Primary Objective:

• Calculate the price-elasticity of alcoholic beverages.

Secondary Objectives:

• Compare price-elasticity between spirits and beer.

• Compare price-elasticity between quantiles based on price.

• Data mining of two data sets: Monthly sales and Retail price for the state of Karnataka.
  

• Time Period: March 2011 to March 2022.
  

• Source: Karntaka State Beverage Corporation Limited (KSBCL).
  ! ##

• The data is in PDF files - requiring extraction of tables.

• The item names are mentioned in a different way every month.
  

• Further fragmentation due to inclusion of packaging information in the item name. See example:

#Show all Teacher's Highland Cream Scotch Whisky
unique(temp$item[str_detect(temp$item,
regex("Teacher'?s Highland Cream Scotch Whisky",ignore_case = T))])

## [1] "Teachers Highland Cream Scotch Whisky 375 Ml (0134)"       
## [2] "Teacher's Highland Cream Scotch Whisky 180ML(0134)"        
## [3] "Teachers Highland Cream Scotch Whisky 60 Ml (0134)"        
## [4] "Teachers Highland Cream Scotch Whisky 750 Ml (0134)"       
## [5] "Teachers Highland Cream Scotch Whisky 1000 Ml (0134)"      
## [6] "Teacher's Highland Cream Scotch Whisky 1000MLx12Btls(0134)"

Steps:
- Extract beverage name.

• Extract packaging volume, month-year and place of sale. 

• Collapse items which are same but spelt differently.
  Tools Used:
  

• ‘R’ platform with stringR, pdfTools packages.
  

• Deep learning model - Sentence Transformers based on BERT, to calculate similarity between names and manually collapse them.

Output:

#Number of initial items: 
nrow(temp)

## [1] 3742095

#Number of items with cleaning but no collapsing:  
length(unique(temp$itemName))

## [1] 14471

#Number of items with cleaning and collapsing:
length(unique(temp$itemNameNew))

## [1] 2137

• Four team members coded alcoholic beverages in to ‘beer’,‘wine’ and ‘spirit’ category.
  
• Essential step as they differ in alcohol concentration.
  
• Due to almost negligible contribution of wine to alcohol sales this category has not been included in the analyses.

Overall approach: Linear Mixed Effects Models to explain sales of a particular beverage as a function of prevalent price, controlling for confounding variables.

• All models have ‘time’ and ‘monthly seasonality’ as predictors (confounding variables).

• All inference is based on bootstrapped confidence intervals.

• All models are log-log i.e. sales and price both are on log scale.

• Random effects are chosen based on the question we wish to answer:

What is the overall price-elasticity?

• Random Intercept for beverages.

Is there variation between districts?

• Random slopes of price for district.

Is there a difference between beer and spirits?

• Random slopes of price for beverage type.

Is there a difference based on how costly is a standard drink?

• Random slopes of price for each quantile (based on average price).

Implications for price-elasticity.
The base price is ‘seemingly low’ and increase does not have sufficient dis-incentivising effect.

What is the overall Price Elasticity of Alcoholic Beverages

## totalMlsSoldLog ~ (1 | itemNameNew) + district + lockdownEffect + 
##     month + linerTime + effectivePriceLog

## # Indices of model performance
## 
## AIC       |      AICc |       BIC | R2 (cond.) | R2 (marg.) |   ICC |  RMSE | Sigma
## -----------------------------------------------------------------------------------
## 2.407e+06 | 2.407e+06 | 2.408e+06 |      0.759 |      0.361 | 0.623 | 1.428 | 1.431

The model is a decent fit explaining 76% of variability in sales

## # Fixed Effects
## 
## Parameter         | Coefficient |       SE |         95% CI | t(674438) |      p
## --------------------------------------------------------------------------------
## (Intercept)       |        9.67 |     0.04 | [ 9.58,  9.76] |    220.89 | < .001
## lockdownEffect    |       -0.76 |     0.05 | [-0.87, -0.66] |    -14.57 | < .001
## linerTime         |   -2.91e-03 | 6.59e-05 | [ 0.00,  0.00] |    -44.22 | < .001
## effectivePriceLog |       -1.61 |     0.03 | [-1.66, -1.55] |    -57.23 | < .001

Interpretation: For a 1% increase in price, sales decrease by 1.6% after controlling for seasonality, time effect, lockdown and allowing itemwise variation in expected sales as well as variation between districts (coefs not shown).
## Comparison between Districts

• To be done