pdexplorer 0.0.32

A Stata emulator for Python/pandas

"...succinctness is power... we take the trouble to develop high-level languages... so that we can say (and more importantly, think) in 10 lines of a high-level language what would require 1000 lines of machine language... -Paul Graham, Succinctness is Power

Installation

pdexplorer is available on PyPI. Run pip to install:

pip install pdexplorer

Then import the pdexplorer commands with

from pdexplorer import *

pdexplorer

pdexplorer is a Stata emulator for Python/pandas. In contrast to raw Python/pandas, Stata syntax achieves succinctness by:

• Using spaces and commas rather than parentheses, brackets, curly braces, and quotes (where possible)
• Specifying a set of concise commands on the "current" dataset rather than cluttering the namespace with multiple datasets
• Being verbose by default i.e., displaying output that represents the results of the command
• Having sensible defaults that cover the majority of use cases and demonstrate common usage
• Allowing for namespace abbreviations for both commands and variable names
• Employing two types of column names: Variable name are concise and used for programming. Variable labels are verbose and used for presentation.
• Packages are imported lazily e.g., import torch is loaded only when it's first used by a command. This ensures that from pdexplorer import * runs quickly.

Examples

Load Stata dataset and perform exploratory data analysis

webuse('auto')
browse()

See https://www.stata.com/manuals/dwebuse.pdf

Summarize Data by Subgroups

webuse('auto')
with by('foreign'):
    summarize('mpg weight')

See https://www.stata.com/manuals/rsummarize.pdf

Ordinary Least Squares (OLS) Regression

webuse('auto')
regress('mpg weight foreign')
ereturnlist()

Return Values

In the last example, note the use of ereturnlist(), corresponding to the Stata command ereturn list. Additionally, a Python object may also be available as the command's return value. For example,

webuse('auto')
results = regress('mpg weight foreign')

Here, results is a RegressionResultsWrapper object from the statsmodels package.

Similarly,

results = regress('mpg weight foreign', library='scikit-learn')

Now, results is a LinearRegression object from the scikit-learn package.

Finally,

results = regress('mpg weight foreign', library='pytorch')

Here, results is a torch.nn.Linear object from the PyTorch package.

LLM Fine Tuning Examples

Sentiment Analysis using HuggingFace

from pdexplorer import *
from pdexplorer.tests.fixtures import yelp_reviews
df = pd.DataFrame.from_records(yelp_reviews)
use(df) # load examples into pdexplorer
fthuggingface("stars text", task="sentiment-analysis", model_name="distilbert-base-uncased") # slow
askhuggingface("I absolutely loved Burgerosity!", task="sentiment-analysis")

Next Word Prediction using HuggingFace

from pdexplorer import *
from pdexplorer.tests.fixtures import eli5
df = pd.DataFrame.from_records(eli5)
use(df) # load examples into pdexplorer
fthuggingface("text", task="text-generation", model_name="distilgpt2") # slow
askhuggingface("A poem about Mickey Mouse in iambic pentameter:\n", task="text-generation")

Next Word Prediction using OpenAI (gpt-3.5-turbo)

from pdexplorer import *
from pdexplorer.tests.test_ftgpt import df
use(df)
ftgpt("assistant user system") # slow; requires OPENAI_API_KEY environment variable
askgpt("A poem about Mickey Mouse in iambic pentameter:\n")

Why not use Stata instead of pandas?

Stata is great, but Python/pandas is free and easier to integrate with other applications. For example, you can build a web server in Python, but not Stata; You can run Python in AWS SageMmaker, but not Stata.

Additionally, even for devout Stata users, there is utility in being able to run Stata commands through a Python stack for comparison purposes.

How pdexplorer fulfills the Zen of Python (relative to pandas)

YES	NO
Beautiful is better than ugly.	Explicit is better than implicit.
Simple is better than complex.	In the face of ambiguity, refuse the temptation to guess.
Flat is better than nested.
Readability counts.
Although practicality beats purity.
There should be one-- and preferably only one --obvious way to do it.
Now is better than never.

