Pandas DataFrames

Overview

Teaching: 60 min
Exercises: 20 min

Questions

• How can I do statistical analysis of tabular data?

Objectives

• Select individual values from a Pandas dataframe.

• Select entire rows or entire columns from a dataframe.

• Select a subset of both rows and columns from a dataframe in a single operation.

• Select a subset of a dataframe by a single Boolean criterion.

Note about Pandas DataFrames/Series

A DataFrame is a collection of Series; The DataFrame is the way Pandas represents a table, and Series is the data-structure Pandas use to represent a column.

Pandas is built on top of the Numpy library, which in practice means that most of the methods defined for Numpy Arrays apply to Pandas Series/DataFrames.

What makes Pandas so attractive is the powerful interface to access individual records of the table, proper handling of missing values, and relational-databases operations between DataFrames.

If you don’t have the data files already (e.g., using Google Colab), you will want to download them.

# Same download function as last time

import urllib.request
import os

def download_data(filename):
  if not os.path.exists('data'):
          os.mkdir('data')
  url = 'https://raw.githubusercontent.com/ualberta-rcg/python-intro/gh-pages/data/' + filename
  output_file = 'data/' + filename
  urllib.request.urlretrieve(url, output_file)
  print("Downloaded " + filename + " to the data directory")

# Run the function to download

filenames = ['gapminder_gdp_europe.csv',
             'gapminder_all.csv']

for filename in filenames:
  download_data(filename)

Selecting values

To access a value at the position [i,j] of a DataFrame, we have two options, depending on what is the meaning of i in use. Remember that a DataFrame provides a index as a way to identify the rows of the table; a row, then, has a position inside the table as well as a label, which uniquely identifies its entry in the DataFrame.

Use DataFrame.iloc[..., ...] to select values by their (entry) position

• Can specify location by numerical index analogously to 2D version of character selection in strings.

import pandas as pd
europe_df = pd.read_csv('data/gapminder_gdp_europe.csv', index_col='country')
print(europe_df.iloc[0, 0])

1601.056136

Use DataFrame.loc[..., ...] to select values by their (entry) label.

• Can specify location by row name analogously to 2D version of dictionary keys.

europe_df = pd.read_csv('data/gapminder_gdp_europe.csv', index_col='country')
print(europe_df.loc["Albania", "gdpPercap_1952"])

1601.056136

Use : on its own to mean all columns or all rows.

• Just like Python’s usual slicing notation.

print(europe_df.loc["Albania", :])

gdpPercap_1952    1601.056136
gdpPercap_1957    1942.284244
gdpPercap_1962    2312.888958
gdpPercap_1967    2760.196931
gdpPercap_1972    3313.422188
gdpPercap_1977    3533.003910
gdpPercap_1982    3630.880722
gdpPercap_1987    3738.932735
gdpPercap_1992    2497.437901
gdpPercap_1997    3193.054604
gdpPercap_2002    4604.211737
gdpPercap_2007    5937.029526
Name: Albania, dtype: float64

• Would get the same result printing europe_df.loc["Albania"] (without a second index).

print(europe_df.loc[:, "gdpPercap_1952"])

country
Albania                    1601.056136
Austria                    6137.076492
Belgium                    8343.105127
⋮ ⋮ ⋮
Switzerland               14734.232750
Turkey                     1969.100980
United Kingdom             9979.508487
Name: gdpPercap_1952, dtype: float64

• Would get the same result printing europe_df["gdpPercap_1952"]
• Also get the same result printing europe_df.gdpPercap_1952 (since it’s a column name)

Select multiple columns or rows using DataFrame.loc and a named slice.

print(europe_df.loc['Italy':'Poland', 'gdpPercap_1962':'gdpPercap_1972'])

             gdpPercap_1962  gdpPercap_1967  gdpPercap_1972
country
Italy           8243.582340    10022.401310    12269.273780
Montenegro      4649.593785     5907.850937     7778.414017
Netherlands    12790.849560    15363.251360    18794.745670
Norway         13450.401510    16361.876470    18965.055510
Poland          5338.752143     6557.152776     8006.506993

In the above code, we discover that slicing using loc is inclusive at both ends, which differs from slicing using iloc, where slicing indicates everything up to but not including the final index.

