial infection among patients in Intensive Care Units (ICUs), which are sequentially moved to a regime of specialised cleaning. In this case, we could assume that the response for the $$k-$$th individual in cluster $$i$$ at time $$j$$ is $$Y_{ijk}\sim\mbox{Bernoulli}(\mu_{ij},n_{ijk})$$, where $$\mu_{ij}$$ and $$n_{ij}$$ are the cluster- and time-specific probability of infection and sample size, respectively. In this case, the model specified in (1) can be generalised~to $$$\phi_{ij} = g(\mu_{ij}) = \mbox{logit}(\mu_{ij}) = \log\left(\frac{\mu_{ij}}{1-\mu_{ij}}\right) = \alpha_i + \beta_j + X_{ij}\theta. \tag{2}$$$ The model in (2) is a simple generalised mixed linear model; more complex structures may be necessary (eg to account for additional layers of clustering, or include other covariates). To fit this model in SWSamp and determine the optimal sample size under given, pre-specified circumstances, it is possible to use the function HH.binary, for example using the following commands. # Defines the basic parameters p1 <- 0.26 # baseline probability of the outcome in the control group OR <- 0.56 # assumed treatment effect (in terms of odds ratio) J <- 5 # number of time points (excluding the baseline measurement) K <- 20 # assumed average cluster size rho <- 0.3 # assumed intra-class correlation sig.level <- 0.05 # assumed significance level which.var <- "within" # specifies whether the variance is within or total # Computes the power using the analytic formula HH.binary(p1,OR,I=8,J,K,rho,sig.level,which.var,X=NULL)$power
[1] 0.5276896

$p1 [1] 0.26$p2
[1] 0.1644083

$sigma.y [1] 0.485341$sigma.e
[1] 0.4060654

$sigma.a [1] 0.2658322$setting
$setting$n.clusters
[1] 8

$setting$n.time.points
[1] 5

$setting$avg.cluster.size
[1] 20

$setting$design.matrix
Baseline Time 1 Time 2 Time 3 Time 4 Time 5
1        0      1      1      1      1      1
8        0      0      1      1      1      1
4        0      0      1      1      1      1
5        0      0      0      1      1      1
2        0      0      0      0      1      1
3        0      0      0      0      1      1
6        0      0      0      0      0      1
7        0      0      0      0      0      1

This works pretty much in the same way as the function HH.normal and in fact most of the inputs are the same. What does change is the set of input values that are specific to the type of outcome: p1 instead of muand OR instead of b.trt. In addition, the Bernoulli model does not require the specification of the residual variance and thus there is no need to provide the argument sigma to the call to the function HH.binary.

As suggested by Baio et al. (2015), a related issue concerns the fact that, in the binary and count outcome cases, it is more cumbersome to assume that information is provided in terms of the total variance. This is because, unlike the Normal distribution, the Bernoulli and Poisson distributions are characterised by a single parameter, which simulataneously determines both the linear predictor and the variance. Consequently, because the linear predictor includes the cluster-specific random effects $$\alpha_i$$, assuming a fixed total variance $$\sigma^2_y$$ implies in (2) a re-scaling of the baseline value $$\mu$$ to guarantee that the resulting total variance approximates the required value. For this reason, when using a simulation-based approach for non-normally distributed outcomes it is easier to provide information on the within-cluster variance $$\sigma^2_e$$ as input.

In the case of count outcomes, SWSamp allows analytic computations using the function HH.count, which works in a very similar way. A call to this function would be done by using the following code.

HH.count(lambda1,RR,I,J,K,rho,sig.level,which.var,X)

where lambda1 is the baseline (controls) rate of occurrence of the event under investigation; RR is the treatment effect measured in terms of the relative risk; and similar arguments apply to the other inputs.

### Analytic sample size calculations based on the Design Effect

The SWSamp function DE.woert can be also used to compute the “Design Effect” (DE) for a SWT, based on the (correct) formulation of the equation provided by Woertman et al. (2013) and presented in Baio et al. (2015). In a nutshell, the DE is an inflation factor that corrects the analytic sample size based on an individual-randomised trial to account for the specific form of correlation induced by the SW design.

