1.1 Measuring “Unhealthiness”

The variables that are used to measure “healthiness” are:

• Total fat
• Total sugar
• Total caffeine

An unhealthy index (referred to as unhealthy_index in the dataset) was constructed to combine the effects of the 3 variables to signify how “unhealthy” a beverage was. Therefore - as common sense also alludes to - the more fat, sugar and caffeine a beverage contains, the more unhealthy it is.

With reference to Table 1. below, the following steps were taken to construct the unhealthy_index in Table 2.:

1. total_fat_g, sugar_g and caffeine_mg are expressed as a proportion of the serving size of the beverage (serv_size_m_l)

2. unhealthy_index is defined as the sum of the proportions. The range of this value is \(0 < unhealthy{\_}index < 3\). It is expected that \(unhealthy{\_}index << 3\) as caffeine is in units of milligrams and it is highly improbable that the proportions of all these variables will be 1.

These quantities are included in Table 2. and are highlighted in grey.

1.2 Handling Inconsistent/Missing Data

It was noticed that espresso beverages had to be handled as they contained differing values for size which may have contributed to their serving sizes being recorded as 0 ml. Size in the context of espressos are considered as the number of shots, where each shot is 30 ml (referenced from here). Thus, serv_size_m_l for espressos was mutated accordingly, from this information.

1.3 Filtering Relevant Data

It was noticed that there were 3 observations in the dataset which were not beverages and hence, irrelevant for my analysis. These were: - Protein & Fibre Powder - Matcha Green Tea Powder - Espresso shot

They also had no serving size values and a different metric for measuring size (“scoop”/“shot”), thus, they were filtered out of the dataset.

Furthermore, other variables which were also irrelevant to the analysis were filtered out, such as:

• trans_fat_g
• saturated_fat_g
• cholesterol_mg
• sodium_mg
• fiber_g

1.4 Modifying Variables’ Data Types

The milk variable - which is a categorical variable - took on values from 0-5. This variable’s values were changed to include the type of milk in words to be more descriptive. The data type was also changed from a double to a factor. This aided the process of data modelling (in part ), given we want the milk regressor to be a categorical dummy rather than a continuous variable.

Similarly, the whip variable - which is also a binary categorical dummy - was changed from a double data type to a factor data type.

Finally, table 2. shows the final result after data cleaning and wrangling. The observations are arranged in order of descending unhealthy_index.