The Crosstab Kite

Decision-makers often care how long it takes for important events to happen. In this article, I show how to compute Kaplan-Meier survival curves and Nelson-Aalen cumulative hazard curves directly in SQL, so you can answer time-to-event questions directly in your SQL-based analytics tables and dashboards.

How to construct survival tables from duration tables

July 15, 2021

The survival table is the workhorse of univariate survival analysis. Previously, I showed to build a duration table from an event log; now I show how to take the next steps, from duration table to survival table and estimates of survival and hazard curves.

A review and how-to guide for Microsoft Form Recognizer

July 13, 2021

Information Extraction

Form Recognizer is Microsoft Azure's answer to Amazon Textract and Google Form Parser for information extraction from form documents. I put all three to the test on a challenging set of invoice documents and was surprised to see Microsoft's service come up short.

A review and how-to guide for Amazon Textract

July 2, 2021

Information Extraction

At just two years old, Amazon Textract is already the grizzled veteran of the fast-moving information extraction-as-a-service market. I put it to the test against Google Form Parser and Microsoft Form Recognizer to see which service has the best speed, accuracy, and developer experience.

How to plot survival curves with Plotly and Altair

June 29, 2021

Survival curve plots are an essential output of survival analysis. The main Python survival analysis packages show only how to work with Matplotlib and hide plot details inside convenience functions. This article shows how to draw survival curves with two other Python plot libraries, Altair and Plotly.

Google Form Parser, a review and how-to

June 15, 2021

Information Extraction

Google Form Parser is a new challenger in the information extraction arena, offering general-purpose off-the-shelf form extraction. We compared Form Parser to Amazon Textract and Microsoft Form Recognizer in terms of accuracy, speed, and ease of use. Take our code snippets and try it yourself!

How to build duration tables from event logs with SQL

June 9, 2021

Duration tables are a common input format for survival analysis but they are not trivial to construct. In our last article, we used Python to convert a web browser activity log into a duration table. Event logs are usually stored in databases, however, so in this article we do the same conversion with SQL.

How to convert event logs to duration tables for survival analysis

June 7, 2021

Survival models describe how much time it takes for some event to occur. This is a natural way to think about many applications but setting up the data can be tricky. In this article, we use Python to turn an event log into a duration table, which is the input format for many survival analysis tools.

A checklist for professionalizing machine learning models

May 27, 2021

Productionization

Data scientists are drawn to the latest and greatest machine learning tasks and models, even though tabular binary classification remains the industry workhorse. We should take more pride in professionalizing the models that we know to work, rather than reflexively chasing every new thing.

Review: Statistical Rethinking, by Richard McElreath

May 21, 2021

This is an absolute gem of a book. McElreath has found an elusive combination: Statistical Rethinking is not only one of the best intro textbooks for both causal and Bayesian modeling, it's also highly readable, even entertaining.

Streamlit review and demo: best of the Python data app tools

May 12, 2021

Presentation

Streamlit has quickly become the hot thing in data app frameworks. We put it to the test to see how well it stands up to the hype. Come for the review, stay for the code demo, including detailed examples of Altair plots.

How to analyze a staged rollout experiment

May 7, 2021

Recently we argued that confidence intervals are a poor choice for analyzing staged rollout experiments. In this article, we show a better way: a Bayesian approach that gives decision-makers the answers they really need. Check out our interactive Streamlit app first, then the article for the details.

Research digest: what does cross-validation really estimate?

April 26, 2021

Model Evaluation

A new paper by Bates, Hastie, and Tibshirani reminds us that estimating a model's predictive performance is tricky. For linear models at least, cross-validation does not estimate the generalization error of a specific model, as you would assume. How much does this matter for data science in practice?

What we're reading, April edition

April 7, 2021

The data science content firehose can be overwhelming; these are the pieces we think might be worth your time to check out. This month we're focusing on causal inference.

No, your confidence interval is not a worst-case analysis

March 31, 2021

Confidence intervals are one of the most misunderstood concepts in statistics. Common sense says the lower bound of a confidence interval is a good estimate of the worst-case outcome, but the definition of the confidence interval doesn't allow us to make this claim. Or does it? Let's take a look.

Announcing ABGlossary, an experiment vocab translator

March 24, 2021

Defining terms is a key part of experimentation culture. The consequence, however, is that every community has its own experiment jargon, which makes it hard to spot patterns, let alone communicate across groups. We've created a lightweight tool called ABGlossary to help translate experiment vocab.

Choose your experiment platform wisely

March 17, 2021

To build a culture of fast, reliable, evidence-based innovation, you need an experiment platform. These tools support each stage of the experiment process and, done well, become the beating heart of your infrastructure.

In defense of statistical modeling

March 10, 2021

Data science remains hot but there is a persistent stream of articles that says the field is overhyped and that hiring managers and aspiring data scientists should focus more on engineering. Let's remember why data science's core skill of statistical modeling is so valuable.

Our experimentation roadmap

March 3, 2021

Experimentation is the gold standard for predicting the impact of a new idea. Simple designs like A/B testing sound easy but are fiendishly hard to get right. Data scientists are uniquely positioned to solve the challenge and help their companies develop the experimentation muscle.

What we're reading: February 22

February 22, 2021

What we're reading for the week of February 22nd. The data science content firehose can be overwhelming. These are the pieces we think might be a good use of your time to read and study.

Data before models, but problem formulation first

January 28, 2021

Recent tweets highlighted the importance of data annotation and curation in applied machine learning vs. model perfection. Data is indeed critical, but formulating a business problem as a data science task is even more foundational.

Conversion rate modeling: worth the effort?

January 9, 2021