Consider for example a SWT in which the outcome is a binary variable and the treatment effect is assumed to translate to an odds ratio of 0.53 for the outcome of interest. The main inputs that are required for DE.woert are:

• outcome: a string (default 'cont' to indicate a continuous outcome, but possible values are 'bin' or 'count', for binary or count data, respectively);
• input: a R list containing the following arguments:
• For a continuous outcome
1. delta: treatment effect;
2. sd: pooled population standard deviation.
• For a binary outcome
1. p1: baseline probability of outcome;
2. either p2 (treatment probability of outcome), or OR (treatment effect asoddsratio).
• For a count outcome
1. r1: baseline rate of outcome;
2. either r2 (treatment rate of outcome), or RR (treatment effect as relative risk).
• K: the average cluster size;
• J: the number of time points (excluding baseline);
• B: the number of baseline measurement times (default = 1);
• T: the number of measurement times during each crossover point;
• rho: the ICC;
• sig.level: the significance level (default at 0.05);
• power: the required power (default at 0.8).

The following is a simple example of how the function DE.woert can be used.

input <- list(OR=.53,p1=.26)
K <- 20
J <- 5
rho <- .2
x <- DE.woert(outcome="bin",input=input,K=K,J=J,rho=rho)
x
$n.cls.swt [1] 11$n.pts.swt
[1] 1221.568

$DE.woert [1] 2.513514$CF
[1] 0.4189189

$n.rct [1] 486 The output of this function is: n.cls.swt, the number of clusters required for a SWT (in this case, 11); n.pts.swt, the total number of subjects that need to be observed under the current design (1221); DE.woert, the design effect due to clustering (2.51); CF, the correction factor attributable to the use of the SWT (0.41); and finally n.rct, the number of subjects that would be included in a RCT with the same power as that of the SWT. ### Simulation-based sample size calculations The SWSamp function sim.power can be used for different choices of the parameter I, indicating the number of clusters and assess which of these is the minimum required to hit the pre-specified power. For example, the following code can be used to estimate the power corresponding to such a design including $$I=14$$ clusters, using the following R call: library(SWSamp) x <- sim.power(I=14,J=5,K=20,rho=0.5,mu=0.3,sigma.e=1.55,b.trt=-0.3875,n.sims=100) which would simulate and analyse n.sims=100 “virtual trials” with the given design. The object x is a R list containing the following elements: • power: the estimated power for the current configuration. If the model does include random effects (which is the case for a SWT), then SWSamp assesses whether the “true” effect is correctly detected by computing the $$100\times(1-sl)$$% interval around the point estimate for the “treatment effect” and checking whether it is entirely above or below 0 (in which case, the results are deemed as “significant” for that particular simulated trial) The average of the simulations is the estimate of the power. The choice of focusing on estimation rather than hypothesis testing is also justified because it is in general difficult to assess the correct degrees of freedom of the resulting (linear) mixed model. The p-value could be computed using the approximation in Satterthwaite (1946), or by using a rougher Normal approximation. Neither however is free from problems, as suggested by Pinheiro and Bates (2000). • time2run: the computational time (in seconds); • ci.power: an estimated $$100\times(1-sl)$$% interval for the power. This is based on a relatively crude Normal approximation and computed using the results from the simulation procedure; • theta: the estimated treatment effect from the simulation procedure, together with an estimate of its variability; • rnd.eff.sd: the estimated standard deviation components, including the cluster and residual values indicated by $$\sigma_a$$ and $$\sigma_e$$, respectively, in the model described in equation (1); • setting: a R list summarising the assumptions in terms of number of clusters, time points, type of model, formula used. These can be explored by typing in R the following command. x$power
[1] 0.81

$time2run Time to run (secs) 0.834$ci.power
[1] 0.7327231 0.8872769

$theta Estimate Standard Error -0.3834323 0.1362724$rnd.eff.sd
cluster (Intercept)            Residual
1.497543            1.550484

