makeRandomDataT: makeRandomData

In benkeser/haltmle.sim: Simulation study for HAL-TMLE

Description

Simulate a random data generating distribution. See Details to see how it's done.

Usage

makeRandomDataT(n, maxD = 2, D = NULL, minObsA = 30,
  func.distW = c("uniformW", "normalW", "bernoulliW", "binomialW", "gammaW",
  "normalWCor", "bernoulliWCor", "uniformWCor", "gammaPointMassW",
  "binomialFracW", "normalPointMassW"), funcG0.uni = c("linUni", "polyUni",
  "sinUni", "jumpUni", "pLogisUni", "dNormUni", "qGammaUni", "dNormMixUni",
  "cubicSplineUni", "linSplineUni"), funcG0.biv = c("linBiv", "polyBiv",
  "sinBiv", "jumpBiv", "dNormAddBiv", "dNormMultBiv", "polyJumpBiv",
  "sinJumpBiv", "dNormMultJumpBiv", "linSplineBiv"), funcG0.tri = c("linTri",
  "polyTri", "sinTri", "jumpTri", "linJump2Tri", "linJump1Tri", "polyJump2Tri",
  "polyJump1Tri", "sinJump2Tri", "sinJump1Tri", "dNormAddTri", "dNormMultTri",
  "dNormMultJump2Tri", "dNormMultJump1Tri", "pLogisAddTri"),
  funcG0.quad = c("linQuad", "polyQuad", "sinQuad", "jumpQuad",
  "linJump2Quad", "linJump1Quad", "polyJump2Quad", "polyJump1Quad",
  "sinJump2Quad", "sinJump1Quad", "dNormAddQuad", "dNormMultQuad",
  "dNormMultJump2Quad", "dNormMultJump1Quad", "pLogisAddQuad"),
  funcQ0.uni = c("linUni", "polyUni", "sinUni", "jumpUni", "pLogisUni",
  "dNormUni", "qGammaUni", "dNormMixUni", "cubicSplineUni", "linSplineUni"),
  funcQ0.biv = c("linBiv", "polyBiv", "sinBiv", "jumpBiv", "dNormAddBiv",
  "dNormMultBiv", "polyJumpBiv", "sinJumpBiv", "dNormMultJumpBiv",
  "linSplineBiv"), funcQ0.tri = c("linTri", "polyTri", "sinTri", "jumpTri",
  "linJump2Tri", "linJump1Tri", "polyJump2Tri", "polyJump1Tri", "sinJump2Tri",
  "sinJump1Tri", "dNormAddTri", "dNormMultTri", "dNormMultJump2Tri",
  "dNormMultJump1Tri", "pLogisAddTri"), funcQ0.quad = c("linQuad", "polyQuad",
  "sinQuad", "jumpQuad", "linJump2Quad", "linJump1Quad", "polyJump2Quad",
  "polyJump1Quad", "sinJump2Quad", "sinJump1Quad", "dNormAddQuad",
  "dNormMultQuad", "dNormMultJump2Quad", "dNormMultJump1Quad", "pLogisAddQuad"),
  errY = c("normalErr", "uniformErr", "gammaErr", "normalErrAW",
  "uniformErrAW"), minG0 = 0.001, minR2 = 0.01, maxR2 = 0.99,
  pos = 0.005, skewing = c(-1, 1), ...)

Arguments

n	A numeric sample size
maxD	A numeric indicating maximum number of covariates
funcG0.uni	A vector of characters that are functions in the global environment that are used to generate main terms for logit of the propensity score.
funcG0.biv	Ditto above, but for bivariate functions.
funcG0.tri	Ditto above, but for trivariate functions.
funcQ0.uni	Ditto funcG0.uni, but for the outcome regression
funcQ0.biv	Ditto
funcQ0.tri	Ditto
errY	A vector of characters that are functions in the global environment that are used to generate error terms for the outcome regression.
minG0	The minimum value for the propensity score (default 0.01).
pos	(1-2*pos)100% of pscores fall withing minG0 and 1- minG0, default is .005
distW.uncor	A vector of characters that are functions in the global environment that generate covariates. These functions should have an associated parameter function with "Parm" added to the end of their name. See the format of the functions provided to understand their structure.
distW.cor	Same as above, but generates a correlated covariate
skewage	default is c(-1,1). random add a skewing constant to the logit(pscores)

Details

Draw random data generating distribution. The code simulates as follows:
Simulate W

• Choose D, a random number of covariates from Uniform(1,maxD)

• Randomly choose d1, the distribution for first covariate, from distW.uncor

• Call paste0(d1,".Parm") (which is assumed to exist in the global environment) to get parameters for d1

• Call d1 with parameters chosen in previous step to generate first covariate

• Randomly choose d2, the distribution for second covariate, from c(distW.uncor, distW.cor) (now including correlated distributions).