Result of slicing can be used in further operations.

• Usually don’t just print a slice.
• All the statistical operators that work on entire dataframes work the same way on slices.
• E.g., calculate max of a slice.

print(europe_df.loc['Italy':'Poland', 'gdpPercap_1962':'gdpPercap_1972'].max())

gdpPercap_1962    13450.40151
gdpPercap_1967    16361.87647
gdpPercap_1972    18965.05551
dtype: float64

print(europe_df.loc['Italy':'Poland', 'gdpPercap_1962':'gdpPercap_1972'].min())

gdpPercap_1962    4649.593785
gdpPercap_1967    5907.850937
gdpPercap_1972    7778.414017
dtype: float64

Use comparisons to select data based on value.

• Comparison is applied element by element.
• Returns a similarly-shaped dataframe of True and False.

# Use a subset of data to keep output readable.
subset = europe_df.loc['Italy':'Poland', 'gdpPercap_1962':'gdpPercap_1972']
print('Subset of data:\n', subset)

# Which values were greater than 10000 ?
print('\nWhere are values large?\n', subset > 10000)

Subset of data:
             gdpPercap_1962  gdpPercap_1967  gdpPercap_1972
country
Italy           8243.582340    10022.401310    12269.273780
Montenegro      4649.593785     5907.850937     7778.414017
Netherlands    12790.849560    15363.251360    18794.745670
Norway         13450.401510    16361.876470    18965.055510
Poland          5338.752143     6557.152776     8006.506993

Where are values large?
            gdpPercap_1962 gdpPercap_1967 gdpPercap_1972
country
Italy                False           True           True
Montenegro           False          False          False
Netherlands           True           True           True
Norway                True           True           True
Poland               False          False          False

Select values or NaN using a Boolean mask.

• A frame full of Booleans is sometimes called a mask because of how it can be used.

mask = subset > 10000
print(subset[mask])

             gdpPercap_1962  gdpPercap_1967  gdpPercap_1972
country
Italy                   NaN     10022.40131     12269.27378
Montenegro              NaN             NaN             NaN
Netherlands     12790.84956     15363.25136     18794.74567
Norway          13450.40151     16361.87647     18965.05551
Poland                  NaN             NaN             NaN

• Get the value where the mask is true, and NaN (Not a Number) where it is false.
• Useful because NaNs are ignored by operations like max, min, average, etc.

print(subset[subset > 10000].describe())

       gdpPercap_1962  gdpPercap_1967  gdpPercap_1972
count        2.000000        3.000000        3.000000
mean     13120.625535    13915.843047    16676.358320
std        466.373656     3408.589070     3817.597015
min      12790.849560    10022.401310    12269.273780
25%      12955.737547    12692.826335    15532.009725
50%      13120.625535    15363.251360    18794.745670
75%      13285.513523    15862.563915    18879.900590
max      13450.401510    16361.876470    18965.055510

Selection of Individual Values

Assume Pandas has been imported into your notebook and the Gapminder GDP data for Europe has been loaded:

import pandas as pd

europe_df = pd.read_csv('data/gapminder_gdp_europe.csv', index_col='country')

Write an expression to find the Per Capita GDP of Serbia in 2007.

Solution

The selection can be done by using the labels for both the row (“Serbia”) and the column (“gdpPercap_2007”):

print(europe_df.loc['Serbia', 'gdpPercap_2007'])

The output is

9786.534714

Extent of Slicing

1. Do the two statements below produce the same output?
2. Based on this, what rule governs what is included (or not) in numerical slices and named slices in Pandas?

print(europe_df.iloc[0:2, 0:2])
print(europe_df.loc['Albania':'Belgium', 'gdpPercap_1952':'gdpPercap_1962'])