$setting$setting$n.clusters [1] 14$setting$n.time.points [1] 5$setting$avg.cluster.size [1] 20$setting$design [1] "cross-sec"$setting$formula y ~ treatment + factor(time) + (1 | cluster) <environment: 0x9d63ce0>$setting$method [1] "lmer"$setting$family [1] "gaussian" The command sim.power has also some additional arguments that can be used to produce variants of the default setting or extra outputs. For example, the user can specify a string family to define the type of outcome: the default value is "gaussian", but other possibilities are "binomial" and "poisson", in which case the procedure runs a suitable generalised linear model. Another option is related to the type of design to be considered, which can be specified in the input design; currently, the options are "cross-sec" (the default choice) or "cohort" (which would consider a repeated measures design). Similarly, the user can specify a particular formula to describe the relationship between the outcome and the predictors. If nothing is specified, SWSamp assumes that the user wants to consider a model formula=y~treatment+factor(time)+(1|cluster). This uses the notation of the R package lme4 and considers an outcome named y; a “fixed effect” for a variable named treatment; a “fixed effect” for a categorical (factor) variable named time; and a “random effect” to account for correlation within the levels of a variable named cluster. The notation 1|cluster instructs lme4 that this model should consider a varying intercept for each levels of the clustering variable. When the user defines a custom formula, the name of the “treatment” variable can be specified as a string in the extra argument treat.name. If this is not present, then it is assumed that the name of the treatment variable is, unsurprisingly, "treatment". By default, sim.power performs an analysis based on a frequentist approach; if the model does contain clustering this is done using lme4, while for simpler structures (eg without random effects), the standard (g)lm function is used. However, if the user specifies method="inla", SWSamp performs the analysis of the simulated datasets fitting a Bayesian model using Integrated Nested Laplace Approximation. In this case, the resulting values are estimated from the relevant posterior distributions. Future developments will explore the use of Hamiltonian Monte Carlo to perform Bayesian analysis. To help in the computationally intensive task of analysing a very large number of simulated dataset, the user can take advantage of R multi-core capability, by specifying the option n.cores=K, where K is the number of CPUs to be used for parallel processing. By default, n.cores is set to NULL, which means that R will figure out the number of available cores and use all but one, to keep the machine sufficiently responsive. Finally, the option plot (default value at FALSE) instructs SWSamp on whether a plot showing the performance of the simulation procedure should be printed or not. If plot=TRUE, then a plot is created with the moving average of the computed power at successive batches of iteration (every 10%). This can be useful to assess whether the procedure has reached some sort of “convergence” to a stable value of the computed power. ## User-defined data generating processes Of course, each trial may have some characteristic that is just specific to the particular setting under consideration. Thus, SWSamp allows the user to specify a data generating process to be used in the analysis of the “virtual trials” and thus the estimation of the resulting power/required sample size. As a first example, consider a very simple trial based on a continuous outcome with two treatment groups: the first one (coded as $$t=0$$) is the control arm of the trial (eg individuals subject to the status quo, or a placebo), while the second one ($$t=1$$) is the group of individuals given the active treatment. This situation can be again modelled by assuming $$y\sim\mbox{Normal}(\mu+\theta t,\sigma^2)$$, where $$\mu$$ is the baseline outcome for the controls and $$\theta$$ is the treatment effect. The user could specify a function in R, something like simple.trial <- function(n,mu=0,theta,sigma) { x <- rbinom(n,1,.5) linpred <- mu+theta*x y <- rnorm(n,linpred,sigma) return(data.frame(y=y,x=x)) } The function simple.trial takes as arguments the overall sample size n, the baseline mean outcome mu (for which the default value is specified as 0), the treatment effect theta and the common standard deviation in the two groups sigma. For example, the following R call creates a dataset with 10 observations. data <- simple.trial(n=10,mu=0,theta=1,sigma=1) data y x 1 0.8795421 0 2 0.4308724 1 3 0.4312428 0 4 -1.3926546 0 5 0.7979760 1 6 0.5796582 1 7 2.2146694 1 8 1.1364324 1 9 -0.7489955 1 10 1.5454978 1 For such a design, it is possible to determine the optimal sample size analytically, for instance using the R function power.t.test, gives the following result power.t.test(delta=1,sd=1,sig.level=0.05,power=0.8) Two-sample t test power calculation n = 16.71477 delta = 1 sd = 1 sig.level = 0.05 power = 0.8 alternative = two.sided NOTE: n is number in *each* group which suggests that a total sample size of $$n=34$$ individuals is required (notice that the function power.t.test indicates the treatment effect by means of the parameter delta). The user can also perform this calculation based on simulations. For example as in the following. x <- sim.power(data=simple.trial,inpts=list(n=34,theta=1,sigma=1), formula=y~x,treat.name="x",n.sims=1000) x$power
[1] 0.805

