BANOVA: Hierarchical Bayesian ANOVA Models

Description

This package includes several hierarchical Bayes Analysis of Variance models. These models are suited for the analysis of experimental designs in which both within- and between- subjects factors are manipulated, and account for a wide variety of distributions of the dependent variable. Floodlight analysis and mediation analysis basaed on these models are also provided. The package uses 'Stan' and 'JAGS' as the computational platform.

Details

Model:
E(y_i) = g^{-1}(\eta_i)
where \eta_i = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^p\beta_{j,s_i}^p, s_i is the subject id of data response i. Missing values (NAs) of y_i are allowed. The within-subjects factors and their interactions are indexed by p (p = 1,2,.,P). Each index p represents a batch of J_p coefficients: \beta_{j,s}^p, j = 1,.,J_p;s = 1,.,S indexes subjects. Note that if the subject-level covariate is continuous, J_p=1, so that ANCOVA models are also accommodated (relaxing their "constant slope" assumption).

The population-level model allows for heterogeneity among subjects, because the subject-level coefficients \beta_{j,s}^p are assumed to follow a multivariate normal distribution.The between-subjects factors and their interactions are indexed by q,(q = 1,2,.,Q), q = 0 denotes the constant term. The population-level ANOVA can be written as:
\beta_{j,s}^p = \sum_{q = 0}^Q \theta_{j,k_s^q}^{pq} + \delta_{j,s}^p

The population-level ANCOVA model can be expressed as a linear model with a design matrix Z that contains all between-subjects factors and their interactions and a constant term:
\beta_{j,s}^p = \sum_{k = 1}^Q Z_{s,k}\theta_{j,k}^{p} + \delta_{j,s}^p
where Z_{s,k} is an element of Z, a S \times Q matrix of covariates. \theta_{j,k}^p is a hyperparameter which captures the effects of between-subjects factor q on the parameter \beta_{j,s}^p of within-subjects factor p. The error \delta_{j,s}^p is assumed to be normal: \delta_{j,s}^p ~ N(0,\sigma_p^{-2} ). Proper, but diffuse priors are assumed: \theta_{j,k}^p ~ N(0,\gamma), and \sigma_p^{-2} ~ Gamma(a,b), where \gamma,a,b are hyper-parameters. The default setting is \gamma = 10^{-4}, a = 1, b = 1.

Note that missing values of independent variables are currently not allowed in the package.

Author(s)

Chen Dong; Michel Wedel

Maintainer: Chen Dong <cdong@math.umd.edu>

References

Dong, C. and Wedel, M. (2017) BANOVA: An R Package for Hierarchical Bayesian ANOVA, Journal of Statistical Software, Vol. 81, No.9, pp. 1-46.

McCullagh, P., Nelder, JA. (1989) Generalized linear models, New York, NY: Chapman and Hall.

Gelman, A. (2005) Analysis of variance-why it is more important than ever, Ann. Statist., Vol. 33, No. 1, pp. 1-53.

Rossi, P., Allenby,G., McCulloch, R. (2005) Bayesian Statistics and Marketing, John Wiley and Sons.

Gill, J. (2007) Bayesian Methods for the Social and Behavioral Sciences, Chapman and Hall, Second Edition.

Gelman, A., Carlin, J., Stern, H. and Dunson, D. (2013) Bayesian Data Analysis, London: Chapman and Hall.

Wedel, M. and Dong, C. (2016) BANOVA: Bayesian Analysis of Variance for Consumer Research. Submitted.

Estimation of BANOVA with a Bernoulli dependent variable

Description

BANOVA.Bernoulli implements a Bayesian ANOVA for binary dependent variable, using a logit link and a normal heterogeneity distribution.

Usage

BANOVA.Bernoulli(l1_formula = "NA", l2_formula = "NA", data,
  id, l2_hyper = c(1, 1, 0.0001), burnin = 5000, sample = 2000, thin = 10, 
  adapt = 0, conv_speedup = F, jags = runjags.getOption('jagspath'))

## S3 method for class 'BANOVA.Bernoulli'
summary(object, ...)
## S3 method for class 'BANOVA.Bernoulli'
predict(object, newdata = NULL,...)
## S3 method for class 'BANOVA.Bernoulli'
print(x, ...)

Arguments

Details

Level 1 model:
y_i ~ Binomial(1,p_i), p_i = logit^{-1}(\eta_i)
where \eta_i = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^p\beta_{j,s_i}^p, s_i is the subject id of data record i. see BANOVA-package

Value

BANOVA.Bernoulli returns an object of class "BANOVA.Bernoulli". The returned object is a list containing:

Examples

data(bernlogtime)
# model with the dependent variable : response
res <- BANOVA.Bernoulli(response~typical, ~blur + color, bernlogtime,
bernlogtime$subject, burnin = 5000, sample = 2000, thin = 10)
summary(res)

Estimation of BANOVA with a Binomial dependent variable

Description

BANOVA.Binomial implements a Hierarchical Bayesian ANOVA for a binomial response variable using a logit link and a normal heterogeneity distribution.

Usage

BANOVA.Binomial(l1_formula = "NA", l2_formula = "NA", data,
  id, num_trials, l2_hyper = c(1, 1, 0.0001), burnin = 5000, sample = 2000, 
  thin = 10, adapt = 0, conv_speedup = F, jags = runjags.getOption('jagspath'))
## S3 method for class 'BANOVA.Binomial'
summary(object, ...)
## S3 method for class 'BANOVA.Binomial'
predict(object, newdata = NULL,...)
## S3 method for class 'BANOVA.Binomial'
print(x, ...)

Arguments

Details

Level 1 model:
y_i ~ Binomial(ntrials,p_i), p_i = logit^{-1}(\eta_i)
where ntrials is the binomial total for each record i, \eta_i = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^p\beta_{j,s_i}^p, s_i is the subject id of response i. see BANOVA-package

Value

BANOVA.Binomial returns an object of class "BANOVA.Bin". The returned object is a list containing:

Examples

data(colorad)

# mean center Blur for effect coding
colorad$blur <- colorad$blur - mean(colorad$blur)
res <- BANOVA.Binomial(y~typic, ~color*blur, colorad, colorad$id, as.integer(16), 
burnin = 5000, sample = 2000, thin = 10)
summary(res)
# or use BANOVA.run
require(rstan)
res0 <- BANOVA.run(y~typic, ~color*blurfac, data = colorad, model_name = 'Binomial',
id = 'id', num_trials = as.integer(16), iter = 100, thin = 1, chains = 2)
summary(res0)
table.predictions(res0)
# only in-model variables(except numeric variables) will be used
predict(res0, c(1, 0, 8, 2, 1, 0.03400759))

Estimation of BANOVA with a Multinomial dependent variable

Description

BANOVA.Multinomial implements a Hierarchical Bayesian ANOVA for multinomial response variable using a logit link and a normal heterogeneity distribution.

Usage

BANOVA.Multinomial(l1_formula = "NA", l2_formula = "NA",
  dataX, dataZ, y, id, l2_hyper = c(1, 1, 0.0001), burnin = 5000, sample = 2000, 
  thin = 10, adapt = 0, conv_speedup = F, jags = runjags.getOption('jagspath'))
## S3 method for class 'BANOVA.Multinomial'
summary(object, ...)
## S3 method for class 'BANOVA.Multinomial'
predict(object, Xsamples = NULL, Zsamples = NULL,...)
## S3 method for class 'BANOVA.Multinomial'
print(x, ...)

Arguments

Details

Level 1 model:
P(y_i = \ell) = \frac{exp(\eta_{i\ell})}{\sum_{\ell=1}^L exp(\eta_{i\ell})}
where \eta_{i\ell} = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^{k,p} \beta_{j,s_i}^p, s_i is the subject id of response i, see BANOVA-package. X_{i,j}^{k,p} is the design matrix corresponding to each class \ell(\ell=1,.,L) of y_i. The first level of the response is the base level, thus the intercept corresponding to this level will not be included.

