The most successful joke among statisticians (a low bar, I know) is that we spend 80 percent of our time cleaning data, and the other 20 percent complaining about cleaning data.The joke resonates with analysts because it rings true. We do spend a nontrivial amount of time going through the busy work of data munging, when we’d much rather be actually analyzing the data. When doing data analysis, you’ll often have assumptions that should be true but might not be.
For example, every person has one (and only one) email address for your service. Getting these assumptions wrong can potentially bias your analysis, so it’s good to check.
However, the process of writing and checking those assumptions is tedious.
Read how to make D&A Part of Your DNA!
Enter Engarde
To streamline the process of reading data, checking assumptions, applying transformations, and back to checking assumptions, I’ve written a simple Python library called Engarde, which builds on top of pandas.
For this post, we’ll work with a data set of customer preferences on trains, which I’ve saved to data/trains.csv.
We can start by making some very basic assertions – that the dataset is the correct shape and that a few columns are the correct data types. Assertions are made as decorators to functions that return a DataFrame.
In [1]: import pandas as pd
In [2]: import engarde.decorators as ed
In [3]: pd.set_option('display.max_rows', 10)
In [4]: dtypes = dict(
...: price1=int,
...: price2=int,
...: time1=int,
...: time2=int,
...: change1=int,
...: change2=int,
...: )
In [5]: @ed.is_shape((None, 11))
...: @ed.has_dtypes(items=dtypes)
...: def unload():
...: trains = pd.read_csv("data/trains.csv", index_col=0)
...: return trains
In [6]: unload()
Out[6]:
id choiceid choice price1 time1 change1 comfort1 price2 time2 change2 comfort2
1 1 1 choice1 2400 150 0 1 4000 150 0 1
2 1 2 choice1 2400 150 0 1 3200 130 0 1
3 1 3 choice1 2400 115 0 1 4000 115 0 0
4 1 4 choice2 4000 130 0 1 3200 150 0 0
5 1 5 choice2 2400 150 0 1 3200 150 0 0
.. .. ... ... ... ... ... ... ... ... ... ...
347 30 7 choice1 2100 135 1 1 2800 135 1 0
348 30 8 choice1 2100 125 1 1 3500 125 1 0
349 30 9 choice1 2100 150 0 0 2800 125 0 1
350 30 10 choice1 2800 125 0 1 2800 135 1 0
351 30 11 choice2 3500 125 1 0 2800 135 1 0
[351 rows x 11 columns]One very important part of the design of Engarde is that your code, the code actually doing the work, shouldn’t have to change. I don’t want a bunch of asserts cluttering up the logic of what’s happening. This is a perfect case for decorators.
The order of execution here is unload returns the DataFrame, trains. Next, ed.has_dtypes asserts that trains has the correct dtypes as specified with dtypes. Once that assert passes, has_dtypes passes trains along to the next check, and so on, until the original caller gets back trains.
Each row of this dataset contains a passenger’s preference over two routes. Each route has an associated cost, travel time, comfort level, and number of changes.
Like any good economist, we’ll assume people are rational. Their first choice is surely going to be better in at least one way than their second choice (faster, more comfortable, etc).
This is fundamental to our analysis later on, so we’ll explicitly state it in our code, and check it in our data analysis.
We write our custom assumption as a function rational that takes a DataFrame and returns one or more boolean (true / false) values.
In [7]: def rational(df):
...: """
...: Check that at least one criteria is better.
...: """
...: r = ((df.price1 < df.price2) | (df.time1 < df.time2) |
...: (df.change1 < df.change2) | (df.comfort1 > df.comfort2))
...: return r
...:
In [8]: @ed.is_shape((None, 11))
...: @ed.has_dtypes(items=dtypes)
...: @ed.verify_all(rational)
...: def unload():
...: trains = pd.read_csv("data/trains.csv", index_col=0)
...: return trains
...:
In [9]: df = unload()
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-9-b108f050ce4e> in <module>()
----> 1 df = unload()
...
AssertionError: ('rational not true for all',
id choiceid choice price1 time1 change1 comfort1 price2 time2 \
13 2 3 choice2 2450 121 0 0 2450 93
18 2 8 choice2 2975 108 0 0 2450 108
27 3 6 choice2 1920 106 0 0 1440 96
28 3 7 choice1 1920 106 0 0 1920 96
33 4 1 choice2 545 105 1 1 545 85
.. .. ... ... ... ... ... ... ... ...
306 27 9 choice1 3920 140 1 1 3920 125
319 28 8 choice2 2450 133 1 1 2450 108
325 28 14 choice2 2450 123 0 1 2450 108
[42 rows x 11 columns])So, our check failed. Apparently people aren’t rational.
Engarde has printed the name of the failed assertion, and the rows that violated our assumption. We’ll “fix” this problem by ignoring those people. Although, in reality, we’d dig into why those people are different.
In [16]: @ed.verify_all(rational)
....: def drop_nonrational_people(df):
....: r = df.query("price1 < price2 | time1 < time2 |"
....: "change1 < change2 | comfort1 > comfort2")
....: return r
....:
In [17]: @ed.is_shape((None, 11))
....: @ed.has_dtypes(items=dtypes)
....: def unload():
....: trains = pd.read_csv("data/trains.csv", index_col=0)
....: return trains
In [18]: df = unload().pipe(drop_nonrational_people)
In [19]: df.head()
Out[19]:
id choiceid choice price1 time1 change1 comfort1 price2 time2 change2 comfort2
1 1 1 choice1 2400 150 0 1 4000 150 0 1
2 1 2 choice1 2400 150 0 1 3200 130 0 1
3 1 3 choice1 2400 115 0 1 4000 115 0 0
4 1 4 choice2 4000 130 0 1 3200 150 0 0
5 1 5 choice2 2400 150 0 1 3200 150 0 0All of our assertions have now “passed,” so we’re happy, and our data analysis can proceed.
By using (and reusing) simple libraries like Engarde, we were able to do the data analysis efficiently while maintaining correctness.
– Tom Augspurger, Lead Data Analysis Scientist at MITTERA