$time2run Time to run (secs) 0.423$ci.power
[1] 0.7804314 0.8295686

$theta Estimate Standard Error 0.9948904 0.3444156$rnd.eff.sd
NULL

$setting$setting$formula y ~ x$setting$method [1] "lm"$setting$family [1] "gaussian" This time, the call to the function sim.power has different arguments than in the standard case. In fact, it is possible to specify the optional argument data which is simply the name of the user-defined function telling R what the data generating process to be used to create the “virtual trials” is; this can be complemented by the argument inpts, a R list including the values that the user wants to assign to the inputs of the function defined in data. In this case, simple.trial requires that the values for n, theta and sigma are given (and if nothing is specified for mu, the default value is taken). If the function specified as data had no input, then the optional argument inpts would not be necessary; if the user does not specify it, sim.power will assume that it is not needed and will automatically set it to NULL. When considering user-defined functions for the data generating process, it is probably necessary to specify a formula to instruct R and SWSamp as to what is the underlying model that should be used for the analysis of each simulated trial. In this simple case, we specify formula=y~x. In addition to this, SWSamp needs to know which of the covariate(s) is the “treatment” - this is because it needs to compute the “treatment effect” $$\theta$$ and then the resulting power based on whether it is detected as “significantly” different from 0. If the optional argument treat.name is not specified, sim.power assumes that one of the covariates is called 'treatment'; this is not the case in the current example and thus the user needs to tell sim.power that in fact the variable called 'x' is the one for which the treatment effect should be computed. Finally, we can specify the number of simulations to be used, in this case we selected n.sims=1000 - notice that since this is the default value, the inclusion of this extra argument is not essential. While analysing the simulated data, SWSamp checks that the model specification contains structured (random) effects, in the syntax of the R package lme4, ie using the form (1|x) to indicate a varying intercept for a covariate x, or (z|x) to indicate a random slope for the covariate z, depending on the levels of the covariate x. If random effects are present, SWSamp will analyse the simulated dataset using (g)lmer, the lme4 functions to fit (generalised) linear mixed models. If there are no random effects in the model specification (eg in the present example), then SWSamp will analyse the simulated datasets using a simple (generalised) linear model, ie the R command (g)lm. The choice of the distributional assumption is governed by the extra argument family that the user can specify to instruct SWSamp on the nature of the outcome. The default value is family='gaussian', indicating that the outcome is associated with a Normal distribution and thus the resulting analysis will be performed using a linear model (ie using lmer, in the presence of random effects). Other options include family='binomial' or family='poisson', in which cases a suitable generalised linear (mixed) model is used. For example, consider again a very simple DGP in which the outcome variable is $$y \sim \mbox{Binomial}(\pi,n)$$, with $$\mbox{logit}(\pi) = \mu + \theta t$$ and where $$\mu$$ and $$\theta$$ represent the baseline log-OR for the event described by $$y$$ (which applies to the controls) and the log-OR for the treatment effect, assuming $$t=0,1$$ is the treatment, respectively. A simple R function to simulate this DGP is the following bin.trial <- function(n,p1,OR) { p2 <- OR*(p1/(1-p1))/(1+OR*(p1/(1-p1))) x <- rbinom(n,1,.5) linpred <- log(p1/(1-p1)) + log(OR)*x pi <- exp(linpred)/(1+exp(linpred)) y <- rbinom(n,1,pi) return(data.frame(y=y,x=x)) } The function bin.trial takes as inputs the sample size n, the baseline probability of the outcome p1 and the odds ratio OR - obviously $\begin{eqnarray*} \mu =\log\left(\frac{p_1}{1-p_1}\right) \qquad \mbox{and} \qquad \theta = \log(\mbox{OR}). \end{eqnarray*}$ The power for this trial can be computed using sim.power by specifying x <- sim.power(data=bin.trial,inpts=list(n=n,p1=p1,OR=OR), formula=y~x,treat.name="x",family="binomial") assuming that suitable values for n, p1 and OR are specified by the user. For instance, setting n = 100, p1 = 0.54 and OR = 1.3 and running sim.power would return the following results $power
[1] 0.098