Value

BANOVA.Multinomial returns an object of class "BANOVA.Multinomial". The returned object is a list containing:

Examples

# see 'choicedata'
data(choicedata)
# generate dataX(convert the within-subject variables to a list)
dataX <- list()
for (i in 1:nrow(choicedata)){
  logP <- as.numeric(log(choicedata[i,3:8]))
  # all numeric variables must be mean centered
  dataX[[i]] <- as.data.frame(logP) - mean(logP)
}
dataZ <- choicedata[,9:13]

res <- BANOVA.Multinomial(~ logP, ~ college, dataX, dataZ, 
  choicedata$choice, choicedata$hhid, burnin = 100, sample = 100, thin = 10)
# or use BANOVA.run based on 'Stan'
require(rstan)
res <- BANOVA.run(~ logP, ~ college, dataX = dataX, dataZ = dataZ, 
                    model_name = 'Multinomial', y_value = choicedata$choice, 
                    id = choicedata$hhid, iter = 100, thin = 1, chains = 2)

Estimation of BANOVA with a normally distributed dependent variable

Description

BANOVA.Normal implements a Hierarchical Bayesian ANOVA for linear models with normal response and a normal heterogeneity distribution.

Usage

BANOVA.Normal(l1_formula = "NA", l2_formula = "NA", data,
  id, l1_hyper = c(1, 1), l2_hyper = c(1, 1, 0.0001), burnin = 5000, 
  sample = 2000, thin = 10, adapt = 0, conv_speedup = F, 
  jags = runjags.getOption('jagspath'))
## S3 method for class 'BANOVA.Normal'
summary(object, ...)
## S3 method for class 'BANOVA.Normal'
predict(object, newdata = NULL,...)
## S3 method for class 'BANOVA.Normal'
print(x, ...)

Arguments

Details

Level 1 model:
y_i ~ Normal(\eta_i,\sigma^{-2})
where \eta_i = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^p\beta_{j,s_i}^p, s_i is the subject id of response i, \sigma^{-2} ~ Gamma(\alpha,\beta). see BANOVA-package

Value

BANOVA.Normal returns an object of class "BANOVA.Normal". The returned object is a list containing:

Examples

# Use the ipadstudy data set
data(ipadstudy)
# mean center covariates
ipadstudy$age <- ipadstudy$age - mean(ipadstudy$age) 
ipadstudy$owner <- ipadstudy$owner - mean(ipadstudy$owner)
ipadstudy$gender <- ipadstudy$gender - mean(ipadstudy$gender)

# or use BANOVA.run based on 'Stan'
require(rstan)
res <- BANOVA.run(attitude~owner + age + gender + selfbrand*conspic, 
data = ipadstudy, model_name = 'Normal', id = 'id', 
iter = 100, thin = 1, chains = 2)

Estimation of BANOVA with Poisson dependent variables

Description

BANOVA.Poisson implements a Hierarchical Bayesian ANOVA for models with a count-data response variable and normal heterogeneity distribution.

Usage

BANOVA.Poisson(l1_formula = "NA", l2_formula = "NA",
  data, id, l2_hyper = c(1, 1, 0.0001), burnin = 5000, sample = 2000, thin = 10, 
  adapt = 0, conv_speedup = F, jags = runjags.getOption('jagspath'))
## S3 method for class 'BANOVA.Poisson'
summary(object, ...)
## S3 method for class 'BANOVA.Poisson'
predict(object, newdata = NULL,...)
## S3 method for class 'BANOVA.Poisson'
print(x, ...)

Arguments

Details

Level 1 model:
y_i ~ Poisson(\lambda_i), \lambda_i = exp(\eta_i + \epsilon_i)
where \eta_i = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^p\beta_{j,s_i}^p, s_i is the subject id of response i, see BANOVA-package. \epsilon_i is a dispersion term.

Value

BANOVA.Poisson returns an object of class "BANOVA.Poisson". The returned object is a list containing:

Examples

# use the bpndata dataset
data(bpndata)
# within-subjects model using the dependent variable : PIC_FIX 
res1 <- BANOVA.Poisson(PIC_FIX ~ AD_ID + PIC_SIZE+ PAGE_NUM 
+ PAGE_POS, ~1, bpndata, bpndata$RESPONDENT_ID, burnin = 500, 
sample = 200, thin = 5)
summary(res1)

# use the goalstudy dataset
data(goalstudy)
goalstudy$bid <- as.integer(goalstudy$bid + 0.5)
res2<-BANOVA.Poisson(bid~1, ~progress*prodvar,  goalstudy, goalstudy$id, 
burnin = 5000, sample = 2000, thin = 10)
summary(res2)

# or use the BANOVA.run based on 'Stan'
require(rstan)
res3 <- BANOVA.run(bid~progress*prodvar, data = goalstudy, 
model_name = 'Poisson', id = 'id', iter = 100, thin = 1, chains = 2)

Estimation of BANOVA with T-distributin of the dependent variable

Description

BANOVA.T implements a Hierarchical Bayesian ANOVA for linear models with T-distributed response.

Usage

BANOVA.T(l1_formula = "NA", l2_formula = "NA", data, id, l1_hyper = c(1, 1, 1), 
l2_hyper = c(1, 1, 0.0001), burnin = 5000, sample = 2000, thin = 10, 
adapt = 0, conv_speedup = F, jags = runjags.getOption('jagspath'))
## S3 method for class 'BANOVA.T'
summary(object, ...)
## S3 method for class 'BANOVA.T'
predict(object, newdata = NULL,...)
## S3 method for class 'BANOVA.T'
print(x, ...)

Arguments

Details

Level 1 model:
y_i ~ t(\nu, \eta_i,\sigma^{-2})
where \eta_i = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^p\beta_{j,s_i}^p, s_i is the subject id of response i, see BANOVA-package. The hyper parameters: \nu is the degree of freedom, \nu ~ Piosson(\lambda) and \sigma is the scale parameter, \sigma^{-2} ~ Gamma(\alpha, \beta).

Value

BANOVA.T returns an object of class "BANOVA.T". The returned object is a list containing:

Examples

# Use the ipadstudy data set
data(ipadstudy)
res <- BANOVA.T(attitude~1, ~owner + age + gender + selfbrand*conspic, ipadstudy,
ipadstudy$id, burnin = 5000, sample = 2000, thin = 10)
summary(res)

# or use BANOVA.run based on 'Stan'
require(rstan)
res19 <- BANOVA.run(attitude~owner + age + gender + selfbrand*conspic, 
data = ipadstudy, model_name = 'T', id = 'id', iter = 100, 
thin = 1, chains = 2)

Build BANOVA models

Description

BANOVA.build builds(compiles) BANOVA models.

Usage

BANOVA.build(BANOVA_model)

Arguments

Value

BANOVA.build returns an object of class "BANOVA.build". The returned object is a list containing:

Examples

model <- BANOVA.model('Poisson', single_level = FALSE) 
Poisson_model <- BANOVA.build(model)
# visualize the model
cat(model$model_code)
# modify the model code and rebuild
# be careful to change any parameters
model$model_code <-"
data {
  int<lower=0> N;
  int<lower=0> J;
  int<lower=0> M;
  int<lower=0> K;
  matrix[N, J] X;
  matrix[M, K] Z;
  int<lower=0> id[N];
  int y[N];
}

parameters {
  matrix[J, M] beta1;
  matrix[K, J] beta2; 
  vector<lower=0>[J] tau_beta1Sq;
} 

model {
  vector[N] y_hat;
  matrix[M, J] mu_beta1;
  vector[J] tau_beta1;
  tau_beta1 = sqrt(tau_beta1Sq);
  for (i in 1:N){
    y_hat[i] = X[i,]*beta1[,id[i]];
  }
  y ~ poisson_log(y_hat);
  mu_beta1 = Z*beta2;
  for (i in 1:J){
    beta1[i,] ~ normal(mu_beta1[,i], tau_beta1[i]);
  }
  tau_beta1Sq ~ inv_gamma(1, 1);
  for (i in 1:J){
    beta2[,i] ~ normal(0, 10);
  }
}
"
Poisson_model_new <- BANOVA.build(model)

Floodlight analysis based on BANOVA models

Description

BANOVA.floodlight conducts floodlight analysis based on various BANOVA models.

Usage

BANOVA.floodlight(sol, var_numeric, var_factor, flood_values = list())
## S3 method for class 'BANOVA.floodlight'
print(x, ...)

