Idomatic Pandas

Q: How do I make my pandas code faster with parallelism?

A: You don’t need parallelism, you can use Pandas better.

-- Matthew Rocklin

Now that we have been exposed to the basic functionality of pandas, lets explore some more advanced features that will be useful when addressing more complex data management tasks.

As most statisticians/data analysts will admit, often the lion's share of the time spent implementing an analysis is devoted to preparing the data itself, rather than to coding or running a particular model that uses the data. This is where Pandas and Python's standard library are beneficial, providing high-level, flexible, and efficient tools for manipulating your data as needed.

As you may already have noticed, there are sometimes mutliple ways to achieve the same goal using pandas. Importantly, some approaches are better than others, in terms of performance, readability and ease of use. We will cover some important ways of maximizing your pandas efficiency.

Reshaping DataFrame objects

In the context of a single DataFrame, we are often interested in re-arranging the layout of our data.

This dataset in from Table 6.9 of Statistical Methods for the Analysis of Repeated Measurements by Charles S. Davis, pp. 161-163 (Springer, 2002). These data are from a multicenter, randomized controlled trial of botulinum toxin type B (BotB) in patients with cervical dystonia (spasmodic torticollis) from nine U.S. sites.

• Randomized to placebo (N=36), 5000 units of BotB (N=36), 10,000 units of BotB (N=37)
• Response variable: total score on Toronto Western Spasmodic Torticollis Rating Scale (TWSTRS), measuring severity, pain, and disability of cervical dystonia (high scores mean more impairment)
• TWSTRS measured at baseline (week 0) and weeks 2, 4, 8, 12, 16 after treatment began

This dataset includes repeated measurements of the same individuals (longitudinal data). Its possible to present such information in (at least) two ways: showing each repeated measurement in their own row, or in multiple columns representing multiple measurements.

The stack method rotates the data frame so that columns are represented in rows:

Have a peek at the structure of the index of the stacked data (and the data itself).

To complement this, unstack pivots from rows back to columns.

Exercise

Which columns uniquely define a row? Create a DataFrame called cdystonia2 with a hierarchical index based on these columns.

If we want to transform this data so that repeated measurements are in columns, we can unstack the twstrs measurements according to obs.

We can now merge these reshaped outcomes data with the other variables to create a wide format DataFrame that consists of one row for each patient.

A slightly cleaner way of doing this is to set the patient-level information as an index before unstacking:

To convert our "wide" format back to long, we can use the melt function, appropriately parameterized. This function is useful for DataFrames where one or more columns are identifier variables (id_vars), with the remaining columns being measured variables (value_vars). The measured variables are "unpivoted" to the row axis, leaving just two non-identifier columns, a variable and its corresponding value, which can both be renamed using optional arguments.

This illustrates the two formats for longitudinal data: long and wide formats. Its typically better to store data in long format because additional data can be included as additional rows in the database, while wide format requires that the entire database schema be altered by adding columns to every row as data are collected.

The preferable format for analysis depends entirely on what is planned for the data, so it is imporant to be able to move easily between them.

Method chaining

In the DataFrame reshaping section above, you probably noticed how several methods were strung together to produce a wide format table:

This approach of seqentially calling methods is called method chaining, and despite the fact that it creates very long lines of code that must be properly justified, it allows for the writing of rather concise and readable code. Method chaining is possible because of the pandas convention of returning copies of the results of operations, rather than in-place operations. This allows methods from the returned object to be immediately called, as needed, rather than assigning the output to a variable that might not otherwise be used. For example, without method chaining we would have done the following:

This necessitates the creation of a slew of intermediate variables that we really don't need.

Let's transform another dataset using method chaining. The measles.csv file contains de-identified cases of measles from an outbreak in Sao Paulo, Brazil in 1997. The file contains rows of individual records:

The goal is to summarize this data by age groups and bi-weekly period, so that we can see how the outbreak affected different ages over the course of the outbreak.

