The RGremlinsConjoint package provides the tools and utilities to estimate a the model described in “Gremlin’s in the Data: Identifying the Information Content of Research Subjects” (https://doi.org/10.1177%2F0022243720965930) using conjoint analysis data such as that collected in Sawtooth Software’s Lighthouse or Discover Products. The packages also contains utility functions for formatting the input data and extracting the relevant results.
You can install the development version from GitHub with:
The package exposes basically one function You can use it like:
# Read in the data
truck_design_file <- system.file("extdata", "simTruckDesign.csv", package = "RGremlinsConjoint")
truck_data_file <- system.file("extdata", "simTruckData.csv", package = "RGremlinsConjoint")
truckDesign <- read.csv(truck_design_file)
truckData <- read.csv(truck_data_file)
outputSimData_burn <- estimateGremlinsModel(truckData,
R = 10,
keepEvery = 1,
num_lambda_segments = 2)
#> Finding Starting Values
#> Beginning MCMC Routine
#> Completing iteration : 1
#> Accept rate slopes: 0
#> Accept rate lambda: 0
#> Mu_adapt lambda: 50
#> Gamma_adapt lambda: 10
#> metstd lambda: 10
#> current lambda: 50
The first two parameters represent the responses and the the design file. The data file format follows the Sawtooth Software .csv file format with one row per respondent. The first column holds a respondent identified, the second column is an index to the block of version number that the respondent saw in the conjoint survey. The remaining columns hold the option index for each of the conjoint screens that the respondent saw. There should be one column for each screen and the responses should correspond to the row in the design file that the respondent chose. (Indexes starting at 1 are assumed.)
R specifies the number of iterations that
the MCMC algorithm should run. With this model we recommend using a
relatively large number of iterations as it can take a while for the
variance parameters to fully stabilize since they directly influence the
keepEvery controls the thinning of the
MCMC chain. For long chains it is recommend that you increase the
default as the MCMC chain can quickly grow in size.
The final parameter
num_lambda_segments determines how
many “gremlin” segments are in your model. We anticipate hat this will
be 2 for the vast majority of projects. This corresponds to one
information rich segment and one information poor segment. Using more
than two segments can make interpretation difficult.
The output of this function is a data structure containing the complete MCMC chain for each parameter. You will need to subset the chain after deciding where in the chain convergence occurs. We do not provide any additional functions for diagnosing or evaluating the MCMC chains.
estimateGremlinsModel takes a number of
optional arguments. See the documentation for full details. There are
three that are worth calling out specifically. If you need to continue
the chain it is possible to pass the function a list of
startingValues. This allows you to override the default
starting value calculation if you need to continue a previous chain.
The algorithm also uses an automatic step size tuning algorithm to optimize the Metropolis-Hasting proposal distribution for the slopes and the lambda parameters. You will need to monitor the acceptance rate for these parameters and adjust the settings if the acceptance rate it too high or too low. We recommend that you target an acceptance rate between 20%-80%. Lowering the tuning factor will increase the acceptance rate.