Arguments

Details

A floodlight analysis (Spiller et al. 2013; Johnson and Neyman 1936) based on BANOVA models is conducted, which identifies regions of the numeric variable for which differences between the levels of the factor are significant. The endpoints of the 95% credible interval of the numeric variable provide the Johnson-Neyman points; for values outside of that interval there is 'strong' evidence that there is a difference between the levels of the factor.

Value

BANOVA.floodlight returns an object of class "BANOVA.floodlight". The returned object is a list containing:

References

Spiller, S., Fitzsimons, G., Lynch Jr., J. and McClelland, G. (2013) Spotlights, Floodlights, and the Magic Number Zero: Simple Effects Tests in Moderated Regression. Journal of Marketing Research, Vol. L, pp. 277-288.

Wedel, M. and Dong, C. (2016) BANOVA: Bayesian Analysis of Variance for Consumer Research. Submitted.

Examples

data(condstudy_sub)

library(rstan)
# use BANOVA.run
model <- BANOVA.model('Normal')
stanmodel <- BANOVA.build(model)
res <- BANOVA.run(att~cond+pict, ~type, fit = stanmodel, data = condstudy_sub, 
                  id = 'id', iter = 500, thin = 1, chains = 2)
BANOVA.floodlight(res, var_factor = 'type', var_numeric = 'pict')

Mediation analysis based on BANOVA models

Description

BANOVA.mediation conducts mediation and moderated mediation analysis based on various BANOVA models with a single mediator and a casual variable.

Usage

BANOVA.mediation(sol_1, sol_2, xvar, mediator, individual = F,
return_posterior_samples = F,  multi_samples_beta1_raw_m = NULL)

Arguments

Details

This function conducts a correlational mediation and moderated mediation analysis (Baron and Kenny 1986; Zao, Lynch and Chen 2010; Zhang, Wedel and Pieters 2008) based on BANOVA models. Based on the samples from posterior distributions, the function calculates the direct effect and indirect effect for which posterior means and 95% credible intervals are reported. The effect size of the indirect effect is computed as a generalized partial eta-squared. For details about this metric, see the publication of Wedel and Dong (2019).

When the algorithm is calculating the effects of a given causal variable it first identifies all moderators which are interacting with the investigated variable. Based on each interaction, moderated coefficients are computed and summarized in a table. If the causal variable is a part of an interaction term with three or more variables, separate results are computed for each of the moderators and all of their possible combinations. This results in multiple tables with the effects. If a continuous variable is involved in the interaction the effects are evaluated at its mean value, which is zero by default. This is equivalent to omitting the continuous variable from the interaction.

The function combines the effects of the mediator on the dependent variable with the effect of the causal variable on the mediator in a multiplicative manner to obtain the indirect effect of the treatment. If multiple tables with moderated effects of the mediator or the causal variable on mediator are obtained in the previous steps of the analysis, the indirect effects are computed for each combination of these table.

For models with a Normal outcome variable, it is possible to interpret the effects as causal by explicitly encoding the causal variable with dummy coding and including an interaction between the causal and mediating variables in the model. For further details, see the publication of MacKinnon et al. (2020).

Value

BANOVA.mediation returns an object of class "BANOVA.mediation". The returned object is a list containing:

References

Baron, R.M., and Kenny, D.A. (1986) Moderator Mediator Variables Distinction in Social Psychological Research: Conceptual, Strategic, and Statistical Considerations, Journal of Personality and Social Psychology, Vol. 51, No. 6, pp. 1173-82.

Zhang, J., Wedel, M. and Pieters, R.G.M. (2009) Sales Effects of Attention to Feature Advertisements: A Bayesian Mediation Analysis, Journal of Marketing Research, Vol.46, No.5, pp. 669-681.

Ying, Y. and MacKinnon, D.P. (2009) Bayesian Mediation Analysis, Psychological Methods, Vol. 14, No.4, pp. 301-322.

Zhao, X., John G.L., and Chen, Q. (2010) Reconsidering Baron and Kenny: Myths and Truths About Mediation Analysis, Journal of Consumer Research, Vol.37, No.2, pp. 197-206.

Wedel, M., and Dong, C. (2019) BANOVA: Bayesian Analysis of Variance for Consumer Research, Journal of Consumer Psychology, Vol. 30, No. 1, pp. 3-23.

MacKinnon, D.P., Valente, M.J., and Gonzalez, O. (2020) The correspondence between causal and traditional mediation analysis: the link is the mediator by treatment interaction, Prevention Science, Vol. 21, No. 2, pp. 147-157.

Examples

data(condstudy_sub)

# use BANOVA.run based on 'Stan'
model <- BANOVA.model('Normal')
banova_model <- BANOVA.build(model)
res_1 <- BANOVA.run(att~cond+pict, ~type, fit = banova_model, data = condstudy_sub, 
                  id = 'id', iter = 500, thin = 1, chains = 2)
res_2 <- BANOVA.run(pict~cond, ~type, fit = banova_model, data = condstudy_sub, 
                  id = 'id', iter = 500, thin = 1, chains = 2)
# (moderated) mediation
sol <- BANOVA.mediation(res_1, res_2, xvar='cond', mediator='pict')
print(sol)
print(sol$dir_effects)

Extract BANOVA models

Description

BANOVA.model extracts BANOVA models from the package.

Usage

BANOVA.model(model_name, single_level = F)

Arguments

Details

The function loads a pre-specified 'Stan' model for the analysis in BANOVA.

'Normal' model: A model suitable for a continuous dependent variable, which follows a Normal distribution.

'T' model: A model suitable for a continuous dependent variable, which might be prone to 'outliers' or fatter tails than the Normal.

'Bernoulli' model: A model suitable for a binary dependent variable, which can take values 0 and 1.

'Binomial' model: A model suitable for a dependent variable, which represents a number of successes in a sequence of B independent Bernoulli experiments.

'Poisson' model: A model suitable for a dependent variable, which represents count data. A Poisson distributed dependent variable can take values 0, 1, 2 ....

'ordMultinomial' model: A model suitable for an ordered categorical (ordinal) dependent variable, which follows an ordered Multinomial distribution. This dependent variable can take values from 1 to K, where possible alternatives are ordered according to some principal.

'Multinomial' model: A model suitable for a categorical (nominal) dependent variable, which follows a Multinomial distribution. This dependent variable can take values from 1 to K, where possible alternatives are unordered.

'multiNormal' model: A model suitable for a Multivariate Normal dependent variable, which represents L possibly correlated Normal dependent variables with shared predictors. The analysis corresponds to the seemingly unrelated regressions (SUR) technique.

'truncNormal' model: A model suitable a dependent variable, which values can only be observed if they lie within a certain range. The variable can be bounded from below, above, or from two sides.

Value

BANOVA.model returns an object of class "BANOVA.model". The returned object is a list containing:

Examples

model <- BANOVA.model('Poisson', single_level = FALSE) 
cat(model$model_code)

Mediation analysis with multiple possibly correlated mediators

Description

BANOVA.multi.mediation is a function for analysis of multiple possibly correlated mediators. These mediators are assumed to have no causal influence on each other. Both single-level and multi-level models can be analyzed.

Usage

BANOVA.multi.mediation(sol_1, sol_2, xvar, mediators, individual = FALSE)

Arguments

Details

The function extends BANOVA.mediation to the case with multiple possibly correlated mediators. For details about mediation analysis performed in BANOVA see the help page for the BANOVA.mediation.

BANOVA.multi.mediation estimates and tests specific indirect effects of the causal variable conveyed through each mediator. Furthermore, the total indirect effect of the causal variables are computed as a sum of the specific indirect effects.

The function prints multiple tables with mediated effects. Tables with direct effects of the causal variable and mediators on the outcome variable, as well as direct effects of the causal variable on the mediators include a posterior mean and 95% credible intervals of the effects. Next, the function displays on the console tables with specific indirect effects and effect sizes of the mediators, followed by the TIE of the causal variable. These tables include the mean, 95% credible intervals, and two-sided Bayesian p-values.