The best approach is to build up the chain incrementally. We can begin by generating the age groups (using cut) and grouping by age group and the date (ONSET):

What we then want is the number of occurences in each combination, which we can obtain by checking the size of each grouping:

This results in a hierarchically-indexed Series, which we can pivot into a DataFrame by simply unstacking:

Now, fill replace the missing values with zeros:

Finally, we want the counts in 2-week intervals, rather than as irregularly-reported days, which yields our the table of interest:

From this, it is easy to create meaningful plots and conduct analyses:

Pivoting

The pivot method allows a DataFrame to be transformed easily between long and wide formats in the same way as a pivot table is created in a spreadsheet. It takes three arguments: index, columns and values, corresponding to the DataFrame index (the row headers), columns and cell values, respectively.

For example, we may want the twstrs variable (the response variable) in wide format according to patient, as we saw with the unstacking method above:

Exercise

Try pivoting the cdystonia DataFrame without specifying a variable for the cell values:

A related method, pivot_table, creates a spreadsheet-like table with a hierarchical index, and allows the values of the table to be populated using an arbitrary aggregation function.

For a simple cross-tabulation of group frequencies, the crosstab function (not a method) aggregates counts of data according to factors in rows and columns. The factors may be hierarchical if desired.

Data transformation

There are a slew of additional operations for DataFrames that we would collectively refer to as transformations which include tasks such as:

• removing duplicate values
• replacing values
• grouping values.

Dealing with duplicates

We can easily identify and remove duplicate values from DataFrame objects. For example, say we want to remove ships from our vessels dataset that have the same name:

These rows can be removed using drop_duplicates

Value replacement

Frequently, we get data columns that are encoded as strings that we wish to represent numerically for the purposes of including it in a quantitative analysis. For example, consider the treatment variable in the cervical dystonia dataset:

A logical way to specify these numerically is to change them to integer values, perhaps using "Placebo" as a baseline value. If we create a dict with the original values as keys and the replacements as values, we can pass it to the map method to implement the changes.

Alternately, if we simply want to replace particular values in a Series or DataFrame, we can use the replace method.

An example where replacement is useful is replacing sentinel values with an appropriate numeric value prior to analysis. A large negative number is sometimes used in otherwise positive-valued data to denote missing values.

In such situations, we can use replace to substitute nan where the sentinel values occur.

We can also perform the same replacement that we used map for with replace:

Inidcator variables

For some statistical analyses (e.g. regression models or analyses of variance), categorical or group variables need to be converted into columns of indicators--zeros and ones--to create a so-called design matrix. The Pandas function get_dummies (indicator variables are also known as dummy variables) makes this transformation straightforward.

Let's consider the DataFrame containing the ships corresponding to the transit segments on the eastern seaboard. The type variable denotes the class of vessel; we can create a matrix of indicators for this. For simplicity, lets filter out the 5 most common types of ships.

Exercise

Create a subset of the vessels DataFrame called vessels5 that only contains the 5 most common types of vessels, based on their prevalence in the dataset.

We can now apply get_dummies to the vessel type to create 5 indicator variables.

Discretization

Pandas' cut function can be used to group continuous or countable data in to bins. Discretization is generally a very bad idea for statistical analysis, so use this function responsibly!

Lets say we want to bin the ages of the cervical dystonia patients into a smaller number of groups:

Let's transform these data into decades, beginnnig with individuals in their 20's and ending with those in their 80's:

The parentheses indicate an open interval, meaning that the interval includes values up to but not including the endpoint, whereas the square bracket is a closed interval, where the endpoint is included in the interval. We can switch the closure to the left side by setting the right flag to False:

Since the data are now ordinal, rather than numeric, we can give them labels:

A related function qcut uses empirical quantiles to divide the data. If, for example, we want the quartiles -- (0-25%], (25-50%], (50-70%], (75-100%] -- we can just specify 4 intervals, which will be equally-spaced by default:

Alternatively, one can specify custom quantiles to act as cut points:

Exercise

Use the discretized segment lengths as the input for get_dummies to create 5 indicator variables for segment length:

Categorical Variables

One of the keys to maximizing performance in pandas is to use the appropriate types for your data wherever possible. In the case of categorical data--either the ordered categories as we have just created, or unordered categories like race, gender or country--the use of the categorical to encode string variables as numeric quantities can dramatically improve performance and simplify subsequent analyses.

