Hire Me (as a Data Scientist!), Part IV

Since this is my last post in the beer review series, I’ll keep this short and sweet in terms of analysis. Having done all of this, I do have a few reflections I would like to share after doing the One-Size-Fits-All-Data-Science Interview that I have included at the end.

Our final question to answer is:

Lastly, if I typically enjoy a beer due to its aroma and appearance, which beer style should I try?

This is a pretty broad question and should be able to be answered without many pitfalls, so let’s get started.

library(ggplot2)
library(data.table)
beer <- fread("data/beer_reviews.csv") 

From a bird’s eye view, it seems like the most sensible thing to do would be to look at our data from the highest level, then just zoom in until we have the level of granularity we feel answers the question well. Let’s first average all the beer styles to get a rough estimate of how a beer style fairs on the variables we are interested in, and additionally see how much variability is associated with that measurement.

sem <-function(x) {sd(x)/sqrt(length(x))}

question4.means <- beer[, .(mean_aroma = mean(review_aroma), mean_appearance = mean(review_appearance), mean_overall = mean(review_overall),
                            sem_aroma = sem(review_aroma), sem_appearance = sem(review_appearance), sem_overall = sem(review_overall),
                            sd_aroma = sd(review_aroma), sd_appeareance = sd(review_appearance),sd_overall = sd(review_overall)), 
                        by = beer_style]
question4.means

##                           beer_style mean_aroma mean_appearance mean_overall
##   1:                      Hefeweizen   3.761735        3.828293     3.929626
##   2:              English Strong Ale   3.749427        3.852469     3.783288
##   3:          Foreign / Export Stout   3.828366        4.039015     3.877679
##   4:                 German Pilsener   3.387159        3.572444     3.731573
##   5:  American Double / Imperial IPA   4.097782        4.078916     3.998017
##  ---                                                                        
## 100:                          Gueuze   4.117574        4.034864     4.086287
## 101:                            Gose   3.783528        3.908163     3.965015
## 102:                        Happoshu   2.595436        2.925311     2.914938
## 103:                           Sahti   3.827992        3.655985     3.700283
## 104: Bière de Champagne / Bière Brut   3.734704        4.045889     3.648184
##        sem_aroma sem_appearance sem_overall  sd_aroma sd_appeareance sd_overall
##   1: 0.003668940    0.003595729 0.004051038 0.6129217      0.6006912  0.6767538
##   2: 0.008118012    0.007674182 0.009323636 0.5623738      0.5316275  0.6458931
##   3: 0.007222404    0.006830955 0.008163490 0.5581381      0.5278874  0.6308640
##   4: 0.004611304    0.004323963 0.005097580 0.6863721      0.6436027  0.7587521
##   5: 0.001937927    0.001600133 0.002171618 0.5682357      0.4691883  0.6367582
##  ---                                                                           
## 100: 0.007225256    0.006450020 0.008273156 0.5600855      0.4999910  0.6413163
## 101: 0.019413627    0.015860893 0.023754558 0.5084740      0.4154222  0.6221699
## 102: 0.048722437    0.051373864 0.063538226 0.7563756      0.7975368  0.9863785
## 103: 0.019516104    0.017677122 0.021691778 0.6356980      0.5757968  0.7065662
## 104: 0.021782360    0.018920362 0.026781986 0.7044834      0.6119209  0.8661809

Knowing how each beer style fluctuates on our variables of interest (with our overall score thrown in for good measure!), let’s plot our results. Note that I have included standard error of the mean error bars as a sanity check to make sure that each beer’s ratings is not going to bleed into the others’ dimensions. By doing this, we can have a bit more confidence that we are looking at actually has some meaning. This plot shows the standard error on each of the two variables we are interested in, and for the most part it looks as if they are relatively well contained.

ggplot(question4.means, aes(x = mean_aroma, y = mean_appearance, color = beer_style)) + 
  geom_point() +
  geom_errorbar(aes(ymin=mean_appearance-sem_appearance, ymax=mean_appearance+sem_appearance), width=.1) +
  geom_errorbarh(aes(xmin=mean_aroma-sem_aroma, xmax=mean_aroma+sem_aroma)) + theme(legend.position="none") +
  labs(title = "Mean Appearance and Aroma", y = "Mean Aroma", x = "Mean Appearance") 

This graph has a lot of beers, but what we are interested in is that top right quadrant where both the average appearance and aroma are maxed out. Let’s zoom in on it and throw in a sizing variable of the overall rating and inspect our graph.

ggplot(question4.means[mean_appearance > 4 & mean_aroma > 4], 
       aes(x = mean_aroma, y = mean_appearance, color = beer_style, size = mean_overall)) + 
  geom_point() + xlim(4,4.2) + ylim(4,4.25) +  theme(legend.position="none") +
 # geom_errorbar(aes(ymin=mean_appearance-sem_appearance, ymax=mean_appearance+sem_appearance), width=.1) +
#  geom_errorbarh(aes(xmin=mean_aroma-sem_aroma, xmax=mean_aroma+sem_aroma)) + theme(legend.position="none") +
  labs(title = "Mean Appearance and Aroma", y = "Mean Aroma", x = "Mean Appearance") + 
  geom_text(aes(label=beer_style, hjust = .5, vjust = -.75)) 

Looking at this subsection, it appears we have a few choices for beers that score highly on both Appearance and Aroma. The American Double / Imperial Stout looks like a good option as it scores higher on how it looks, but our Russian Imperial Stout has a higher Aroma score. We could start crunching more numbers here to find “the best” option, or at this point we could take off our data science hats and put our psychology ones back on (assuming that’s what we were wearing…) and run some double-blind experiments on ourselves to make sure that our data is actually lining up with something in the real world.