BACKGROUND: Info & Data Munging Code

These results reference the following RMDs: http://tinyurl.com/ycb5pfgr (HW 1 instructions & survey details) and http://tinyurl.com/yc6v2ahy (in-class example).

For easy reference, below is the code from our in-class example – which cleaned and prepped the data to for this analysis. When running this RMD, key elements to remember are: 1) set working directory, 2) load the below libraries (reshape2, dplyr & corrplot) and 3) read the CSV into mcNum data frame.

## [1] "library(reshape2)"
## [1] "library(dplyr)"
## [1] "library(corrplot)"
## [1] "Q32M" "Q32N" "Q32O" "Q32P"

BONUS CODE: Renaming Attributes

To make it easier to read and interpret my visuals, I’ve renamed the response optoin variables (e.g. r1, which was ‘Looks exciting and fun to ride’ in the survey, will now read as ‘exciting’).

QUESTION 1: Key Results?

QUESTION: Using the motorcycle perception data prepared as above, do a perceptual mapping analysis using principal components analysis and biplots as described in CF, and as we reviewed in class. Include your best figures and any other summaries of results.

RESPONSE: Below are the visuals that best help me understand the response dataset, which include a correlation plot, scree plot and perceptual map.

Correlation plot: Provides good context on the relationships between data attributes. For example, it’s reasonable to expect that European and Japanese have a low correlation, while trendsetting and exciting are more highly correlated. Nothing way off. It also shows us points that might surprise us, areas we wouldn’t have closely linked (e.g. exciting and aggressive, exciting and high performance both positively correlated) or areas we’d expect (e.g. boring and exciting are negatively correlated).

Scree plot: Shows that the bulk of variation is just in the first two principal components, which is great for perceptual map making. This is also in the summary stats of the mean PCA.

Perceptual map: Shows Honda owns over the other brands is wimpy, lightweight, European; Kawakasaki is overdone, Japanese, young; BMW is high performance; and Yamaha is tough and American. The positive scores are related to larger means, and using the means is in the PCA analysis is a ‘quick and easy’ way to compare brands.

Aside: It’s important to note the broader context of the responses’ summary statistics. For example, the highest means are exciting, easy and tough while the lowest means are boring, wimpy and overdone. It’s critical to remember that a single adjective doesn’t on the perceptual map doesn’t reflect the strength of brand on that attribute (e.g. easy doesn’t have a brand near to it, but it is still a highly important adjective). [Not listed here due to space constraints, but the command –for future reference– is summary(mcNumCastOrd).]

BikeCor=cor(mcNumCastOrd[,3:19])
typeof(BikeCor) #DOUBLE checking that that data is two dimensional or DOUBLE :) 
## [1] "double"
corrplot::corrplot(BikeCor)

sc.Bike <- scale(mcNumCastOrd[,3:19])
BikePCA <- prcomp(mcNumCastOrd[,3:19]) #Data reduction re: dimensions -- but graph is really messy, looks like an *, unable to tease out much meaning from it 
ratings.means=aggregate(.~brandName,data=mcNumCastOrd[,c(2:19)],mean) #using mean (over mode, median has sim results)
rownames(ratings.means)=ratings.means[,1]  
ratings.means=ratings.means[,-1]  
ratings.means.pca=prcomp(ratings.means,scale=TRUE)
summary(ratings.means.pca)
## Importance of components%s:
##                           PC1    PC2     PC3       PC4
## Standard deviation     3.4036 2.1755 0.82623 2.441e-15
## Proportion of Variance 0.6814 0.2784 0.04016 0.000e+00
## Cumulative Proportion  0.6814 0.9598 1.00000 1.000e+00
plot(ratings.means.pca, type='l')

biplot(ratings.means.pca,main="Brand Positioning in 2D", cex=c(2,1))

QUESTION 2: Meaningful?

QUESTION: Your map will be in two dimensions. It’s supposed to be a map, right? How well does your best map capture all the variation in the ratings? All maps are not created equal. What is your assessment based on?

RESPONSE: Yes, this map is in two dimensions – which is sensible given the dataset’s results. The first two principal components capture the bulk of variance (~95%), which shows that the observations greatly relate to the underlying dimensions. Additionally, the brands are all fairly distinctly plotted in the perceptual map. This indicates that the perceptual map created has accounted for the bulk of the variation and, to me, that it’s a reasonable visual to use to spark conversation about the competitive landscape and brand identity.

QUESTION 3: BMW Asks, “So What?”

QUESTION: Suppose BMW would like to differentiate itself visually from the three other brands as much as possible. What possibilities would you suggest it investigate? What steps would you suggest it take, next?

RESPONSE: Due to the low sample size of the survey, I would emphasize that the results from this data are directional. I would highly recommend additional research –both quant (e.g. 1,000 person survey) and qual (e.g. focus groups, deep dive interviews)– in order to confirm and build out these findings. Specific research recommendations would be subject to and tailored for BMW’s budget, timing, existing research and stakeholder needs.

With that caveat in mind, I would then recommend to BMW that they have the opportunity to really lean into their existing strength, which is an identity around high-performance that no other brand owns as distinctly. BMW could also better own the brand space around trendsetting and excitement. No brand is greatly own those attributes – and they are somewhat associated to high-performance. Again, this is an initial recommendation and needs fine tuning.

In addition to further brand research, I would also suggest begin small message testing (A|B bucket testing in CRM data, for example) to track how the messaging resonates via behavior as well as to track early indicators of marketing effectiveness.

QUESTION 4: Key Considerations?

QUESTION: Based on what you know from the information provided here, what issues do you think need to be considered when intepreting your results?

RESPONSE: When interpreting the results, it is important to consider the following:

  1. The population surveyed – small sample size, potential bias, recency of survey

  2. The survey tool used – respondent fatigue, right set of attributes (anything missing? like german?), any other qs asked to add context (share of consideration, awareness of the brand, likelihood to use)

  3. The statistical techniques used – how does removing a variable impact the results/perceptual map (very different or similar)? how would just using 80% of the data impact the results (very different or similar)? how does leaving in incomplete response set impact the results? how does using the median or mode in the aggregation step shift the results? (which are generally better for ordinal or categorical data)

  4. The existing BMW data – these survey results live in a broader context of research (e.g. third-party, internal behavioral data) BMW has conducted, where do our findings results diverge or dovetail?

  5. The stakeholders – will people viewing this recognize that the perceptual map isn’t about the strength of the brand to an attribute but the relative broader competitve landscape? will they understand the response set limitations? What is their agenda?

NOTE: This HTML RMD appears to print three pages | http://rpubs.com/tatoonie/309048