When text data are imported into a DataFrame, they are endowed with an object dtype. This will result in relatively slow computation because this dtype runs at Python speeds, rather than as Cython code that gives much of pandas its speed. We can ameliorate this by employing the categorical dtype on such data.

This creates an unordered categorical variable. To create an ordinal variable, we can specify order=True as an argument to astype:

However, this is not the correct order; by default, the categories will be sorted alphabetically, which here gives exactly the reverse order that we need.

To specify an arbitrary order, we can used the set_categories method, as follows:

Notice that we obtained set_categories from the cat attribute of the categorical variable. This is known as the category accessor, and is a device for gaining access to Categorical variables' categories, analogous to the string accessor that we have seen previously from text variables.

Additional categoried can be added, even if they do not currently exist in the DataFrame, but are part of the set of possible categories:

To complement this, we can remove categories that we do not wish to retain:

Or, even more simply:

For larger datasets, there is an appreciable gain in performance, both in terms of speed and memory usage.

Data aggregation and GroupBy operations

One of the most powerful features of Pandas is its GroupBy functionality. On occasion we may want to perform operations on groups of observations within a dataset. For exmaple:

• aggregation, such as computing the sum of mean of each group, which involves applying a function to each group and returning the aggregated results
• slicing the DataFrame into groups and then doing something with the resulting slices (e.g. plotting)
• group-wise transformation, such as standardization/normalization

This grouped dataset is hard to visualize

However, the grouping is only an intermediate step; for example, we may want to iterate over each of the patient groups:

A common data analysis procedure is the split-apply-combine operation, which groups subsets of data together, applies a function to each of the groups, then recombines them into a new data table.

For example, we may want to aggregate our data with with some function.

We can aggregate in Pandas using the aggregate (or agg, for short) method:

Notice that the treat and sex variables are not included in the aggregation. Since it does not make sense to aggregate non-string variables, these columns are simply ignored by the method.

Some aggregation functions are so common that Pandas has a convenience method for them, such as mean:

The add_prefix and add_suffix methods can be used to give the columns of the resulting table labels that reflect the transformation:

Exercise

Use the quantile method to generate the median values of the twstrs variable for each patient.

If we wish, we can easily aggregate according to multiple keys:

Alternately, we can transform the data, using a function of our choice with the transform method:

It is easy to do column selection within groupby operations, if we are only interested split-apply-combine operations on a subset of columns:

Or, as a DataFrame:

If you simply want to divide your DataFrame into chunks for later use, its easy to convert them into a dict so that they can be easily indexed out as needed:

By default, groupby groups by row, but we can specify the axis argument to change this. For example, we can group our columns by dtype this way:

Its also possible to group by one or more levels of a hierarchical index. Recall cdystonia2, which we created with a hierarchical index:

The level argument specifies which level of the index to use for grouping.

Apply

We can generalize the split-apply-combine methodology by using apply function. This allows us to invoke any function we wish on a grouped dataset and recombine them into a DataFrame.

The function below takes a DataFrame and a column name, sorts by the column, and takes the n largest values of that column. We can use this with apply to return the largest values from every group in a DataFrame in a single call.

To see this in action, consider the vessel transit segments dataset (which we merged with the vessel information to yield segments_merged). Say we wanted to return the 3 longest segments travelled by each ship:

Notice that additional arguments for the applied function can be passed via apply after the function name. It assumes that the DataFrame is the first argument.

Exercise

Load the dataset in titanic.xls. It contains data on all the passengers that travelled on the Titanic.

Women and children first?

1. Use the groupby method to calculate the proportion of passengers that survived by sex.
2. Calculate the same proportion, but by class and sex.
3. Create age categories: children (under 14 years), adolescents (14-20), adult (21-64), and senior(65+), and calculate survival proportions by age category, class and sex.

References

Python for Data Analysis Wes McKinney