Conjoint Analysis

Conjoint Analysis is a survey based technique to identify how customers value various attributes that make up an individual product. Products and services are bundles of features that customers consider jointly and while making a purchase decision they must make trade offs. Conjoint analysis helps companies and business owners determine importance of different features and their money value. This also helps them identify optimal price for their products and services.

Basic Setup

In this example, we will discover how to design a survey to get the information necessary, analyze and quantify importance of various features in a product and how to interpret the results. But first , let’s begin by setting up R environment and loading required libraries.

#set working directory
setwd("C:/Users/awani/Documents/GitHub/50daysofAnalytics/Day 10 - Conjoint Analysis")

# load libraries
if (!require("pacman")) install.packages("pacman")
pacman::p_load(conjoint, DoE.base, knitr, dplyr, kableExtra, ggplot2)

options(scipen = 999)
options(digits = 3)

Experiment Design

A small business owner wants to use conjoint analysis to understand what features in his product are most attractive to his customer and how should he price them. He sells chocolates with three different attributes.

Chocolate - Milk, Dark Organic
Center - Soft, Chewy or Plain
Nuts - Mixed, Almonds or None

Let first figure out how should the survey for conjoint be designed. In this example, 27 different varieties of chocolates can be manufactured and it might to present all those options to a customer and ask their preference but in real world scenario, we will have enormous possible combinations. Asking a customer to rate all different combinations will not only be expensive but inaccurate. To avoid this issue, we can design a survey to just ask questions about few combinations and then predict their response for rest of the combinations. This process of often referred as experiment design.

###Experiment Design for asking Conjoint Questions

#- identify number of questions required. Define level and factors
NumeberOfQuestions = nrow(oa.design(nlevels=c(3,3,3)))

#- create dummy data
data = expand.grid(Chocolate = c("Milk","Dark","Organic"),
                   Center = c("Plain", "Chewy", "Soft"),
                   Nuts = c("Mixed", "Almonds","None"))

#- Combinations to enquire
selectedComb = caFactorialDesign(
  data = data,
  type='fractional',
  cards=NumeberOfQuestions)

#- print Selected Combinations
kable(selectedComb) %>%
  kable_styling(bootstrap_options = c("striped", "hover", "condensed", "responsive"))

	Chocolate	Center	Nuts
3	Organic	Plain	Mixed
5	Dark	Chewy	Mixed
7	Milk	Soft	Mixed
11	Dark	Plain	Almonds
13	Milk	Chewy	Almonds
18	Organic	Soft	Almonds
19	Milk	Plain	None
24	Organic	Chewy	None
26	Dark	Soft	None

Predicting Responses for other combinations

Using orthogonal factorial design, we identified 9 combinations to include in customer survey out of possible 27 combinations. It saves time, money and effort. Survey takers was asked if they like the combination or not and their reposes were noted for all 9 combinations. We will use these 9 responses to train out logistic regression model and use to predict rest of combinations.

# add response column
selectedComb$Response = c(0,0,1,1,1,1,1,0,0)

# logistic regression
logit=glm(Response ~ factor(Chocolate) + factor(Center) + factor(Nuts),
            family=binomial(link='logit'), data=selectedComb)

# predict
data$response = ifelse(predict(logit,data,type="response") > 0.5,1,0)