updates to JAGS models

- removing ExpPower hierarchical model. Hierarchical inference on (k,
tau) needs more thought #147
- introduces (epsilon_alpha, epsilon_beta) which will help with the
hierarchical / mixed / separate models #161

ddToolbox/models/nonparametric_models/separateNonParametric.jags

@@ -20,10 +20,16 @@
 
 model{
 
+groupALPHAmu <- 0           # TODO: NEEDS ATTENTION
+groupALPHAsigma <-0.01      # TODO: NEEDS ATTENTION
+
+epsilon_alpha <- 1+1
+epsilon_beta <- 1+100
+
 # priors
 for (p in participantIndexList){
-    epsilon[p]  ~ dbeta(1+1 , 1+100) T(,0.5)
-    alpha[p]    ~ dnorm(0, 1/0.1^2) T(0,) # <--------------- FIGURE OUT BEST PRIOR. 
+    epsilon[p]  ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
+    alpha[p]    ~ dnorm(groupALPHAmu, 1/(groupALPHAsigma^2)) T(0,)
 
     # PRIOR OVER DISCOUNT FRACTION
     # prior over A/B should be centered on 1 if we are expecting discounting and anti-discounting, and can be truncates

ddToolbox/models/parametric_models/bens_new_model/ModelMixedExpPower.m

@@ -18,8 +18,8 @@
 			% Generate initial values of the root nodes
 			nExperimentFiles = obj.data.getNExperimentFiles();
 			for chain = 1:nchains
-				%initialParams(chain).k = unifrnd(0, 0.5, [nExperimentFiles,1]);
-                %initialParams(chain).tau = unifrnd(0.01, 2, [nExperimentFiles,1]);
+% 				initialParams(chain).k = unifrnd(0.01, 0.5, [nExperimentFiles,1]);
+%                 initialParams(chain).tau = unifrnd(0.01, 2, [nExperimentFiles,1]);
 
 				initialParams(chain).groupW             = rand;
 				initialParams(chain).groupALPHAmu		= rand*100;

ddToolbox/models/parametric_models/bens_new_model/mixedExpPower.jags

@@ -28,16 +28,22 @@ for (t in 1:length(ID)) {
 
 
 # RESPONSE ERROR PARAMETERS ====================================================
+
 # comparison acuity (alpha) hyperprior
 groupALPHAmu        ~ dnorm(0,1/(100^2)) T(0,)
 groupALPHAsigma     ~ dunif(0,100) 
+
 # error rates (epsilon) hyperprior
 groupW          ~ dbeta(1.1, 10.9)  # mode for lapse rate
 groupKminus2    ~ dgamma(0.5,0.5) # concentration parameter ## UPDATED SINCE PAPER
 groupK          <- groupKminus2+2
+
+epsilon_alpha <- groupW*(groupK-2)+1
+epsilon_beta <- (1-groupW)*(groupK-2)+1
+
 for (p in 1:(nRealExperimentFiles+1)){ # +1 for unobserved participant
 	alpha[p]    ~ dnorm(groupALPHAmu, 1/(groupALPHAsigma^2)) T(0,)
-    epsilon[p]  ~ dbeta(groupW*(groupK-2)+1 , (1-groupW)*(groupK-2)+1 ) T(,0.5)
+    epsilon[p]  ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
 }
 
 

ddToolbox/models/parametric_models/bens_new_model/separateExpPower.jags

@@ -26,8 +26,11 @@ for (t in 1:length(ID)) {
 
 
 # RESPONSE ERROR PARAMETERS ====================================================
+epsilon_alpha <- 1.1
+epsilon_beta <- 10.9
+
 for (p in 1:(nRealExperimentFiles)){     # no +1 because no shrinkage hyperprior
-    epsilon[p]  ~ dbeta(1.1 , 10.9) T(,0.5)
+    epsilon[p]  ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
 	alpha[p]    ~ dexp(0.01)
 }
 

ddToolbox/models/parametric_models/exponential_models/hierarchicalExp1.jags

@@ -10,11 +10,12 @@ model{
 # DISCOUNT FUNCTION PARAMETERS =================================================
 # RANDOM (BY PARTICIPANT) FACTORS; HYPER-PRIORS = YES 
 
-groupKmu      ~ dnorm(0.01, 1/(0.5^2))
-groupKsigma   ~ dgamma(0.1, 0.1)
+K_MEAN      ~ dnorm(0.01, 1/(0.5^2))
+Ksigma      ~ dgamma(0.1, 0.1)
+K_PRECISION <- 1/((Ksigma+0.0001)^2)
 
 for (p in 1:(nRealExperimentFiles+1)){      # +1 for unobserved participant
-	k[p]        ~ dnorm(groupKmu, 1/((groupKsigma+0.0001)^2))
+	k[p] ~ dnorm(K_MEAN, K_PRECISION)
 }
 