Value

Returns an object of class "BANOVA.multi.mediation". The returned object is a list containing:

Author(s)

Anna Kopyakova

Examples

# Use the colorad data set
data(colorad)
# Add a second mediator to the data set
colorad$blur_squared <- (colorad$blur)^2
# Prepare mediators to be analyzed in the Multivariate Normal model
mediators <- cbind(colorad$blur, colorad$blur_squared)
colnames(mediators) <- c("blur", "blur_squared")
colorad$mediators <- mediators

# Build and analyze the model for the outcome variable
model <- BANOVA.model('Binomial')
banova_binom_model <- BANOVA.build(model)
res_1 <- BANOVA.run(y ~ typic, ~ color + blur + blur_squared, fit = banova_binom_model,
                    data = colorad, id = 'id', num_trials = as.integer(16), 
                    iter = 2000, thin = 1, chains = 2)
# Build and analyze the model for the mediators
model <- BANOVA.model('multiNormal')
banova_multi_norm_model <- BANOVA.build(model)
res_2 <- BANOVA.run(mediators ~ typic, ~ color, fit = banova_multi_norm_model,
                    data = colorad, id = 'id', iter = 2000, thin = 1, chains = 2)
                    
# Calculate (moderated) effects of "typic" mediated by "blur" and "blur_squared"
results <- BANOVA.multi.mediation(res_1, res_2, xvar='typic', mediators=c("blur", "blur_squared"))

Estimation of BANOVA with a ordered Multinomial response variable

Description

BANOVA.ordMultinomial implements a Hierarchical Bayesian ANOVA for ordered multinomial responses, with a normal heterogeneity distribution.

Usage

BANOVA.ordMultinomial(l1_formula = "NA",
  l2_formula = "NA", data, id, l1_hyper = c(0.0001, 100), 
  l2_hyper = c(1, 1, 0.0001, 100), burnin = 5000, 
  sample = 2000, thin = 10, adapt = 0, conv_speedup = F, 
  jags = runjags.getOption('jagspath'))
## S3 method for class 'BANOVA.ordMultinomial'
summary(object, ...)
## S3 method for class 'BANOVA.ordMultinomial'
predict(object, newdata = NULL,...)
## S3 method for class 'BANOVA.ordMultinomial'
print(x, ...)

Arguments

Details

Level 1 model:
y_i = 1, if l_i < 0
y_i = 2, if 0 < l_i < c_2
...
y_i = \ell, if c_{\ell-1} < l_i < \infty
l_i = \eta_i + \epsilon_i where \epsilon_i ~ logistic (0,1), c_\ell, (\ell = 2,...L-1) are cut points, c_\ell ~ N(0, \bar{\sigma}_\ell^2), and \bar{\sigma}_\ell^2 ~ Uniform(0, d), with d a hyper-parameter.
\eta_i = \sum_{p = 0}^{P}\sum_{j=1}^{J_p}X_{i,j}^p\beta_{j,s_i}^p, s_i is the subject id of response i. see BANOVA-package

Value

BANOVA.ordMultinomial returns an object of class "BANOVA.ordMultinomial". The returned object is a list containing:

Examples

data(goalstudy)

res <- BANOVA.ordMultinomial (perceivedsim~1, ~progress*prodvar,  goalstudy,
goalstudy$id, burnin = 1000, sample = 1000, thin = 2)
summary(res)
# or use BANOVA.run based on 'Stan'
require(rstan)
res <- BANOVA.run(perceivedsim~progress*prodvar, data = goalstudy, 
model_name = 'ordMultinomial', id = 'id', iter = 100, thin = 1, chains = 2)

Estimation of BANOVA models

Description

BANOVA.run implements Hierarchical Bayesian ANOVA models using 'Stan'

Usage

BANOVA.run(l1_formula = "NA", l2_formula = "NA", fit = NULL, model_name = 'NA', 
dataX = NULL, dataZ = NULL, data = NULL, y_value = NULL, id, iter = 2000, 
num_trials = 1, contrast = NULL, y_lowerBound = -Inf, y_upperBound = Inf, ...)
## S3 method for class 'BANOVA'
summary(object, ...)
## S3 method for class 'BANOVA'
predict(object, newdata = NULL, Xsamples = NULL, Zsamples = 
                        NULL, ...)
## S3 method for class 'BANOVA'
print(x, ...)

Arguments

Value

BANOVA.run returns an object of class "BANOVA". The returned object is a list containing:

Examples

# Analysis of a single-level Normal dependent variable
# Use the ipadstudy data set
data(ipadstudy)
library(rstan)
# build the BANOVA model first so that it can be reused
model <- BANOVA.model('Normal', single_level = TRUE)
banova_model <- BANOVA.build(model)
res_1 <- BANOVA.run(attitude~owner + age + gender + selfbrand*conspic, 
fit = banova_model, data = ipadstudy, id = 'id', iter = 2000, chains = 2)
summary(res_1)
# or call the function directly without specifying the fit argument
# but it needs compilation
res_1 <- BANOVA.run(attitude~owner + age + gender + selfbrand*conspic,
model_name = 'Normal', data = ipadstudy, id = 'id', iter = 2000, chains = 2)

# Hierarchical analysis of multiple dependent variables (Multivariate Normal distribution)
# Use the colorad data set
data(colorad)
# Prepare dependent variables to be analyzed
colorad$blur_squared <- (colorad$blur)^2
dv <- cbind(colorad$blur, colorad$blur_squared)
colnames(dv) <- c("blur", "blur_squared")
colorad$dv <- dv
# Build and analyze the model for the dependent variables
model <- BANOVA.model('multiNormal')
banova_multi_norm_model <- BANOVA.build(model)
res_2 <- BANOVA.run(dv ~ typic, ~ color, fit = banova_multi_norm_model,
                    data = colorad, id = 'id', iter = 2000, thin = 1, chains = 2)

Simple effects calculation

Description

BANOVA.simple is a function for probing interaction effects in models where both moderator and explanatory variables are factors with an arbitrary number of levels. The function estimates and tests simple or partial effects, also known as simple main or conditional effects. Both single-level and multi-level models with any of the distributions accommodated in the package can be analyzed.

Usage

BANOVA.simple(BANOVA_output, base = NULL, quantiles = c(0.025, 0.975), 
dep_var_name = NULL, return_posterior_samples = FALSE)

Arguments

Details

The function identifies all factors and their combinations that are interacting with a moderating of "base" variable. For each interaction, it determines all possible level combinations of the involved regressors, which are further used to combine the posterior samples of the selected regression coefficients to calculate simple effects.

When the default effect coding scheme is used the simple effects are calculated for all levels of the interacting variables, as specified in the data. If a user specifies different contrasts for any of the interacting variables the simple effects for these variables are reported for the user-defined regressors. This distinction is reflected in the labels of the reported results: in the default case labels from the original factors are displayed; in the case of user-defined contrasts, the name of the regressor is displayed instead.

The summary of the posterior distribution of each simple effect contains the mean, standard deviation, posterior interval, which by default reports a central 95% interval, but can also be specified by the user, and a two-sided Bayesian p-value.

Note that for a Multinomial model intercepts and between-subject regressors have choice specific coefficients and thus simple effects are reported for each possible choice outcome. To perform the calculation for a Multinomial model an additional argument dep_var_name with a name of the dependent variable must be specified.

Value

Returns a list with the summary tables of the results; optionally returns the samples drawn from the posterior simple effects distributions.

Author(s)

Anna Kopyakova

Examples

# Use the colorad data set
data(colorad)

# Build and analyze the model
model <- BANOVA.model('Binomial')
banova_model <- BANOVA.build(model)
res_1 <- BANOVA.run(y ~ typic, ~ color*blurfac, fit = banova_model,
                    data = colorad, id = 'id', num_trials = as.integer(16), 
                    iter = 2000, thin = 1, chains = 2)
# Calculate simple effects with "blurfac" as a moderating vriable
simple_effects <- BANOVA.simple(BANOVA_output = res_1, base = "blurfac")

Function to print the table of effect sizes

Description

The analysis of variance is performed at level 1 (for the single level model) and level 2 equation of the Bayesian ANOVA see BANOVA-package. This makes it possible to capture the effects of level-1 or level-2 variables on the heterogeneity distribution of subjects, and compute sums of squares and effect sizes.