How pdexplorer differs from Stata

• Commands are implemented as Python functions and hence require at least one set of parentheses
• pdexplorer uses Python libraries under the hood. (The result of a command reflects the output of those libraries. See above.)
• There is no support for mata. Under the hood, pdexplorer is just the Python data stack.
• The API for producing charts is based on Altair, not Stata.
• pdexplorer adds commands for machine learning (using sklearn, PyTorch, and huggingface)

Syntax summary

With few exceptions, the basic Stata language syntax (as documented here) is

[by varlist:] command [varlist] [=exp] [if exp] [in range] [weight] [, options]

where square brackets distinguish optional qualifiers and options from required ones. In this diagram, varlist denotes a list of variable names, command denotes a Stata command, exp denotes an algebraic expression, range denotes an observation range, weight denotes a weighting expression, and options denotes a list of options.

The by varlist: prefix causes Stata to repeat a command for each subset of the data for which the values of the variables in varlist are equal. When prefixed with by varlist:, the result of the command will be the same as if you had formed separate datasets for each group of observations, saved them, and then gave the command on each dataset separately. The data must already be sorted by varlist, although by has a sort option.

In pdexplorer, this gets translated to

with by('varlist'):
    command("[varlist] [=exp] [if exp] [in range] [weight] [, options]", *args, **kwargs)

where *args, and **kwargs represent additional arguments that are available in a pdexplorer command but not in the equivalent Stata command.

Sometimes, Stata commands are two words. In such cases, the pdexplorer command is a concatenation of the two words. For example,

label data "label"

becomes

labeldata("label")

Module Dependencies

File location	Description	Dependencies
/*.py	commands that are native to Stata related to data wrangling or statistics	pandas
/_altair_mapper.py	commands that are native to Altair for charting	altair
shortcuts/*.py	shortcut commands related to data wrangling, statistics, or charting	all of the above
finance/*.py	commands that are specific to financial applications	all of the above
ml/*.py	commands that use machine learning techniques (and are outside the scope of Stata)	scikit-learn
nn/*.py	commands that use neutral networks (primarily built using PyTorch)	PyTorch
data/*.py	python scripts that collect data from various sources	Data suppliers
experimental/*.py	commands that are current under development and not yet stable	N/A

Command Dependencies

pdexplorer command	package dependency
cf	ydata-profiling or sweetviz
browse	dtale
regress	statsmodels or scikit-learn or PyTorch

Charts

pdexplorer departs somewhat from Stata in producing charts. Rather than emulating Stata's chart syntax, pdexplorer uses Altair with some syntactic sugar. Take the example "Simple Scatter Plot with Tooltips":

import webbrowser
import altair as alt
from vega_datasets import data
source = data.cars()
chart = alt.Chart(source).mark_circle(size=60).encode(
    x="Horsepower",
    y="Miles_per_Gallon",
    color='Origin',
    tooltip=["Name", "Origin", "Horsepower", "Miles_per_Gallon"],
)
chart.save('mychart.html')
webbrowser.open('mychart.html')

In pdexplorer, this becomes:

from pdexplorer import *
webuse("cars", "vega")
circlechart("miles_per_gallon horsepower, \
    color(origin) tooltip(name origin horsepower miles_per_gallon)",
    size=60,
)

In other words, pdexplorer supports a varlist parameter for y/x encodings. Additional encodings are specified via the Stata options syntax.

In the above example, circlechart automatically displays the chart in a web browser. However, sometimes it's necessary to add features to the alt.Chart Altair object. To use the alt.Chart object, we can use circlechart_ instead of circlechart. For example,

import webbrowser
from pdexplorer import *
webuse("cars", "vega")
altair_chart_object = circlechart_(
    "miles_per_gallon horsepower, color(origin) \
    tooltip(name origin horsepower miles_per_gallon)",
    size=60,
)
altair_chart_object.configure_legend(
    strokeColor='gray',
    fillColor='#EEEEEE',
    padding=10,
    cornerRadius=10,
    orient='top-right'
) # See https://altair-viz.github.io/user_guide/configuration.html#legend-configuration
altair_chart_object.save('mychart.html')
webbrowser.open('mychart.html')

Instead of saving the chart, explicitly, we can also write

altair_chart_object()

The () at the end tells Altair to open the chart in a web browser. This method is not available in alt.Charts itself, but we monkey patched this into the class for convenience.