$time2run Time to run (secs) 0.542$ci.power
[1] 0.07956337 0.11643663

$theta Estimate Standard Error 0.2660182 0.4114141$rnd.eff.sd
NULL

$setting$setting$formula y ~ x$setting$method [1] "glm"$setting$family [1] "binomial" As a more complex example, we could actually define a function that creates data from a SW design. SWSamp has a built-in function for data generating processes that is suitable for a SWT; this is called make.swt and basically takes as inputs many of the arguments that are specified in the default call to sim.power. For example, we could define sw.trial <- function(){ make.swt(I=8,J=5,K=10,mu=.3,b.trt=-.3875,sigma.e=1.55,rho=.4) } data <- sw.trial() rbind(head(data),tail(data)) y person time cluster treatment linpred b.trt 1 2.8606711 1 0 4 0 3.747343 -0.3875 2 3.8102603 2 0 4 0 3.747343 -0.3875 3 5.2833223 3 0 4 0 3.747343 -0.3875 4 4.9194180 4 0 4 0 3.747343 -0.3875 5 1.1869019 5 0 4 0 3.747343 -0.3875 6 3.6239423 6 0 4 0 3.747343 -0.3875 475 -0.9707193 5 5 6 1 -0.685079 -0.3875 476 -0.2939763 6 5 6 1 -0.685079 -0.3875 477 2.5747190 7 5 6 1 -0.685079 -0.3875 478 0.5972963 8 5 6 1 -0.685079 -0.3875 479 -0.3935788 9 5 6 1 -0.685079 -0.3875 480 -2.8377084 10 5 6 1 -0.685079 -0.3875 This function generates a dataset that could arise from a SWT with $$I=8$$ clusters, $$J=5$$ time point (excluding the baseline), an average cluster size of $$K=10$$ individuals, a baseline outcome of $$\mu=0.3$$, a treatment effect of $$-0.3875$$, a cluster-level ICC $$\rho=0.4$$ and assuming a variance $$\sigma_e = 1.55$$. Executing the command sw.trial() generates the full dataset (which, incidentally contains 480 rows). Notice that since all the relevant inputs are defined inside the function sw.trial, there are no arguments to be added to the call to the function. As no argument design is specified in the call to make.swt, SWSamp assumes that the default cross-sec should be used, implying that the resulting dataset is constructed under a cross-sectional design, with individuals observed only once. The function sw.trial can be used as an argument to sim.power to estimate the power for this particular SWT. For example, the user can type x <- sim.power(data=sw.trial) x$power
[1] 0.335

$time2run Time to run (secs) 5.342$ci.power
[1] 0.3057316 0.3642684

$theta Estimate Standard Error -0.3868762 0.2527232$rnd.eff.sd
cluster (Intercept)            Residual
1.228314            1.547958

$setting$setting$formula y ~ treatment + factor(time) + (1 | cluster) <environment: 0xa831908>$setting$method [1] "lmer"$setting\$family
[1] "gaussian"

In this case, n.sims=1000 simulations are used (since no value is selected for this argument), which combined with the more complex DGP associated with the SWT implies a longer running time (about 11 seconds). The power for this configuration is estimated at 0.34 and likely to range between 0.31 and 0.36. This is in line with the fact that the estimates of the treatment effect theta (-0.39) is not obtained with large precision - the standard error is 0.25, which means an approximate 95% interval is [-0.89; 0.12], ie it crosses the threshold value of 0.