• Call paste0(d2,".Parm") to get parameters for d2

• Call d2 with parameters chosen in previous step and first covariate (to possibly induce correlation with previous covariate) to generate second covariate.

• Continue until maxD is reached.

Simulating A

• Draw MG1, a random number of main terms for propensity score, from Uniform(1,maxD)

• For i = 1,...,MG1, draw f a random function from funcG0.uni

• Call paste0(f,"Parm") (assumed to exist) to generate parameters for f

• Call f with x = W[,i] to generate i-th main term.

• If D > 1, draw MG2, a random number of bivariate interactions for propensity, from Uniform(1,MG1-1)

• For i = 1,...,MG2, draw f a random function from funcG0.biv

• Call paste0(f,"Parm") (assumed to exist) to generate parameters for f

• Draw j and k, two random columns of W

• Call f with x1 = W[,j], x2 = W[,k] to generate i-th bivariate interaction term.

• If D > 2, draw MG3, a random number of trivariate interactions for propensity, from Uniform(1,MG2-1)

• For i = 1,...,MG3, draw f a random function from funcG0.tri

• Call paste0(f,"Parm") (assumed to exist) to generate parameters for f

• Draw j,k,l, three random columns of W

• Call f with x1 = W[,j], x2 = W[,k], x3 = W[,l] to generate i-th trivariate interaction term.

• Sum together all main, bivariate, and trivariate terms to get the logit propensity score.

• Divide the logit propensity score by 1.01 until the minimum propensity score is bigger than minG0.

• Draw A from Bernoulli distribution with conditional probability equal to the adjusted propensity score.

Simulating Y

• Draw MQ1, a random number of main terms for outcome regression score, from Uniform(2,maxD)

• For i = 1,...,MQ1, draw f a random function from funcQ0.uni

• Call paste0(f,"Parm") (assumed to exist) to generate parameters for f

• Call f with x = AW[,i] to generate i-th main term, where AW is matrix with first column A, other columns W.

• If D > 1, draw MQ2, a random number of bivariate interactions for outcome regression, from Uniform(1,MQ1-1)

• For i = 1,...,MQ2, draw f a random function from funcQ0.biv

• Call paste0(f,"Parm") (assumed to exist) to generate parameters for f

• Draw j and k, two random columns of W

• Call f with x1 = W[,j], x2 = W[,k] to generate i-th bivariate interaction term.

• If D > 2, draw MQ3, a random number of trivariate interactions for propensity, from Uniform(1,MQ2-1)

• For i = 1,...,MQ3, draw f a random function from funcQ0.tri

• Call paste0(f,"Parm") (assumed to exist) to generate parameters for f

• Draw j,k,l, three random columns of W

• Call f with x1 = W[,j], x2 = W[,k], x3 = W[,l] to generate i-th trivariate interaction term.

• Sum together all main, bivariate, and trivariate terms to get the conditional mean outcome, Q0.

• Draw errf, arandom function from errY from which to draw errors

• Call paste0(errf,"Parm" (assumed to exist) to generate random parameters for error distribution.

• Call errf with randomly chosen parameters to draw e, an error for the outcome.

• Let Y = Q0 + e.

Value

An object of class "makeRandomData" with the following entries

W	A matrix of covariates
A	A vector of binary treatments
Y	A vector of continuously valued outcome
distW	A list containing relevant information needed to reproduce data sets
fnG0	A list of lists containing relevant information needed to reproduce data sets
fnQ0	A list of lists containing relevant information needed to reproduce data sets
distErrY	A list containing relevant information needed to reproduce data sets
divideLogitG0	A numeric of the scaling factor for the propensity

benkeser/haltmle.sim documentation built on May 12, 2019, noon