Food Nutrition and Price

Explore a supermarket of food nutrition and price (renders slowly)

After feasting, thoughts turn to diets. Diets mean optimisation. Optimisation requires data. For myself, I want to know:

• What is the cheapest protein, fat and carbohydrate per 100g?
• What is the general distribution of nutritional value?
• What is the sweeter condiment: sauce or jam?

So I scraped a supermarket.

Specifically, I scraped the grocery section of the website of a major UK supermarket, and then I sat on the data for nearly two years because it’s so untidy.

Getting nutritional information in bulk

There are strict labelling requirements on food, its nutritional value in particular, so almost all food packaging gives carbohydrate, fat and protein content per 100g or 100ml. This should make it straightforward to learn about food, because the data already exists.

Table of nutritional information for Sainsbury’s Brocolli Cheese ready meal

Fun fact: Brocolli Cheese is the food with the most equal balance between fat, protein and carbohydrate. It’s in the middle of the ternary plot below.

Unfortunately, the government doesn’t collect the data, though the Office for National Statistics scraped three supermarked websites for the Consumer Price Index. I tried to do the same and regretted that the ONS hadn’t pubished its code, because it was a pain. A couple of packages have made the task tractable; htmltab is helps with spanning rows and columns html tables, my own unpivotr has improved since 2016, and reticulate makes it trivial to call the Python package quantulum3 for extracting units like grams and kilocalories from free text. The result isn’t perfectly clean data – you’ll notice some obvious glitches – but it’s good for shopping.

This is the least reproducible work I have ever done. If I were to do it again I would use the polite to be nice to the supermarket, and furrr to get it done faster (it ran overnight). I haven’t published the data because I don’t think it belongs to me, but if you really want it then contact me. You could get some of it from this web page anyway, but you won’t get the vitamins and minerals etc. in the raw data.

The cheapest food by nutritional value

The cheapest grocery overall, by mass, was own-brand diet lemonade. All the top ten were soft drinks. But what you’re buying when you buy diet pop is basically water, so it’s far from the cheapest for any kind of nutrition. The cheapest energy and cheapest carbohydrate is flour, with bread and sugar filling out the top ten (wine is probably a mistake). The cheapest fat is lard, with non-butter spread, oil, and mayonnaise (mostly oil) filling out the top ten.

The cheapest protein? Also flour. But you could splash out and get 25kg of chicken portions.

Expand the headings below to see the top 10 in each category.

Price per unit mass

# A tibble: 10 x 10
   product carbohydrate   fat protein energy  price `£/carbohydrate`
   <chr>          <dbl> <dbl>   <dbl>  <dbl>  <dbl>            <dbl>
 1 sainsb…          0       0       0      1 0.02          Inf      
 2 sainsb…          0       0       0     58 0.0285        Inf      
 3 sainsb…         13.7     0       0     57 0.03            0.00219
 4 sainsb…          0       0       0      2 0.03          Inf      
 5 sainsb…          7.5     0       0     30 0.03            0.004  
 6 sainsb…          0       0       0      1 0.03          Inf      
 7 sainsb…          0       0       0      1 0.03          Inf      
 8 sainsb…          0       0       0      1 0.03          Inf      
 9 sainsb…          0       0       0      2 0.03          Inf      
10 sainsb…          0       0       0      1 0.03          Inf      
# … with 3 more variables: `£/fat` <dbl>, `£/protein` <dbl>,
#   `£/energy` <dbl>

Price per unit energy

# A tibble: 10 x 10
   product carbohydrate   fat protein energy  price `£/carbohydrate`
   <chr>          <dbl> <dbl>   <dbl>  <dbl>  <dbl>            <dbl>
 1 sainsb…         71.7   1       9.7    342 0.04           0.000558
 2 sainsb…         69.9   1.1     9.5    335 0.04           0.000572
 3 sainsb…         99.9   0       0      400 0.06           0.000601
 4 sainsb…          0     0       0      400 0.0633       Inf       
 5 sainsb…         71.7   1       9.7    342 0.057          0.000795
 6 sainsb…         45.4   1.9     8.1    238 0.04           0.000881
 7 sainsb…         69.9   1.1     9.5    335 0.057          0.000815
 8 sainsb…         39.7   2.1     9.9    231 0.04           0.00101 
 9 sainsb…         99.9   0       0      400 0.0700         0.000701
10 sainsb…         61.8   2      15.5    340 0.0670         0.00108 
# … with 3 more variables: `£/fat` <dbl>, `£/protein` <dbl>,
#   `£/energy` <dbl>

Price per unit carbohydrate

# A tibble: 10 x 10
   product carbohydrate   fat protein energy  price `£/carbohydrate`
   <chr>          <dbl> <dbl>   <dbl>  <dbl>  <dbl>            <dbl>
 1 sainsb…         71.7   1       9.7    342 0.04           0.000558
 2 sainsb…         69.9   1.1     9.5    335 0.04           0.000572
 3 sainsb…         99.9   0       0      400 0.06           0.000601
 4 sainsb…         99.9   0       0      400 0.0700         0.000701
 5 sainsb…         71.7   1       9.7    342 0.057          0.000795
 6 sainsb…         99.9   0       0      400 0.08           0.000801
 7 sainsb…         69.9   1.1     9.5    335 0.057          0.000815
 8 sainsb…         45.4   1.9     8.1    238 0.04           0.000881
 9 sainsb…         99.9   0       0      400 0.1            0.00100 
10 sainsb…         39.7   2.1     9.9    231 0.04           0.00101 
# … with 3 more variables: `£/fat` <dbl>, `£/protein` <dbl>,
#   `£/energy` <dbl>