## Conclusions

SWSamp` is designed with the specific aim of dealing with sample size calculations for a Stepped Wedge trial; this implies the need to account for complications in the underlying set up, including different layers of structured effects, time trends and other issues that may be complex to comprehensively address using closed-form calculations.

However, we have made an effort to code the core functions so that the user can actually extend the general framework and apply the methods to a wider range of designs. Ideally, suitable functions specifying different data generating process will be made available (and the hope is to create a repository to which user can contribute) so that the process can be made extremely general and widely applicable in practice.

Simulations-based sample size calculations can be very useful in real-world applications, particularly with a view at aligning the model assumed to describe how the data are generated in the context of the trial under investigation and the one used to analyse the data, once they become available. Future extensions to the package involve the use of more complex analysis models (such as GEE or full Bayesian models).

## References

Baio, G, A Copas, G Ambler, J Hargreaves, E Beard, and RZ Omar. 2015. “Sample Size Calculation for a Stepped Wedge Trial.” Trials 16 (August): 354. https://doi.org/10.1186/s13063-015-0840-9.

Beard, E, J Lewis, A Prost, A Copas, C Davey, D Osrin, G Baio, et al. 2015. “Stepped Wedge Randomised Controlled Trials: Systematic Review.” Trials.

Burton, A, D Altman, P Royston, and R. Holder. 2006. “The Design of Simulation Studies in Medical Statistics.” Statistics in Medicine 25: 4279–92.

Copas, A, J. Lewis, J. Thompson, C Davey, K. Fielding, G. Baio, and J. Hargreaves. 2015. “Designing a Stepped Wedge Trial: Three Main Designs, Carry-over Effects and Randomisation Approaches.” Trials.

Cunanan, K. M., B. P. Carlin, and K. A. Peterson. 2016. “A practical Bayesian stepped wedge design for community-based cluster-randomized clinical trials: The British Columbia Telehealth Trial.” Clin Trials 13 (6): 641–50.

Gelman, A., and J. Hill. 2006. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge, UK: Cambridge University Press.

Hargreaves, J., A. Copas, E. Beard, D. Osrin, J. Lewis, C. Davey, J. Thompson, G. Baio, K. Fielding, and A. Prost. 2015. “Five Questions to Consider Before Conducting a Stepped Wedge Trial.” Trials.

Hussey, M, and J Hughes. 2007. “Design and Analysis of Stepped Wedge Cluster Randomised Trials.” Contemporary Clinical Trials 28: 182–91.

Landau, S, and S. Stahl. 2013. “Sample Size and Power Calculations for Medical Studies by Simulation When Closed Form Expressions Are Not Available.” Statistical Methods in Medical Research 22 (3): 324–45.

Pinheiro, J., and D. Bates. 2000. Mixed-effects models in S and S-PLUS. New York, NY: Springer.

Rue, H., S. Martino, and N. Chopin. 2009. “Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations.” Journal of the Royal Statistical Society: Series B (statistical methodology) 71 (2): 319–92.

Satterthwaite, F. E. 1946. “An Approximate Distribution of Estimates of Variance Components.” Biometrics Bulletin 2 (6): 110–14.

Spiegelhalter, D, K Abrams, and J. Myles. 2004. Bayesian Approaches to Clinical Trials and Health-Care Evaluation. London, UK: Wiley; Sons.

Woertman, W, E de Hoop, M Moerbeek, S Zuidema, D Gerritsen, and S. Teerenstra. 2013. “Stepped Wedge Designs Could Reduce the Required Sample Size in Cluster Randomized Trials.” Journal of Clinical Epidemiology 66 (7): 52–58.

Zeger, S., and K. Liang. 1986. “Longitudinal data analysis for discrete and continuous outcomes.” Biometrics 42 (1): 121–30.