Usage

BAnova(x)

Arguments

Details

Measures of effect size in regression are measures of the degree of association between an effect (e.g., a main effect, an interaction, a linear contrast) and the dependent variable. They can be considered as the correlation between a categorical factor(effect) and the dependent variable. They are usually interpreted as the proportion of variance in the dependent variable that is attributable to each effect. In the package, partial Eta squared is calculated and displayed. It is defined as follows,

\eta^2= \frac{(SS_{effect})}{(SS_{effect} + SS_{error})}

Where: SS_effect= the sums of squares for the effect of interest
SS_error= the sums of squares for the error in the regression.

This equation is evaluated at each draw of the parameters, which allows for the calculation of not only the posterior mean, but also the credible interval of the effect size.

References

Fox, J. (2008) Applied Regression Analysis and Generalized Linear Models, Second Edition. Sage.

Fox, J. and Weisberg, S. (2011) An R Companion to Applied Regression, Second Edition, Sage.

Lakens, D. (2013) Calculating and Reporting Effect Sizes to Facilitate Cumulative Science: A Practical Primer for T-tests and ANOVAs, Frontiers in Psychology, Vol. 4, pp.863.

Gelman, A. and Pardoe, I. (2006) Bayesian Measures of Explained Variance and Pooling in Multilevel (Hierarchical) Models, TECHNOMETRICS, Vol. 48, NO. 2.

Examples

data(goalstudy)

library(rstan)
# or use BANOVA.run based on 'Stan'
res2 <- BANOVA.run(bid~progress*prodvar,  model_name = 'Normal', 
data = goalstudy, id = 'id', iter = 1000, chains = 2)
BAnova(res2)

Data for analysis of effects of typicality, blur and color on gist perception of ads

Description

Data from a mixed design experiment, where respondents were exposed to 32 ads, for 100 millisec. The ads were either typical or atypical (typical: 1 or 2). Respondents were exposed to ads that were either in full color or black-and-white (color: 1 or 2), and at different levels of blur (1=normal,5 = very high blur). These are between-subjects factors. The dependent variables are the response 0/1, and the response time. Typicality is a within-subjects variable.

Usage

data(bernlogtime)

Format

This R object contains within-subject variable: $typical is a factor with 2 levels "0" (typical ads) and "1"(atypical ads); between-subjects variables: $blur is a factor with two levels (1=normal,5 = very high blur). $color denotes a factor with 2 levels "1"(full color) and "2"(grayscale). $subject is the ID of subjects. $response denotes if the ad is correctly identified. $logtime is the response time.

$bernlogtime: 'data.frame': 3072 obs. of 6 variables:
... $ subject : int 5 5 5 5 5 5 5 5 5 5 ...
... $ typical : Factor w/ 2 levels "1","2": 1 2 1 1 1 2 2 2 2 1 ...
... $ blur : Factor w/ 2 levels "1","5": 1 1 1 1 1 1 1 1 1 1 ...
... $ color : Factor w/ 2 levels "1","2": 2 2 2 2 2 2 2 2 2 2 ...
... $ response: int 1 1 1 1 1 1 1 1 1 1 ...
... $ logtime : num 0.977 1.73 1.784 1 1.149 ...

References

Wedel, M and R. Pieters (2015). The Buffer Effect: The Role of Color when Advertising Exposures are Brief and Blurred, Marketing Science, Vol. 34, No. 1, pp. 134-143.

Examples

data(bernlogtime)
# model using the dependent variable : log of the response time(logtime) 
res1 <- BANOVA.Normal(logtime~typical, ~blur + color, bernlogtime, 
bernlogtime$subject, burnin = 1000, sample = 1000, thin = 1)
summary(res1)
table.predictions(res1)

# model using the dependent variable : response
res2 <- BANOVA.Bernoulli(response~typical, ~blur + color, bernlogtime, 
bernlogtime$subject, burnin = 1000, sample = 1000, thin = 1)
summary(res2)
table.predictions(res2)

Eye-movement data for analysis of print ad designs

Description

Data were collected in an experimental study in which 88 participants freely paged through a magazine at home or in a waiting room. While flipping through pages at their own pace, participants' eye-movements were recorded with infra-red corneal reflection eye-tracking methodology. In a subsequent memory task, participants were asked to identify the target brand in the ad as soon as possible by touching the correct brand name on the screen. Accuracy (accurate=1, inaccurate =0) of brand memory and response time were recorded for each ad and participant.

Usage

data(bpndata)

Format

This R object contains 3080 observations in the data (35 ads x 88 participants). The goal is to examine the effects of several ad design variables on both eye movements and memory. The variables include:

1. RESPONDENT_ID: ID number of a respondent;
2. AD_ID: ID number of an ad;
3. PAGE_NUM: page number in the magazine where an ad appears (1,2,3,...);
4. PAGE_POS: the right-side vs. left-side position on a page, 1 = right, 0 = left;
5. PIC_FIX: fixation count of the pictorial element (0, 1, 2, 3, ...);
6. PIC_SIZE: surface size of the pictorial element, in inches2;
7. RECALL_ACCU: whether a respondent accurately recalls the brand name, 1= yes, 0 = no;
8. RECALL_TIME: the time it takes a respondent to answer the brand recall question, in seconds.

$ bpndata: 'data.frame': 3080 obs. of 8 variables:
... $ RESPONDENT_ID: int 1 1 1 1 1 1 1 1 1 1 ...
... $ AD_ID : int 1 2 3 4 5 6 7 8 9 10 ...
... $ PAGE_NUM : int 2 5 6 11 13 14 17 18 21 22 ...
... $ PAGE_POS : int 0 1 0 1 1 0 1 0 1 0 ...
... $ PIC_FIX : int 0 2 1 1 1 2 0 3 3 8 ...
... $ PIC_SIZE : num 74.2 52.6 77.6 71.4 52.4 ...
... $ RECALL_ACCU : int 0 0 0 0 0 0 1 1 0 0 ...
... $ RECALL_TIME : num 2.56 1.04 2.76 2.8 2.28 2.32 2.04 2.04 2.48 0.6 ...

References

Wedel, M. and Pieters, R. (Autumn, 2000). Eye Fixations on Advertisements and Memory for Brands: A Model and Findings, Marketing Science, Vol. 19, No. 4, pp. 297-312

Examples

 

data(bpndata)
# within-subjects model using the dependent variable : PIC_FIX
library(rstan)
model <- BANOVA.model('Poisson')
stanmodel <- BANOVA.build(model)
res0 <- BANOVA.run(PIC_FIX ~ PIC_SIZE + PAGE_NUM + PAGE_POS, ~1, 
fit = stanmodel, data = bpndata, id = 'RESPONDENT_ID', 
iter = 200, thin = 1, chains = 2)
res0
# or
res1 <- BANOVA.Poisson(PIC_FIX ~ PIC_SIZE + PAGE_NUM 
+ PAGE_POS, ~1, bpndata, bpndata$RESPONDENT_ID, burnin = 1000, sample = 1000, thin = 1)
res1

# within-subjects model using the dependent variable : RECALL_ACCU
model_bern <- BANOVA.model('Bernoulli')
stanmodel_bern <- BANOVA.build(model_bern)
res2 <- BANOVA.run(RECALL_ACCU ~ RECALL_TIME + PAGE_NUM + PAGE_POS, ~1, 
fit = stanmodel_bern, data = bpndata, id = 'RESPONDENT_ID', 
iter = 200, thin = 1, chains = 2)
res2
# or
res3 <- BANOVA.Bernoulli(RECALL_ACCU ~ RECALL_TIME + PAGE_NUM 
+ PAGE_POS, ~1, bpndata, bpndata$RESPONDENT_ID, burnin = 1000, sample = 1000, thin = 1)
res3

Household Panel Data on Margarine Purchases

Description

Panel data on purchases of margarine by 204 households. Demographic variables are included.