Similarly, the following two statements are identical,

circlechart_("miles_per_gallon horsepower")() # the object-oriented style
circlechart("miles_per_gallon horsepower") # the imperative style

In this example, circlechart_ itself simply returns a alt.Chart object and () indicates that the chart should be displayed in a web browser.

Since we can access the alt.Chart object, we can also specify encodings explicitly using Altair's encode method e.g.,

from pdexplorer import *
webuse("cars", "vega") # note that names are forced to be lower case by default
circlechart_(size=60).encode(
    x="horsepower",
    y="miles_per_gallon",
    color="origin",
    tooltip=["name", "origin", "horsepower", "miles_per_gallon"],
)()  # () indicates that the chart is complete and should be opened in a web browser

Since pdexplorer charts are just alt.Chart objects, Layered and Multi-View Charts are also supported e.g.,

from pdexplorer import *
webuse("cars", "vega")
chartA = circlechart_("miles_per_gallon weight_in_lbs")
chartB = circlechart_("horsepower weight_in_lbs")
(chartA & chartB)() # Vertically concatenate chartA and chartB

Finally, pdexplorer also offers syntactic sugar for charting multiple x/y variables. More specifically, the previous block can be written as

from pdexplorer import *
webuse("cars", "vega")
circlechart("miles_per_gallon horsepower weight_in_lbs", stacked=True)

Note that Stata's varlist interpretation is used here by default i.e., var1 var2 var3 is assumed to represent yvar1 yvar2 xvar. We can change this interpretation with the optional argument yX.

from pdexplorer import *
webuse("cars", "vega")
circlechart("miles_per_gallon horsepower weight_in_lbs", yX=True, stacked=True)

Now var1 var2 var3 is assumed to represent yvar xvar1 xvar2 as it would be for the regress command.

The Stata default is to layer all variables onto a single chart. The stacked=True option allows the graphs to be shown on separate grids. If stacked=False, the charts are all shown on the same grid i.e.,

from pdexplorer import *
webuse("cars", "vega")
circlechart("miles_per_gallon horsepower weight_in_lbs") # stacked=False is the default option

Note that pdexplorer uses Altair's transform_fold method under the hood. For further customization, the Altair methods can be used explicitly e.g.,

import altair as alt
from pdexplorer import *
webuse("cars", "vega")
circlechart_().transform_fold(
    ["Miles_per_Gallon", "Horsepower"] # Note that circlechart_ variable labels are accessible
).encode(
    y="value:Q", x=alt.X("weight_in_lbs", title="Weight_in_lbs"), color="key:N"
)()

alt.Chart.transform_fold is Altair's version of pandas.melt. So another option is to first reshape the data using pandas and then use Altair for charting.

import altair as alt
from pdexplorer import *
webuse("cars", "vega")
melt("miles_per_gallon horsepower,  keep(weight_in_lbs)")
alt.Chart(current.df_labeled).mark_circle().encode(
    y="value:Q", x='Weight_in_lbs', color="variable:N"
)()

Note that the Altair documentation suggests the latter approach in most cases where a data transformation is required.

Abbreviations

As mentioned ealier, Stata supports name abbreviations for both variable names as well as command names. In pdexplorer, all the following regression statements are equivalent:

from pdexplorer import *
webuse("auto")
reg('price mpg weight')
regr('price mpg weight')
regre('price mpg weight')
regres('price mpg weight')
regress('price mpg weight')
reg('pr mpg wei')

Similarly, for charting,

from pdexplorer import *
webuse("cars", "vega")
circle("miles_per_gallon horsepower")
circlec("miles_per_gallon horsepower")
circlech("miles_per_gallon horsepower")
circlecha("miles_per_gallon horsepower")
circlechar("miles_per_gallon horsepower")
circlechart("miles_per_gallon horsepower")
circle("miles horse")

References

• https://aeturrell.github.io/coding-for-economists/coming-from-stata.html
• https://www.stata.com/manuals13/u27.pdf
• https://www.stata.com/manuals13/u11.pdf