Price per unit fat

# A tibble: 10 x 10
   product carbohydrate   fat protein energy price `£/carbohydrate`
   <chr>          <dbl> <dbl>   <dbl>  <dbl> <dbl>            <dbl>
 1 sainsb…          0    99.8     0      898 0.18          Inf     
 2 sainsb…          0    70       0      630 0.153         Inf     
 3 sainsb…          0    70       0      630 0.18          Inf     
 4 sainsb…          3.5  73.2     1      677 0.19            0.0543
 5 sainsb…          3.5  73.2     1      677 0.2             0.0571
 6 sainsb…          0    75       0      675 0.22          Inf     
 7 sainsb…          0    91.9     0      827 0.28          Inf     
 8 sainsb…          3.5  73.2     1      677 0.23            0.0657
 9 sainsb…          3.5  73.2     1      677 0.24            0.0686
10 heinz-…          3    70       0.8    644 0.25            0.0833
# … with 3 more variables: `£/fat` <dbl>, `£/protein` <dbl>,
#   `£/energy` <dbl>

Price per unit protein

# A tibble: 10 x 10
   product carbohydrate   fat protein energy  price `£/carbohydrate`
   <chr>          <dbl> <dbl>   <dbl>  <dbl>  <dbl>            <dbl>
 1 sainsb…         39.7   2.1     9.9    231 0.04           0.00101 
 2 sainsb…         71.7   1       9.7    342 0.04           0.000558
 3 sainsb…         69.9   1.1     9.5    335 0.04           0.000572
 4 sainsb…         61.8   2      15.5    340 0.0670         0.00108 
 5 sainsb…         58.4   2.2    15.9    335 0.073          0.00125 
 6 sainsb…         45.4   1.9     8.1    238 0.04           0.000881
 7 sainsb…         68.2   1.3    13.4    344 0.0700         0.00103 
 8 sainsb…         71.7   1       9.7    342 0.057          0.000795
 9 sainsb…         69.9   1.1     9.5    335 0.057          0.000815
10 sainsb…          0    13.9    24.8    226 0.16         Inf       
# … with 3 more variables: `£/fat` <dbl>, `£/protein` <dbl>,
#   `£/energy` <dbl>

Unsurprisingly, protein is the most expensive nutritional component, then fat, then carbohydrate. But now you can tell when you’re paying over the odds. There’s a lot of choice for less than 10p per gram of protein or fat, or less than 1p per gram of carbohydrate. If you’re paying more than £1 per gram of anything then it isn’t for the nutritional value (unless it’s for something I haven’t explored, such as fibre and vitamins).

Distribution of nutritional value

You can explore an interactive 3d graph at the bottom of the post, but first I want to use ternary plots to compare particular groups of foods. A little light curation before taking the lazy option.

Ternary plots compress three-dimensional data onto two dimensions by exploiting a loophole: when the dimensions are ratios that add to 100%, there are only two degrees of freedom. Once you know the proportions of two dimensions, then you know the third. This doesn’t work for carbohydrate, fat and protein because they don’t necessarily sum to 100% of the mass of the food – diet soda drinks are 0% everything. So to make the food fit the graph, I create fake ratios: carbohydrate divided by (carbohydrate plus fat plus sugar), and so on. When reading the graphs below, a point that is near to the “carbohydrate” corner has more carbohydrate than fat or protein, but it doesn’t necessarily have much of any of them – it could be mostly water. Finally, the colour is the amount of energy, with turquoise for lots of energy and purple for not much (or missing data). Mouse over for the names of the products.

You might wonder about a data glitch in that plot. Why do there seem to be so many foods that despite having proportions of carbohydrate don’t have much energy? That’s because the purple foods on the right are comparitively high in water content – meat, bread, soup, cooked rice and pasta – whereas the turquoise foods on the left are more substantial – biscuits, crisps, sticky cake. Of course there is probably some missing data too.

Ternary plots are good for comparing groups of foods, and here they reveal some surprises. Crisps and biscuits are roughly equivalent, nutritionally. Some crisps are fattier, and some biscuits are more sugary, but there is a substantial overlap.

You might pay closer attention to food labelling when you compare sauces with jams. Fair enough, most of the jams are overplotted in the carbohydrate corner, but lots of sauces are terribly sugary.

The data distinguishes between carbohydrates and “of which, sugar”, but I haven’t here.

‘Lite’ yoghurt is not so much less fatty as less carbohydrate-y than normal yoghurt.

Free exploration

It’s your turn to explore the data. Perhaps you’d like to check that my findings above are still valid as absolutes rather than proportions. Or perhaps you’ve resolved to bulk up and want to sort the data by protein content (egg white powder, then dried meat). The highest energy content? Peppermint extract is in sixth place, being almost 100% fat. You can filter the table, e.g. to find the lowest-sugar sugar (molasses, 94%), and click on rows (hold down shift to select a range) to filter the charts as well. Finally, select a region of the 2d graphs to filter the other graphs and the table. Be patient – the browser has a lot of data to deal with.

The interactivity within graphs plotly and the interactivity between graphs and the table is crosstalk. Everything runs in the browser; there is no server. Matt Dray’s blog post and slides are persuasive and fun.