Usage

data(choicedata)

Format

This is an R object that contains within-subjects variables and between-subjects variables:

$ choicePrice:‘data.frame’: 1500 obs. of 13 variables:
... $ hhid : int 2100016 2100016 2100016 2100016
... $ choice : int 1 1 1 1 1 4 1 1 4 1

Within-subject variables:

... $ PPk_Stk : num 0.66 0.63 0.29 0.62 0.5 0.58 0.29 ...
... $ PBB_Stk : num 0.67 0.67 0.5 0.61 0.58 0.45 0.51 ...
... $ PFl_Stk : num 1.09 0.99 0.99 0.99 0.99 0.99 0.99 ...
... $ PHse_Stk: num 0.57 0.57 0.57 0.57 0.45 0.45 0.29 ...
... $ PGen_Stk: num 0.36 0.36 0.36 0.36 0.33 0.33 0.33 ...
... $ PSS_Tub : num 0.85 0.85 0.79 0.85 0.85 0.85 0.85 ...

Pk is Parkay; BB is BlueBonnett, Fl is Fleischmanns, Hse is house, Gen is generic, SS is Shed Spread. _Stk indicates stick, _Tub indicates Tub form.

Between-subject variables:

... $ Income : num 32.5 17.5 37.5 17.5 87.5 12.5 ...
... $ Fam_Size : int 2 3 2 1 1 2 2 2 5 2 ...
... $ college : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ...
... $ whtcollar: Factor w/ 2 levels "0","1": 0 0 0 0 0 0 0 1 1 1 ...
... $ retired : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ...

Details

choice is a multinomial indicator of one of the 6 brands (in order listed under format). All prices are in $.

Source

Allenby, G. and Rossi, P. (1991), Quality Perceptions and Asymmetric Switching Between Brands, Marketing Science, Vol. 10, No.3, pp. 185-205.

References

Chapter 5, Bayesian Statistics and Marketing by Rossi et al.

Examples

 

data(choicedata)
# generate dataX(convert the within-subjects variables to a list)
dataX <- list()
for (i in 1:nrow(choicedata)){
  logP <- as.numeric(log(choicedata[i,3:8]))
  # Note: Before the model initialization, all numeric variables(covariates) 
  # must be mean centered
  dataX[[i]] <- as.data.frame(logP) - mean(logP)
}
dataZ <- choicedata[,9:13]
res <- BANOVA.Multinomial(~ logP, ~ college, dataX, dataZ, choicedata$choice, 
choicedata$hhid, burnin = 100, sample = 100, thin = 1)
summary(res)
predict(res,dataX[1:4], dataZ[1:4,])

Data for gist perception of advertising, study 1

Description

Data from an experiment in which one hundred and sixteen subjects (53 men; mean age 23, ranging from 21 to 28) were randomly assigned to one condition of a 5 (blur: normal, low, medium, high, very high) x 2 (color: full color, grayscale) between-participants, x 2 (image: typical ads, atypical ads) within-participants, mixed design. Participants were exposed to 40 images, 32 full-page ads and 8 editorial pages. There were 8 ads per product category, with 4 typical and 4 atypical ones, the categories being car, financial services, food, and skincare. Subjects were asked to identify each image being flashed for 100msec. as being an ad or not. The total number of correct ad identifications, for typical and atypical ads, are used as a dependent variable.

Usage

data(colorad)

Format

This R object contains within-subject variable $typic which is a factor with 2 levels "0" (typical ads) and "1"(atypical ads); between-subjects variables: $blur which is a numerical vairable denotes 5 different levels of blur (which must be mean centered), $blurfac is a categorical data corresponding to the levels of $blur, $color which is a factor with 2 levels "0"(full color) and "1"(grayscale). $id is the ID of subjects. $y is the number of correct identifications of the 16 ads of each subject for each level of $typic.

$ colorad:‘data.frame’: 474 obs. of 8 variables:
... $ id : int 1 1 2 2 3 3 4 4 5 5 ...
... $ typic : Factor w/ 2 levels "0","1": 0 1 0 1 0 1 0 1 0 1 ...
... $ y : int 8 6 12 6 11 9 9 11 14 14 ...
... $ blurfac : Factor w/ 5 levels "1","2","3","4",..: 2 2 4 4 2 2 3 3 1 1 ...
... $ color : Factor w/ 2 levels "0","1": 1 1 0 0 0 0 0 0 1 1 ...
... $ blur: num 3.69 3.69 4.79 4.79 3.69 ...

References

Wedel, M and R. Pieters (2015). The Buffer Effect: The Role of Color when Advertising Exposures are Brief and Blurred, Marketing Science, Vol. 34, No. 1, pp. 134-143.

Examples

 

data(colorad)
library(rstan)
# Build the model
model_bin <- BANOVA.model('Binomial')
stanmodel_bin <- BANOVA.build(model_bin)
out0 = BANOVA.run(y ~ typic, ~ color*blurfac, fit = stanmodel_bin,
                   data = colorad, id = 'id', num_trials = as.integer(16), 
                   iter = 100, thin = 2, chains = 1)
summary(out0)
# planned comparison
out0_contra = BANOVA.run(y ~ typic, ~ color*blurfac, fit = stanmodel_bin,
                          data = colorad, id = 'id', num_trials = as.integer(16), 
                          iter = 100, thin = 2, chains = 1, 
                          contrast = list(typic = c(-1,1)))
summary(out0_contra)

Data for gist perception of advertising, study 2

Description

Data from an experiment in which One hundred and forty eight subjects (71 men; age ranging from 21 to 28) were randomly assigned to one condition of a 2 (blur: normal, very high) x 2 (color: full color, grayscale, inverted) between-participants design. Participants were exposed to 25 ads for five brands in each of five categories. Ads were selected to be typical for the category, using the same procedure as in colorad. The product categories used were cars, financial services, food, skincare and fragrance. Images were flashed for 100 msec. and subjects were asked to identify whether the image was an ad or not, and if they identified it correctly as an ad, they were asked to indicate which category (out of five) was advertised. The total number of correct ad identifications and category identifications are used as dependent variables.

Usage

data(colorad2)

Format

This R object contains between-subjects variables: $B is a factor corresponding to the levels of blur (normal = 0, very high = 1), $C1 and $C2 are dummy variables denote 'grayscale' and 'inverted' levels of color. $C is the original factor denote the color with 3 levels. $ID is the ID of subjects. $Y1 is the number of correct identifications of the 25 ads of each subject. $Y2 is the number of correct identifications of the category, given the number of correct ad identifications.

$ colorad2:‘data.frame’: 148 obs. of 7 variables:
... $ ID : int 1 2 3 4 5 6 7 8 9 10 ...
... $ C1 : int 0 1 1 0 0 0 0 0 1 1 ...
... $ C2 : int 0 0 0 1 1 0 0 0 0 0 ...
... $ B : Factor w/ 2 levels "0","1": 1 1 0 0 1 0 0 1 0 1 ...
... $ Y1 : int 14 6 23 21 8 23 24 5 23 6 ...
... $ Y2 : int 2 3 8 8 2 15 10 1 13 0 ...
... $ C : Factor w/ 3 levels "1","2","3": 1 2 2 3 3 1 1 1 2 2 ...

References

Wedel, M and R. Pieters (2015). The Buffer Effect: The Role of Color when Advertising Exposures are Brief and Blurred, Marketing Science, Vol. 34, No. 1, pp. 134-143.

Examples

 

data(colorad2)
# factor C is effect coded
model_bin <- BANOVA.model('Binomial')
stanmodel_bin <- BANOVA.build(model_bin)
res0 <- BANOVA.run(Y1 ~ 1, ~ C + B + C*B, fit = stanmodel_bin, 
data = colorad2, id = 'id', num_trials = as.integer(25), 
iter = 100, thin = 1, chains = 2)
res0
# or use BANOVA.Binomial
res1 <- BANOVA.Binomial(Y1 ~ 1, ~ C + B + C*B, colorad2, colorad2$id, as.integer(25), 
burnin = 100, sample = 100, thin = 1) 

Data for the study of how brand attitudes were influenced by showing brands together with pleasant pictures

Description

The study investigated how brand attitudes were influenced by showing brands together with pleasant pictures. Attitude change via conditioning can result from either a direct transfer of affect from the picture to the brand, or from an indirect association of the brand and the picture in memory. In Sweldens' et al. (2010) experiment 1, indirect conditioning was implemented by presenting a brand repeatedly with the same picture, direct conditioning by presenting it simultaneously with different pictures. The pictures used were either neutral or positive. This study involved a mixed design, with a within-subject factor (cond = neutral, positive), and a between-subject factor (type = indirect, direct), as well as a within-subject mediator. Although the original mediation hypotheses are more intricate, here the mediation of the conditioning effect is investigated by measurements of attitudes towards the pictures that were shown with the brands (pict).

