PredPercentileCNN: Experimental function for getting percentiles using an...

In CambridgeAssessmentResearch/KernEqWPS: Kernel Equating Without Pre-Smoothing

Experimental function for getting percentiles using an approximation to a trained convolutional neural network

Description

This function takes a data frame of test form and anchor scores and estimates the values of the percentiles (1st-99th) for given change in the distribution of anchor scores. See Benton (2017) for more details.

Usage

PredPercentileCNN(
  dx,
  anchortargettable = NA,
  maxx = NA,
  maxa = NA,
  WeightsList = ApproxCNNWeights
)

Arguments

dx	Data frame with variables "x" and "a" representing scores for individual candidates on form X and on the anchor test.
anchortargettable	Table giving distribution of anchor test scores in the target population.
maxx	Maximum score available on form X (calculated from the data by default).
maxa	Maximum score available on anchor test (calculated from the data by default).
WeightsList	A list of neural network parameters used in calculations. Changing this from the default value is not recommended.

References

Benton, T. (2017). Can AI learn to equate?, presented at the International Meeting of the Psychometric Society, Zurich, 2017. Cambridge, UK: Cambridge Assessment.

Examples

#example (compare real and estimated percentiles within a fixed population)
n1=1000
t1=rnorm(n1,0.5,1)
dx=data.frame(x=round(pmin(100,pmax(0,50+20*(0.9*t1+rnorm(n1,0,sqrt(1-0.9^2))))))
	,a=round(pmin(10,pmax(0,5+2*(0.7*t1+rnorm(n1,0,sqrt(1-0.7^2)))))))

percCNN=PredPercentileCNN(dx,table(dx$a),maxx=100)
usualperc=as.vector(quantile(dx$x,seq(0.01,0.99,0.01)))
plot(1:99,usualperc,type='l',xlab="Percentile",ylab="Value")
lines(1:99,percCNN,lty=2)

CambridgeAssessmentResearch/KernEqWPS documentation built on Feb. 23, 2024, 9:34 p.m.