Description
Find the sample size at which a network model recovers a hypothesised network to a target standard with a target probability, by iterating simulation, monotone spline fitting and bootstrapping.
Network models of survey items, such as Gaussian graphical models, need large samples, and there is no closed-form power formula. Constantin and colleagues' algorithm, implemented in the R package powerly, starts from a hypothesised true network, a performance measure with a target (for example sensitivity of 0.6) and a statistic with a target (reach it in 80% of replications). Step 1 simulates and re-estimates data at several sample sizes across a wide range. Step 2 fits a monotone spline to the statistic so it can be read between the sizes tried. Step 3 bootstraps the replications to put confidence bands on the spline. The range then narrows to where the bands cross the target, and the cycle repeats until the range is small. In their simulations the recommendation was on average within three observations of the benchmark, and the method generalises to other models.
Attributes
- Discipline
- Research Design & Rigor
- Loss type
- Unmigrated
- Stage
- Definition, Discovery
- Data
- Analytical / expert
- Effort
- Days
Steps
- Specify the hypothesised network (or nodes and density), the estimation method, a performance measure and target, a statistic and target, and a wide candidate range of sample sizes.
- Step 1: for several equally spaced sample sizes, run many replications of simulate, estimate and score, and compute the statistic at each size.
- Step 2: fit a monotone non-decreasing spline to the statistic and interpolate across the range.
- Step 3: bootstrap replications within each sample size, refit splines and form 95% bands; set the new range where the bands reach the target.
- Repeat until the range is within tolerance or an iteration limit is reached.
- Recommend the median of the final range, validate with extra replications at that size, and repeat for weaker and stronger edge assumptions.
Sources
- A General Monte Carlo Method for Sample Size Analysis in the Context of Network Models (Constantin et al., 2023)
Also known as
- powerly
- Simulation-based sample size analysis for psychological networks
- Monte Carlo power analysis for Gaussian graphical models
- Three-step Monte Carlo sample size algorithm
Categories
- Research Operations
- Research Design & Rigor