Which Product Attributes are Most Important? Full Profile Conjoint Analysis
Conjoint Series — Scene 2
Conjoint Series — Scene 1 — How to Understand Consumer Preference?
Conjoint Series — Scene 3— Designing Orthogonal Survey?
Conjoint Series — Scene 4 — Regression Analysis for All Respondent
Conjoint Series — Scene 5 — Market Segmentation using Conjoint Analysis
“I think the best way to see how conjoint works is by learning from old article published in Harvard Business Review.” Dess starting to explain.
In 1975, Harvard Business Review published an article title “New Way to Measure Customer’s Judgment”.
Let say a company wanted to market new spot remover for carpet and upholstery. The R&D department has created a product that able to handle tough and stubborn sport. Management has decided that there will be five attributes that they think will be the main consideration when the customer purchase the product: package design, brand name, price, a stamped endorsement from “Good Housekeeping” magazine, and money back guarantee.
After few discussion, business meetings and workshop, the team come up with few variable to understand customer preference.
Three package design are considered, pictured below.
They have three brand under considerations: K2R, Glory and Bissel. Out of three for this brand study, two are competitor brand and one are the new brand choice for the product. Three alternative choice are considered: $1.19, $1.39, and $1.59.
You can see that there are three alternatives for each factor, they are called three-level factor. The other two attributes or factor such as the stamp from “Good Housekeeping” and money back guarantee are only have two alternative. Yes, this mean the product is stamped with “Good Housekeeping” or guaranteed to get they money back if the product is broken or not satisfying. The other alternative is this product didn’t stamp with “Good Housekeeping” seal nor money back guarantee. This is called two-level-factor.
“Can you see what is happening here?” asked Dess
“Well, the combination supposed to be 3 x 3 x 3 x 2 x 2 or equal to 108 combinations. I think customer as respondent do not want to rank boring 108 combinations for the benefit of the company without getting something big in return. Is that what happened? The company give something big? Such as ticket concert?” replied Ray.
“Excellent observation,” Dess replied, and then continue explaining.
As the factors and levels to observe from a product grow, the combinations of product will grow exponentially. Luckily, the respondent do not need to rank all combinations of product profile. In the “paper” about “orthogonal design” we can reduce the number of combination into handful that can be accept by normal human being.
For example, in this case, from 108 combinations, users can rank only 18 combination of them. They way we check it is by measure the correlation matrix between variable. They suppose to be uncorrelated. This way, we can be sure that the regression variable will be correctly identify the true value of the coefficient.
“Owh… orthogonal design wanted to make sure that between variable, there is no correlation.” Abe murmuring. “Because when we have correlated variable, the coefficient of the regression variable will not showing the real coefficient”.
“Yes you are correct.” Dess smiling. “The real questions is, how to find correct combination? I will explain later. Now, lets focus on what kind of insight that we can get from doing the regression for this case.
For example, we have one response and the result of the response is below.
What we do next is doing the regression.
”Abe, are you ready for the regression? Lets do it in R”.
“Absolutely”. Abe response opening his laptop and start R Studio. With few strokes, he start to code. But then he hesitate.
“Wait, what is the response variable here?” he asked politely.
Dess explained that the response variable is the rank, and treat is a number instead of factor (categorical) data. So, the program will predict what is the customer rank of the product combination. However, one caveat that Abe need to be aware of is that we want to predict that the better the rank, the weight of the regression result should show higher number. To be able to do this, Abe need to reverse the rank number. Instead of rank number 1 printed as 1, the value will be reverse to 18. The way to do it is by subtracting 19 by the rank.
# load library
library(readxl)
library(janitor)
# read data
spot_remover = read_excel("Spot Remover.xlsx",sheet = "Sheet1")
# inverse rank
spot_remover$Inverse_rank = 19 - spot_remover$Rank
# cleaning column name
spot_remover <- janitor::clean_names(spot_remover)Here is the result.
Now, all ready. Dess start inputting the regression code and showing the result.