Solution

No, they do not produce the same output! The output of the first statement is:

        gdpPercap_1952  gdpPercap_1957
country                                
Albania     1601.056136     1942.284244
Austria     6137.076492     8842.598030

The second statement gives:

        gdpPercap_1952  gdpPercap_1957  gdpPercap_1962
country                                                
Albania     1601.056136     1942.284244     2312.888958
Austria     6137.076492     8842.598030    10750.721110
Belgium     8343.105127     9714.960623    10991.206760

Clearly, the second statement produces an additional column and an additional row compared to the first statement.
What conclusion can we draw? We see that a numerical slice, 0:2, omits the final index (i.e. index 2) in the range provided, while a named slice, ‘gdpPercap_1952’:’gdpPercap_1962’, includes the final element.

Reconstructing Data

Explain what each line in the following short program does: what is in first, second, etc.?

first_df = pd.read_csv('data/gapminder_all.csv', index_col='country')
second_df = first_df[first_df['continent'] == 'Americas']
third_df = second_df.drop('Puerto Rico')
fourth_df = third_df.drop('continent', axis = 1)
fourth_df.to_csv('result.csv')

Solution

Let’s go through this piece of code line by line.

first_df = pd.read_csv('data/gapminder_all.csv', index_col='country')

This line loads the dataset containing the GDP data from all countries into a dataframe called first_df. The index_col='country' parameter selects which column to use as the row labels in the dataframe.

second_df = first_df[first_df['continent'] == 'Americas']

This line makes a selection: only those rows of first_df for which the ‘continent’ column matches ‘Americas’ are extracted. Notice how the Boolean expression inside the brackets, first_df['continent'] == 'Americas', is used to select only those rows where the expression is true. Try printing this expression! Can you print also its individual True/False elements? (hint: first assign the expression to a variable)

third_df = second_df.drop('Puerto Rico')

As the syntax suggests, this line drops the row from second_df where the label is ‘Puerto Rico’. The resulting dataframe third_df has one row less than the original dataframe second_df.

fourth_df = third_df.drop('continent', axis = 1)

Again we apply the drop function, but in this case we are dropping not a row but a whole column. To accomplish this, we need to specify also the axis parameter (we want to drop the second column which has index 1).

fourth_df.to_csv('result.csv')

The final step is to write the data that we have been working on to a csv file. Pandas makes this easy with the to_csv() function. The only required argument to the function is the filename. Note that the file will be written in the directory from which you started the Jupyter or Python session.

Selecting Indices

Explain in simple terms what idxmin and idxmax do in the short program below. When would you use these methods?

europe_df = pd.read_csv('data/gapminder_gdp_europe.csv', index_col='country')
print(europe_df.idxmin())
print(europe_df.idxmax())

Solution

For each column in europe_df, idxmin will return the index value corresponding to each column’s minimum; idxmax will do accordingly the same for each column’s maximum value.

You can use these functions whenever you want to get the row index of the minimum/maximum value and not the actual minimum/maximum value.

Practice with Selection

Assume Pandas has been imported and the Gapminder GDP data for Europe has been loaded. Write an expression to select each of the following:

1. GDP per capita for all countries in 1982.
2. GDP per capita for Denmark for all years.
3. GDP per capita for all countries for years after 1985.
4. GDP per capita for each country in 2007 as a multiple of GDP per capita for that country in 1952.

Solution

1:

europe_df['gdpPercap_1982']

2:

europe_df.loc['Denmark',:]

3:

europe_df.loc[:,'gdpPercap_1985':]

Pandas is smart enough to recognize the number at the end of the column label and does not give you an error, although no column named gdpPercap_1985 actually exists. This is useful if new columns are added to the CSV file later.

