Advanced Pandas: Cleaning data

Probably the best thing about Pandas is its extensive toolset for managing data. Here we describe features of Pandas that allow us to clean data that, for reasons beyond our control, comes in a form that's not immediately amendable to analysis. This is the first of several such notebooks.

Outline:

• Want operator. Start with what we want to end up, then figure out how to get there.
• String methods. Fixing string variables, especially strings that should really be numbers.
• Missing values. Marking, dropping, counting missing values.
• Selecting variables and observations. Choose the variables and observations we want by their labels.
• Boolean selection. This is mostly what we do: choose observations from conditions. We use comparisons to produce Boolean variables and then use the Boolean variables to select observations that are True. The next two methods extend this capability.
• The isin method. Choose observations whose values are in lists you specify.
• The contains method. Flag observations that contain a specific piece of text. Another string method, operates through Booleans.

Note: requires internet access to run.

This Jupyter notebook was created by Dave Backus, Chase Coleman, Spencer Lyon and Balint Szoke for the NYU Stern course Data Bootcamp.

Preliminaries

The want operator

We need to know what we're trying to do -- what we want the data to look like. To borrow a phrase from our friend Tom Sargent, we say that we apply the want operator.

Some problems we've run across that ask to be solved:

• Numerical data is contaminated by commas (marking thousands) or dollar signs.
• Row and column labels are contaminated.
• Missing values are marked erratically.
• We have too much data, would prefer to choose a subset.
• Variables run across rows rather than down columns.

What we want in each case is the opposite of what we have: we want nicely formatted numbers, clean row and column labels, and so on.

We'll solve the first four problems here, the last one in the next notebook.

Example: Chipotle data

This data comes from a New York Times story about the number of calories in a typical order at Chipotle. The topic doesn't particularly excite us, but the data raises a number of issues that come up repeatedly. We adapt some code written by Daniel Forsyth.

Note: The file is a tsv (Tab Separated Values) file, so we need to set the separator accordingly when we call pandas' read_csv method. Remember that the default value of sep is sep=',' (see the docstring). We can change it to tabular by wrinting sep='\t'.

Comment. Note that the variable item_price has dtype object. The reason is evidently the dollar sign. We want to have it as a number, specifically a float.

Example: Data Bootcamp entry poll

This is the poll we did at the start of the course. Responses were collected in a Google spreadsheet, which we converted to a csv and uploaded to our website.

Comments. This is mostly text data, which means it's assigned the dtype object. There are two things that would make the data easier to work with:

First: The column names are excessively verbose. This one's easy: We replace them with single words. Which we do below.

str.title() returns a copy of the string in which first characters of all the words are capitalized.

Second: The second one is harder. The question about special topics of interest says "mark all that apply." In the spreadsheet, we have a list of every choice the person checked. Our want is to count the number of each type of response. For example, we might want a bar chart that gives us the number of each response. The question is how we get there.

Comment. Note the commas separating answers with more than one choice. We want to unpack them somehow.

Example: OECD healthcare statistics

The OECD collects healthcare data on lots of (mostly rich) countries, which is helpful in producing comparisons. Here we use a spreadsheet that can be found under Frequently Requested Data.

This looks bad. But we can always use pd.read_excel?. Let's look into the excel file.

• multiple sheets (want: Physicians)

• The first three lines are empty. Skip those

• Would be nice to have the countries as indices

• The last two columns contain junk

• What about the bottom of the table?

We still have a couple issues.

• The index includes a space and a number: Australia 1, Chile 3, etc. We care about this because when we plot the data across countries, the country labels are going to be country names, so we want them in a better form than this.
• The ..'s in the sheet lead us to label any column that includes them as dtype object. Here we want to label them as missing values.
• If we want to plot each country against time, then we'll need to switch the rows and columns somehow, so that the x axis in the plot (the year) is the index and not the column label.

Example: World Economic Outlook

The IMF's World Economic Outlook database contains a broad range of macroeconomic data for a large number of countries. It's updated twice a year and is a go-to source for things like current account balances (roughly, the trade balance) and government debt and deficits. It also has a few quirks, as we'll see.

Example. Run the following code as is, and with the thousands and na_values parameters commented out. How do the dtypes differ?

Useful columns:

• 1, 2, 3, 4, 6 (indices)
• years, say from 1980 to 2016

Need a list that specifies these

• Look at the bottom

• Missing values

• Notice the , for thousands. As we saw before, there is an easy fix

Comment. This has several issues. Here's what we want:

• We would like the variables to be in columns and have observations labeled 1980, 1981, etc. In other words, we want the years in the index.
• We would like to make sure the data columns (1980, 1981, etc.) have dtype float64

Here's what we might to do achieve what we want:

• We could try to transpose the dataframe to get the years in the index. This is close, but not quite what we want (see below). We'll come back to this in the next notebook.
• To make the data columns have dtype float64 we need to warn pandas about the , thousand separators and n/a for missing data.

Question. Can we transpose the whole thing to get the data running down columns?