 # MODEL-SPECIFIC: CALCULATION OF PRESENT SUBJECTIVE VALUES
@@ -25,16 +26,22 @@ for (t in 1:length(ID)) {
 
 
 # RESPONSE ERROR PARAMETERS ====================================================
+
 # comparison acuity (alpha) hyperprior
 groupALPHAmu        ~ dnorm(0,1/(100^2)) T(0,)
 groupALPHAsigma     ~ dunif(0,100) 
+
 # error rates (epsilon) hyperprior
 groupW          ~ dbeta(1.1, 10.9)  # mode for lapse rate
 groupKminus2    ~ dgamma(0.5,0.5) # concentration parameter ## UPDATED SINCE PAPER
 groupK          <- groupKminus2+2
+
+epsilon_alpha <- groupW*(groupK-2)+1
+epsilon_beta <- (1-groupW)*(groupK-2)+1
+
 for (p in 1:(nRealExperimentFiles+1)){ # +1 for unobserved participant
 	alpha[p]    ~ dnorm(groupALPHAmu, 1/(groupALPHAsigma^2)) T(0,)
-    epsilon[p]  ~ dbeta(groupW*(groupK-2)+1 , (1-groupW)*(groupK-2)+1 ) T(,0.5)
+    epsilon[p]  ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
 }
 
 
@@ -47,7 +54,7 @@ for (t in 1:length(ID)) {
 
 # response likelihood
 for (t in 1:length(ID)) {
-	R[t]  ~ dbern(P[t]) # likelihood of actual response
+	R[t]  ~ dbern(P[t])
 }
 
 # POSTERIOR PREDICTION

ddToolbox/models/parametric_models/exponential_models/mixedExp1.jags

@@ -25,16 +25,22 @@ for (t in 1:length(ID)) {
 
 
 # RESPONSE ERROR PARAMETERS ====================================================
+
 # comparison acuity (alpha) hyperprior
 groupALPHAmu        ~ dnorm(0,1/(100^2)) T(0,)
 groupALPHAsigma     ~ dunif(0,100) 
+
 # error rates (epsilon) hyperprior
 groupW          ~ dbeta(1.1, 10.9)  # mode for lapse rate
 groupKminus2    ~ dgamma(0.5,0.5) # concentration parameter ## UPDATED SINCE PAPER
 groupK          <- groupKminus2+2
+
+epsilon_alpha <- groupW*(groupK-2)+1
+epsilon_beta <- (1-groupW)*(groupK-2)+1
+
 for (p in 1:(nRealExperimentFiles+1)){ # +1 for unobserved participant
 	alpha[p]    ~ dnorm(groupALPHAmu, 1/(groupALPHAsigma^2)) T(0,)
-    epsilon[p]  ~ dbeta(groupW*(groupK-2)+1 , (1-groupW)*(groupK-2)+1 ) T(,0.5)
+    epsilon[p]  ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
 }
 
 
@@ -47,7 +53,7 @@ for (t in 1:length(ID)) {
 
 # response likelihood
 for (t in 1:length(ID)) {
-	R[t]  ~ dbern(P[t]) # likelihood of actual response
+	R[t]  ~ dbern(P[t])
 }
 
 # POSTERIOR PREDICTION

ddToolbox/models/parametric_models/exponential_models/separateExp1.jags

@@ -25,8 +25,11 @@ for (t in 1:length(ID)) {
 
 
 # RESPONSE ERROR PARAMETERS ====================================================
+epsilon_alpha <- 1.1
+epsilon_beta <- 10.9
+
 for (p in 1:(nRealExperimentFiles)){     # no +1 because no shrinkage hyperprior
-    epsilon[p]  ~ dbeta(1.1 , 10.9) T(,0.5)
+    epsilon[p]  ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
 	alpha[p]    ~ dexp(0.01)
 }
 
@@ -48,7 +51,7 @@ for (t in 1:length(ID)) {
 
 # response likelihood
 for (t in 1:length(ID)) {
-	R[t]  ~ dbern(P[t]) # likelihood of actual response
+	R[t]  ~ dbern(P[t])
 }
 
 # POSTERIOR PREDICTION

ddToolbox/models/parametric_models/hyperbolic_magnitude_effect_models/hierarchicalME.jags

@@ -53,8 +53,12 @@ groupALPHAsigma     ~ dunif(0,1000)
 groupW          ~ dbeta(1.1, 10.9)  # mode for lapse rate
 groupKminus2    ~ dgamma(0.01,0.01) # concentration parameter
 groupK          <- groupKminus2+2
+
+epsilon_alpha <- groupW*(groupK-2)+1
+epsilon_beta <- (1-groupW)*(groupK-2)+1
+
 for (p in 1:(nRealExperimentFiles+1)){ # +1 for unobserved participant
-	epsilon[p]  ~ dbeta(groupW*(groupK-2)+1 , (1-groupW)*(groupK-2)+1 ) T(,0.5)
+	epsilon[p] ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
 	alpha[p]    ~ dnorm(groupALPHAmu, 1/(groupALPHAsigma^2)) T(0,)
 }
 

ddToolbox/models/parametric_models/hyperbolic_magnitude_effect_models/hierarchicalMEupdated.jags

@@ -54,8 +54,11 @@ groupW          ~ dbeta(1.1, 10.9)  # mode for lapse rate
 groupKminus2    ~ dgamma(0.5,0.5) # concentration parameter ## UPDATED SINCE PAPER
 groupK          <- groupKminus2+2
 
+epsilon_alpha <- groupW*(groupK-2)+1
+epsilon_beta <- (1-groupW)*(groupK-2)+1
+
 for (p in 1:(nRealExperimentFiles+1)){ # +1 for unobserved participant
-	epsilon[p]  ~ dbeta(groupW*(groupK-2)+1 , (1-groupW)*(groupK-2)+1 ) T(,0.5)
+	epsilon[p]  ~ dbeta(epsilon_alpha , epsilon_beta ) T(,0.5)
 	alpha[p]    ~ dnorm(groupALPHAmu, 1/(groupALPHAsigma^2)) T(0,)
 }