Usage

data(condstudy)

Format

This R object contains a between-subjects variable: type, which denotes a between-subject moderator. It has two levels, "indirect" and "direct". In the "indirect" condition the brands were shown with the same images, in the indirect condition the brands were shown with different images; Within-subject variables: cond, a within-subject factor with 2 levels: "pos", and "xneu", which indicates whether each brand was shown with a neutral (xneu) or positive (pos) emotional image. pict, a within-subject mediator variable measuring the valence (positive/negative) of the emotional image the respondent remembers the brand to have been shown with. att, a dependent variable which denotes the ratings of attitudes toward brands.

$ condstudy: 'data.frame': 888 obs. of 5 variables:
... $ id : int 2 2 2 2 2 2 3 3 3 3 ...
... $ att : num 2.94 2.44 3.44 1.67 1.67 ...
... $ cond: Factor w/ 2 levels "pos","xneu": 1 1 1 2 2 2 1 1 1 2 ...
... $ type: Factor w/ 2 levels "direct","indirect": 2 2 2 2 2 2 2 2 2 2 ...
... $ pict: int 6 7 6 2 4 5 9 3 2 5 ...

References

Sweldens, S., Osselaer, S. and Janiszewski, C. (2010) Evaluative Conditioning Procedures and the Resilience of Conditioned Brand Attitudes. Journal of Consumer Research, Vol. 37.

Wedel, M. and Dong, C. (2016) BANOVA: Bayesian Analysis of Variance for Consumer Research. Submitted.

Examples

# condstudy_sub is a subset of condstudy with 180 obs. and the same variables
data(condstudy_sub)
model <- BANOVA.model('Normal')
stanmodel <- BANOVA.build(model)
out2 <- BANOVA.run(att~cond+pict, ~type, fit = stanmodel, data = condstudy_sub, 
                  id = 'id', iter = 500, thin = 1, chains = 2)    
conv.diag(out2)
summary(out2) 
table.predictions(out2)  
BANOVA.floodlight(out2, var_factor = 'type', var_numeric = 'pict')
cat(out2$model_code)

out3 <- BANOVA.run(pict~cond, ~type, fit = stanmodel, data = condstudy_sub, 
                  id = 'id', iter = 500, thin = 1, chains = 2)
conv.diag(out3)
summary(out3)
BANOVA.mediation(out2, out3, xvar='cond', mediator='pict')

A subset of data for the study of how brand attitudes were influenced by showing brands together with pleasant pictures

Description

This is a subset of the data 'condstudy' with 180 obs.

Usage

data(condstudy_sub)

Format

This R object contains a between-subjects variable: type, which denotes a between-subject moderator. It has two levels, "indirect" and "direct". In the "indirect" condition the brands were shown with the same images, in the indirect condition the brands were shown with different images; Within-subject variables: cond, a within-subject factor with 2 levels: "pos", and "xneu", which indicates whether each brand was shown with a neutral (xneu) or positive (pos) emotional image. pict, a within-subject mediator variable measuring the valence (positive/negative) of the emotional image the respondent remembers the brand to have been shown with. att, a dependent variable which denotes the ratings of attitudes toward brands.

$ condstudy_sub: 'data.frame': 180 obs. of 5 variables:
... $ id : int 2 2 2 2 2 2 3 3 3 3 ...
... $ att : num 2.94 2.44 3.44 1.67 1.67 ...
... $ cond: Factor w/ 2 levels "pos","xneu": 1 1 1 2 2 2 1 1 1 2 ...
... $ type: Factor w/ 2 levels "direct","indirect": 2 2 2 2 2 2 2 2 2 2 ...
... $ pict: int 6 7 6 2 4 5 9 3 2 5 ...

Examples

# condstudy_sub is a subset of condstudy with 180 obs. and the same variables
data(condstudy_sub)
library(rstan)
model <- BANOVA.model('Normal')
stanmodel <- BANOVA.build(model)
out2 <- BANOVA.run(att~cond+pict, ~type, fit = stanmodel, data = condstudy_sub, 
                  id = 'id', iter = 500, thin = 1, chains = 2)    
conv.diag(out2)
summary(out2) 
table.predictions(out2)  
BANOVA.floodlight(out2, var_factor = 'type', var_numeric = 'pict')
cat(out2$model_code)

out3 <- BANOVA.run(pict~cond, ~type, fit = stanmodel, data = condstudy_sub, 
                  id = 'id', iter = 500, thin = 1, chains = 2)
conv.diag(out3)
summary(out3)
BANOVA.mediation(out2, out3, xvar='cond', mediator='pict')

Function to display the convergence diagnostics

Description

The Geweke diagnostic and the Heidelberg and Welch diagnostic are reported. These two convergence diagnostics are calculated based on only a single MCMC chain. Both diagnostics require a single chain and may be applied with any MCMC method. The functions geweke.diag, heidel.diag in coda package is used to compute this diagnostic.

Geweke's convergence diagnostic is calculated by taking the difference between the means from the first n_A iterations and the last n_B iterations. If the ratios n_A/n and n_B/n are fixed and nA +nB < n, then by the central limit theorem, the distribution of this diagnostic approaches a standard normal as n tends to infinity. In our package, n_A= .2*n and n_B= .5*n.

The Heidelberg and Welch diagnostic is based on a test statistic to accept or reject the null hypothesis that the Markov chain is from a stationary distribution. The present package reports the stationary test.The convergence test uses the Cramer-von Mises statistic to test for stationary. The test is successively applied on the chain. If the null hypothesis is rejected, the first 10% of the iterations are discarded and the stationarity test repeated. If the stationary test fails again, an additional 10% of the iterations are discarded and the test repeated again. The process continues until 50% of the iterations have been discarded and the test still rejects. In our package, eps = 0.1, pvalue = 0.05 are used as parameters of the function heidel.diag.

Usage

conv.diag(x)

Arguments

Value

conv.diag returns a list of two diagnostics:

References

Plummer, M., Best, N., Cowles, K. and Vines K. (2006) CODA: Convergence Diagnosis and Output Analysis for MCMC, R News, Vol 6, pp. 7-11.

Geweke, J. Evaluating the accuracy of sampling-based approaches to calculating posterior moments, In Bayesian Statistics 4 (ed JM Bernado, JO Berger, AP Dawid and AFM Smith). Clarendon Press, Oxford, UK.

Heidelberger, P. and Welch, PD. (1981) A spectral method for confidence interval generation and run length control in simulations, Comm. ACM. Vol. 24, No.4, pp. 233-245.

Heidelberger, P. and Welch, PD. (1983) Simulation run length control in the presence of an initial transient, Opns Res., Vol.31, No.6, pp. 1109-44.

Schruben, LW. (1982) Detecting initialization bias in simulation experiments, Opns. Res., Vol. 30, No.3, pp. 569-590.

Examples

data(goalstudy)

library(rstan)
res1 <- BANOVA.run(bid~progress*prodvar, model_name = "Normal", data = goalstudy, 
id = 'id', iter = 100, thin = 1) 
conv.diag(res1)
# might need pairs() to confirm the convergence

Data for the study of the impact of the variety among means on motivation to pursue a goal

Description

The study investigated how the perceived variety (high vs. low) among products, as means to a subjects' goal, affects their motivation to pursue that goal. The hypothesis was that only when progress toward a goal is low, product variety increases motivation to pursue the goal. In the study, one hundred and five subjects were randomly assigned to conditions in a 2 (goal progress: low vs. high) by 2 (variety among means: low vs. high) between-subjects design. The final goal was a "fitness goal", and the products used were protein bars; variety was manipulated by asking subjects to think about how the products were similar (low) or different (high); goal progress was primed by asking subjects questions regarding the frequency of their recent workouts on low (0,1,...,5 or more) versus high (5 or less, 6,7,..., 10) frequency scales. Subjects were asked questions regarding the similarity of protein bars, and the bid they were willing to make for the bars, used as dependent variables in the study.