String methods

We can treat variables as strings in Pandas in much the same way we dealt with strings in core Python. Run the code below to remind yourself how this works.

Pandas string methods. We can do the same thing to all the observations of a variable with so-called string methods. We append .str to a variable in a dataframe and then apply the string method of our choice. If this is part of converting a number-like entry that has mistakenly been given dtype object, we then convert its dtype with the astype method.

Example. Let's use a string method to fix the item_price variable in the Chipotle dataframe. This has three parts:

• Use the method str to identify this as a string method.
• Apply the string method of our choice (here replace) to fix the string.
• Use the astype method to convert the fixed-up string to a float.

We start by making a copy of the chp dataframe that we can experiment with.

Comment. We did everything here in one line: replace the dollar sign with a string method, then converted to float using astype. If you think this is too dense, you might break it into two steps.

Example. Here we use the astype method again to convert the dtypes of weo into float

Want to convert the year variables into float

This error indicates that somewhere in weo['1980'] there is a string value --. We want to convert that into NaN. Later we will see how we can do that directly. For now use read_csv() again

Example. Here we strip off the numbers at the end of the indexes in the OECD docs dataframe. This involves some experimentation:

• Play with the rsplit method to see how it works.
• Apply rsplit to the example country = 'United States 1'.
• Use a string method to do this to all the entries of the variable Country.

Comment. Not quite, we only want to split once.

Comments.

• Note that we need two str's here: one to do the split, the other to extract the first element.
• For reasons that mystify us, we ran into problems when we used maxsplit=1, but it works with n=1.
• This is probably more than you want to know, but file away the possibilities in case you need them.

Missing values

It's important to label missing values, so that Pandas doesn't interpret entries as strings. Pandas is also smart enough to ignore things labeled missing when it does calculations or graphs. If we compute, for example, the mean of a variable, the default is to ignore missing values.

We've seen that we can label certain entries as missing values in read statements: read_csv, read_excel, and so on. Here we do it directly, mostly to remind ourselves what's involved.

Marking missing values

Example. The oecd dataframe contains a number of instances of .. (double period). How can we mark them as missing values?

What to do. We use the replace method on the whole dataframe. To mark something as missing, we replace it as None, which Pandas interprets as missing and labels NaN.

Comment. Replace automatically updates the dtypes. Here the double dots led us to label the variables as objects. After the replace, they're now floats, as they should be.

Comment. Unlike the string methods we described earlier, this use of replace affects complete entries, not elements of string entries. For example, suppose we tried to replace the periods in decimal numbers with an asterisk. We could try the following, but it doesn't work: the decimal numbers don't change.

Working with missing values

Comment. We usually don't have to worry about this, Pandas takes care of missing values automatically.

Comment. Let's try a picture to give us a feeling of accomplishment. What else would you say we need? How would we get it?

Selecting variables and observations

The word selection refers to choosing a subset of variables or observations using their labels or index. Similar methods are sometimes referred to as slicing, subsetting, indexing, querying, or filtering. We'll treat the terms as synonymous.

There are lots of ways to do this. Mostly we do "Boolean" selection, which we address in the next section. We review more direct options here, mostly at high speed because they're not things we use much.

In the outline below, df is a dataframe, var and varn are variable names, n1 and n2 are integers,

• vlist = ['var1', 'var2'] is a list of variable names, and
• nlist = [0, 3, 4] is a list of numerical variable or observation indexes and
• bools is a list or pandas Series of booleans (True and False).

Some of the basic selection/indexing/slicing methods have the form:

• df[var] extracts a variable -- a series, in other words.
• df[vlist] extracts a new dataframe consisting of the variables in vlist.
• df[nlist] does the same thing.
• df[bools]: extracts each row where the corresponding element in bools is true. len(bools) must be equal to df.size[0]
• df[n1:n2] extracts observations n1 to n2-1, the traditional slicing syntax.

Example. Let's try each of these in a different cell and see what they do:

• small[['Country', 'Units']]
• small[[0, 4]]
• small['2011']
• small[1:3]

Can you explain the results?

Series indexing

Indexing a Series is a little different because we only have one column, so all indexing operations interact with rows.

The rules here are a little subtle, so we'll show examples and add comments that explain what each example does

In the list below s is a Series, n is an integer, nlist = [0, 3] is a list of integers, and i is a string, and is is a list of strings

• s[n]: if the index has dtype int, this extracts the row with index n. Otherwise extracts the nth row (starting at zero)
• s[nlist]: if the index has dtype int, this extracts rows with indices in nlist returning NaN if they don't appear. Otherwise extracts the rows at positions in nlist, filling with NaN for invalid positions
• s[i]: if the index has dtype object, this extracts the row with index i, otherwise it is an error
• s[is]:

Boolean selection

We choose observations that satisfy one or more conditions. Boolean selection consists of two steps that we typically combine in one statement:

• Use a comparison to construct a Boolean variable consisting of True and False.
• Compute df[comparison], where df is a dataframe and comparison is a comparison. This will select the observations (rows) for which comparison is true and throw away the others.