4:

europe_df['gdpPercap_2007'] / europe_df['gdpPercap_1952']

Using the dir function to see available methods

Python includes a dir function that can be used to display all of the available methods (functions) that are built into a data object. As an example, the functions available for a list data type are:

potatoes = ["Russet", "Norkota", "Yukon Gold", "Pontiac"]
dir(potatoes)

This command returns:

['__add__',
...
'__subclasshook__',
 'append',
 'clear',
 'copy',
 'count',
'extend',
'index',
'insert',
'pop',
'remove',
'reverse',
'sort']

The double underscore functions can be ignored for now; functions that are not surrounded by double underscores are the public interface of the list type. So, if you want to sort the list of potatoes, according to dir you should try,

potatoes.sort()

Assume Pandas has been imported and the Gapminder GDP data for Europe has been loaded as europe_df. Then, use dir to find the function that prints out the median per-capita GDP across all European countries for each year that information is available.

Solution

Among many choices, dir lists the median() function as a possibility. Thus,

europe_df.median()

Interpretation

Poland’s borders have been stable since 1945, but changed several times in the years before then. How would you handle this if you were creating a table of GDP per capita for Poland for the entire twentieth century?

(Warning: more advanced!) Select-Apply-Combine operations

Pandas vectorizing methods and grouping operations are features that provide users much flexibility to analyse their data.

For instance, let’s say we want to have a clearer view on how the European countries split themselves according to their GDP.

1. We may have a glance by splitting the countries in two groups during the years surveyed, those who presented a GDP higher than the European average and those with a lower GDP.
2. We then estimate a wealthy score based on the historical (from 1962 to 2007) values, where we account how many times a country has participated in the groups of lower or higher GDP

mask_higher = europe_df.apply(lambda x:x>x.mean())
wealth_score = mask_higher.aggregate('sum',axis=1)/len(europe_df.columns)
wealth_score

country
Albania                   0.000000
Austria                   1.000000
Belgium                   1.000000
Bosnia and Herzegovina    0.000000
Bulgaria                  0.000000
Croatia                   0.000000
Czech Republic            0.500000
Denmark                   1.000000
Finland                   1.000000
France                    1.000000
Germany                   1.000000
Greece                    0.333333
Hungary                   0.000000
Iceland                   1.000000
Ireland                   0.333333
Italy                     0.500000
Montenegro                0.000000
Netherlands               1.000000
Norway                    1.000000
Poland                    0.000000
Portugal                  0.000000
Romania                   0.000000
Serbia                    0.000000
Slovak Republic           0.000000
Slovenia                  0.333333
Spain                     0.333333
Sweden                    1.000000
Switzerland               1.000000
Turkey                    0.000000
United Kingdom            1.000000
dtype: float64

Finally, for each group in the wealth_score table, we sum their (financial) contribution across the years surveyed:

europe_df.groupby(wealth_score).sum()

          gdpPercap_1952  gdpPercap_1957  gdpPercap_1962  gdpPercap_1967  \
0.000000    36916.854200    46110.918793    56850.065437    71324.848786   
0.333333    16790.046878    20942.456800    25744.935321    33567.667670   
0.500000    11807.544405    14505.000150    18380.449470    21421.846200   
1.000000   104317.277560   127332.008735   149989.154201   178000.350040   

          gdpPercap_1972  gdpPercap_1977  gdpPercap_1982  gdpPercap_1987  \
0.000000    88569.346898   104459.358438   113553.768507   119649.599409   
0.333333    45277.839976    53860.456750    59679.634020    64436.912960   
0.500000    25377.727380    29056.145370    31914.712050    35517.678220   
1.000000   215162.343140   241143.412730   263388.781960   296825.131210   

          gdpPercap_1992  gdpPercap_1997  gdpPercap_2002  gdpPercap_2007  
0.000000    92380.047256   103772.937598   118590.929863   149577.357928  
0.333333    67918.093220    80876.051580   102086.795210   122803.729520  
0.500000    36310.666080    40723.538700    45564.308390    51403.028210  
1.000000   315238.235970   346930.926170   385109.939210   427850.333420

Key Points

• Use DataFrame.iloc[..., ...] to select values by integer location.

• Use : on its own to mean all columns or all rows.

• Select multiple columns or rows using DataFrame.loc and a named slice.

• Result of slicing can be used in further operations.

• Use comparisons to select data based on value.

• Select values or NaN using a Boolean mask.