# regression
lm_spot_remover = lm(inverse_rank ~ package_design + brand_name +
price + good_housekeeping_seal_y_n +
money_back_guarantee_y_n, data=spot_remover)“Now, can we see the result?” ask Dess
#Review the results
summary(lm_spot_remover)Abe show the result to everyone in the room.
Call:
lm(formula = inverse_rank ~ package_design + brand_name + price +
good_housekeeping_seal_y_n + money_back_guarantee_y_n, data = spot_remover)Residuals:
Min 1Q Median 3Q Max
-1.3333 -0.5000 0.1667 0.4583 1.1667
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 8.0000 0.6351 12.597 5.09e-07 ***
package_designB 8.0000 0.5500 14.546 1.47e-07 ***
package_designC 4.5000 0.5500 8.182 1.85e-05 ***
brand_nameGlory -2.0000 0.5500 -3.637 0.005430 **
brand_nameK2R -1.5000 0.5500 -2.727 0.023323 *
price$1.39 -2.8333 0.5500 -5.152 0.000602 ***
price$1.59 -7.6667 0.5500 -13.940 2.13e-07 ***
good_housekeeping_seal_y_nYes 1.5000 0.4763 3.149 0.011750 *
money_back_guarantee_y_nYes 4.5000 0.4763 9.448 5.73e-06 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.9526 on 9 degrees of freedom
Multiple R-squared: 0.9831, Adjusted R-squared: 0.9682
F-statistic: 65.62 on 8 and 9 DF, p-value: 4.492e-07“Interesting result.” Ray whispered, barely heard.
“What the number tell us Ray?” asked Dess.
“Well… Lets break down how we can analyze the result from different angle. First, we can see that based on p-value of 0.05, all attributes are important. Second, R2 value indicate that the attributes can explain 98% of the variation in consumer ranking.” Ray replied.
“Anything else? Anyone want to volunteer?”
Abe, Ran and Ray stay quiet. So, Dess adding few insights that has been put across by Ray:
- Residual standard error of 0.95 suggest that 95% of predicted rank will be accurate within 1.9 variation.
- We don’t see package design A as the attributes. This is because, the coefficient of package design A considered as 0. Thus the coefficient of package design B implies that selecting package design B will rank 8 higher compare to A and rank 3.5 higher compare to package design C. The 8 difference is calculated from the difference between coefficient value of package design B and package design A, 8–0. While 3.5 coming from the difference between coefficient value of package design B and package design C, 8–4.5. Using the same logic, we can be sure that package design C will lead to rank 4.5 higher compared to package design A.
- Using the same logic from number 2, Bissel is the strongest brand because it can lead to rank 2 higher compared to brand Glory and rank 1.5 higher compared to brand K2R.
- Does approval from “Good Housekeeping” magazine matter? Of course. According to the model, the approval will lead to rank 1.5 higher compared to non sealed product.
- Money back guarantee also very strongly preferred by the consumers because it can lead to rank 4.5 higher to the spot remover without the money back guarantee.
- What about price? Cheaper is more preferred.
From those 5 attributes, can we rank from the most important attribute to the lest important attribute? Certainly. We can sort the importance by calculate the spread within attribute.
“Wow… that was interesting insights. So, the way this customer select the product is based on the design first and price second and Good Housekeeping seal is the lest important”.
“How to design the best product for this customer?” Ran, finally asking. She need to create insightful proposal to the client.
“Well, we can calculate based on the coefficients. The product that would be preferred by the customer is design A, priced at $1.39, with money back guarantee and Good Housekeeping seal”.
“Now, lets continue our discussion. What is orthogonal design? How to create orthogonal design?”
Source:
New Way to Measure Consumers’ Judgments. (1975, July 1). Harvard Business Review. https://hbr.org/1975/07/new-way-to-measure-consumers-judgments.
Winston, W. L. (2014, January 8). Marketing Analytics: Data-Driven Techniques with Microsoft Excel.