We work through this one step at a time:

• Example: apply the want operator
• Comparisons for dataframes
• Boolean selection: select observations for which the comparison is True
• The isin method

This is easier to describe with an example.

Example: Apply the want operator to WEO

Our want here is to take the weo dataframe and extract government debt and deficits for a given set of countries. Putting this to work involves several steps.

Here's the head of the dataframe to remind us what we're dealing with.

Find variable and country codes. Which ones do we want? Let's start by seeing that's available. Here we create special dataframes that include all the variables and their definitions and all the countries.

Note the use of the drop_duplicates method, which does what it sounds like: remove duplicate rows (!)

Exercise.

• Construct a list of countries with countries = weo['Country']; that is, without applying the drop_duplicates method. How large is it? How many duplicates have we dropped?

Comment. Now that we have the country and variable codes, we can be more explicit about what we want. We want observations with those country and variable codes.

We work up to the solution one step at a time.

Comparisons for series

We can construct comparisons for series (dataframe columns) much as we did with simple variables. The difference is that we get a complete column of True/False responses, not just one.

Mutiple comparisons have a different syntax than we saw earlier: and is replaced by &, and or is replaced by |. And when we have more than one comparison, we need to enclose them in parentheses.

Examples. Consider the comparisons:

• small['Units'] == 'National currency'
• small['2011'] >= 100
• (small['Units'] == 'National currency') & (small['2011'] >= 100)
• (small['Units'] == 'National currency') | (small['2011'] >= 100)

Remind yourself what the & and | do.

Boolean selection

Boolean selection simply chooses those observations for which a condition is True. Some people refer to this as filtering. The syntax is

df[comparison]

The result is a new dataframe of observations in which comparison is true.

Example. We choose obervations for which the units are 'National currency'. We do this first in two steps, then in one.

Exercise. Construct dataframes for which

• small['Units'] does not equal 'National currency'.
• small['Units'] equals 'National currency' and small['2011'] is greater than 100.

The isin method

Pay attention now, this is really useful. Suppose we want to extract the data for which weo['Country'] == 'Argentina' or weo['Country'] == 'Greece' (Greece). We could do that by combining the comparisons:

(weo['Country'] == 'Aregentina') | (weo['Country'] == 'Greece')

Remind youself that | stands for "or." (What do we use for "and"?)

A simpler approach is to apply the isin method to a variable. This sets the comparison equal to True if the value of the observation is of weo['Country'] equals any element in a list. We could do the same thing using mulitple comparisons, but this is a lot easier.

Let's see how this works.

Example. Let's apply the same logic to variable codes. If we want to extract the observations with codes

vlist = ['GGXWDG_NGDP', 'GGXCNL_NGDP']

we would use

Comment. We're choosing 2 variables from 45, so there are lots of Falses.

Comment. We can do the same thing with countries. If we want to choose two variables and three countries, the code looks like:

Comments.

• We've now done what we described when we applied the want operator.
• This is a go-to method. Circle it for later reference.
• This is a go-to method. Circle it for later reference.

Exercise. Use the isin method to extract Gross domestic product in US dollars for China, India, and the United States. Assign the result to the dataframe gdp. Hint: You can adapt the code we just ran. The variable code is NGDPD. The country codes are CHN, IND, and USA.

Exercise (challenging). Plot the variable gdp['2015'] as a bar chart. What would you say it needs?

The contains method

Another useful one. The contains string method for series identifies observations that contain a specific string. If yes, the observation is labelled True, if no, False. A little trick converts the True/False outcomes to ones and zeros.

We apply it to the Topics variable of the Entry Poll dataframe entry_poll. You may recall that this variable could have more than one response. We tease them apart with the contains method. Our want is to have a yes/no variable for each response.

Comment. That's pretty good, we now know which students mentioned Machine Learning and which did not. It's more useful, though, to convert this to zeros (False) and ones (True), which we do with this trick: we multiply by 1.

Comment. Now let's do the same for some of the other entries and save them in new variables.

Comment. You might want to think about this a minute. Or two.

Comment. Just for fun, here's a bar graph of the result.

and a pie chart

Review

Let's remind ourselves what we've learned.

Exercise. We explore the Census's Business Dynamics Statistics, a huge collection of data about firms. We've extracted a small piece of one of their databases that includes these variables for 2013:

• Size: size category of firms based on number of employees
• Firms: number of firms in each size category
• Emp: number of employees in each size category

Run the code cell below to load the data.

Use the dataframe bds to:

• Compute the mean number of firms with nbar = bds['Firms'].mean().
• Generate the new variable bds['AboveMean'] that is True if the value of bds['Firms'] is above the mean, False otherwise.
• What dtype is this new variable?
• Select the observations for which the number of firms is above the mean.
• Challenging. Fix the size categories. Specifically, use a string method to eliminate the prefixes a), b), etc. That is, change a) 1 to 4 to 1 to 4, b) 5 to 9 to 5 to 9, and so on. Hint: Use the split method.