Usage

data(goalstudy)

Format

This R object contains between-subjects variables: progress, which denotes the progress toward a goal (1:low , 2: high ); prodvar, which denotes the amount of variety within the means to goal attainment (1:low , 2:high); perceivedsim, which is a seven-point scale dependent variable measuring the perceived similarity of the set of products (1 = not at all similar, 7 = very similar); and bid which denotes the amount that subjects would be willing to pay for the products .

$ goalstudy: 'data.frame': 105 obs. of 5 variables:
... $ id : int 1 2 3 4 5 6 7 8 9 10 ...
... $ perceivedsim : int 5 7 2 2 5 5 5 4 5 7 ...
... $ progress : Factor w/ 2 levels "1","2": 1 1 2 2 2 1 2 1 2 1 ...
... $ prodvar : Factor w/ 2 levels "1","2": 2 1 2 1 1 1 1 2 1 1 ...
... $ bid : num 5 0 1 15 3 10 5 4.5 3 0.75 ...

References

Etkin, J. and Ratner, R. (2012) The Dynamic Impact of Variety among Means on Motivation. Journal of Consumer Research, Vol. 38, No. 6, pp. 1076 - 1092.

Examples

data(goalstudy)
library(rstan)

# single level model
res1 <- BANOVA.run(bid~progress*prodvar, model_name = "Normal", 
data = goalstudy, id = 'id', iter = 1000, thin = 1, chains = 2) 
BAnova(res1)
table.pvalues(res1)
trace.plot(res1)
table.predictions(res1)
# pairs(res1, pars = c("beta1[1]", "tau_ySq"))

Data for the study of relation between Conspicuous, Brand Usage, Self-Brand Connection and attitudes toward the brand

Description

The study is a between-subjects experiment which has factor (conspicuousness: low vs. high) and one measured variable (self-brand connection). The goal is to show that conspicuous brand use negatively affects attitudes toward the user and the brand only for observers low in self-brand connection. One hundred fifty-four participants were exposed to a video manipulating conspicuous brand usage. Participants completed the study by answering several questions which are used to measure the dependent (attitude) and independent (self-brand connection) variables in the model.

Usage

data(ipadstudy)

Format

This R object contains between-subjects variables: $owner is an indicator variable. If the subject owns iPad or iPhone, then owner = 1. It is equal to 0 otherwise. $age denotes the age of subjects. $gender denotes the gender of subjects. gender = 1 if the subject is a female, 0 otherwise. $conspic is an indicator variable related to conspicuousness. conspic = 1 if conspicuousness is high. $selfbrand denotes the self-brand connection for Apple. $id is the id of subjects. $attitude denotes the attitudes towards the brand which is the continuous dependent variable. $apple_dl is a seven-point scale variable which denotes the attitudes (dislike = 1,..., like = 7)

$ ipadstudy: 'data.frame': 154 obs. of 9 variables:
... $ id : int 1 2 3 4 5 6 7 8 9 10 ...
... $ attitude : num 3 5.33 5.67 5.33 6 ...
... $ owner : num 0 0 0 1 1 0 1 0 1 0 ...
... $ age : int 19 33 25 41 38 33 37 46 41 55 ...
... $ gender : num 0 0 1 0 1 1 1 0 1 1 ...
... $ conspic : num 0 1 0 1 1 0 0 1 0 1 ...
... $ selfbrand : num -2.304 1.696 -0.161 -0.447 0.267 ...
... $ apple_dl : int 3 6 6 5 6 4 7 7 5 5 ...

References

Ferraro,R., Kirmani, A. and Matherly, T., (2013) Look at Me! Look at Me! Conspicuous Brand Usage, Self-Brand Connection, and Dilution. Journal of Marketing Research, Vol. 50, No. 4, pp. 477-488.

Examples

data(ipadstudy)

# mean center covariates
ipadstudy$age <- ipadstudy$age - mean(ipadstudy$age) 
ipadstudy$owner <- ipadstudy$owner - mean(ipadstudy$owner )
ipadstudy$gender <- ipadstudy$gender - mean(ipadstudy$gender)

res <- BANOVA.Normal(attitude~1, ~owner + age + gender + selfbrand*conspic, 
ipadstudy, ipadstudy$id, burnin = 100, sample = 100, thin = 1 )
summary(res)
# use apple_dl as the dependent variable
res <- BANOVA.ordMultinomial(apple_dl~1, ~owner + age + gender + selfbrand*conspic, 
ipadstudy, ipadstudy$id, burnin = 100, sample = 100, thin = 2 )
summary(res)
table.predictions(res)

Create a matrix of output plots from a BANOVA object

Description

A pairs method that is customized for MCMC output.

Usage

## S3 method for class 'BANOVA'
pairs(x, ...)

Arguments

Details

For a detailed description see pairs.stanfit

Examples

library(rstan)
data(ipadstudy)
res_1 <- BANOVA.run(attitude~owner + age + gender + selfbrand*conspic, 
model_name = 'Normal', data = ipadstudy, id = 'id', iter = 1000, 
thin = 1, chains = 2)
# pairs(res_1, pars = c("beta1[1]","beta1[2]"))

Function to print the table of means

Description

Output of this function is a table of means for the categorical predictors (and their interactions) at either within- or between- subjects level. Statistics of interest such as credible intervals and standard deviations of the means are also computed. Means of numeric variables and their interactions will not be computed.

Usage

table.predictions(x)

Arguments

Examples

data(goalstudy)

library(rstan)
# or use BANOVA.run based on 'Stan'
res <- BANOVA.run(bid~progress*prodvar, model_name = "Normal", 
data = goalstudy, id = 'id', iter = 1000, thin = 1, chains = 2)
table.predictions(res)

Function to print the table of p-values

Description

Computes the Baysian p-values for the test concerning all coefficients/parameters:

For p = 1,...,P
H_0:\theta_{j,k}^{p,q}=0
H_1:\theta_{j,k}^{p,q} \neq 0

The two-sided P-value for the sample outcome is obtained by first finding the one sided P-value, min(P(\theta_{j,k}^{p,q}<0),P(\theta_{j,k}^{p,q}>0 )) which can be estimated from posterior samples. For example, P(\theta_{j,k}^{p,q}>0) = \frac{n_+}{n}, where n_+ is the number of posterior samples that are greater than 0, n is the target sample size. The two sided P-value is P_\theta(\theta_{j,k}^{p,q}) = 2*min(P(\theta_{j,k}^{p,q}<0),P(\theta_{j,k}^{p,q}>0 )).

If there are \theta_{j,k_1}^{p,q},\theta_{j,k_2}^{p,q},...,\theta_{j,k_J}^{p,q} representing J levels of a multi-level variable, we use a single P-value to represent the significance of all levels. The two alternatives are:

H_0:\theta_{j,k_1}^{p,q} = \theta_{j,k_2}^{p,q} = \cdots = \theta_{j,k_J}^{p,q}=0
H_1 : some \theta_{j,k_j}^{p,q} \neq 0

Let \theta_{j,k_{min}}^{p,q} and \theta_{j,k_{max}}^{p,q} denote the coefficients with the smallest and largest posterior mean. Then the overall P-value is defined as

min(P_\theta (\theta_{j,k_{min}}^{p,q}), P_\theta(\theta_{j,k_{max}}^{p,q})).

Usage

table.pvalues(x)

Arguments

Source

It borrows the idea of Sheffe F-test for multiple testing: the F-stat for testing the contrast with maximal difference from zero. Thank Dr. P. Lenk of the University of Michigan for this suggestion.

Examples

data(goalstudy)

library(rstan)
# or use BANOVA.run
res1 <- BANOVA.run(bid~progress*prodvar, model_name = "Normal", 
data = goalstudy, id = 'id', iter = 1000, thin = 1, chains = 2) 
table.pvalues(res1)

Function to plot the trace of parameters

Description

Function to plot the trace of all coefficients/parameters. The plots can be saved as a pdf file.

Usage

trace.plot(x, save = FALSE)

Arguments

Examples

data(goalstudy)

library(rstan)
# or use BANOVA.run
res1 <- BANOVA.run(bid~progress*prodvar, model_name = "Normal", 
data = goalstudy, id = 'id', iter = 1000, thin = 1, chains = 2)
trace.plot(res1)