Overview
Recent years have witnessed an unprecedented availability of information on social, economic, and health-related phenomena. Researchers, practitioners, and policymakers have nowadays access to huge datasets (the so-called “Big Data”) on people, companies and institutions, web and mobile devices, satellites, etc., at increasing speed and detail.
Machine learning is a relatively new approach to data analytics, which places itself in the intersection between statistics, computer science, and artificial intelligence. Its primary objective is that of turning information into knowledge and value by “letting the data speak”. To this purpose, machine learning limits prior assumptions on data structure, and relies on a model-free philosophy supporting algorithm development, computational procedures, and graphical inspection more than tight assumptions, algebraic development, and analytical solutions. Computationally unfeasible few years ago, machine learning is a product of the computer’s era, of today machines’ computing power and ability to learn, of hardware development, and continuous software upgrading.
This course is a primer to machine learning techniques using Stata. Stata owns today various packages to perform machine learning which are however poorly known to many Stata users. This course fills this gap by making participants familiar with (and knowledgeable of) Stata potential to draw knowledge and value form row, large, and possibly noisy data. The teaching approach will be mainly based on the graphical language and intuition more than on algebra. The training will make use of instructional as well as real-world examples, and will balance evenly theory and practical sessions.
After the course, participants are expected to have an improved understanding of Stata potential to perform some of the most used marching learning techniques, thus becoming able to master research tasks including, among others:
- (i) factor-importance detection
- (ii) signal-from-noise extraction
- (iii) correct model specification
- (iv) model-free classification, both from a data-mining and a causal perspective.
Course Agenda
DAY 1
Session 1 (10:00-12:00 GMT): The basics of Machine Learning
Machine Learning: definition, rational, usefulness
- Supervised vs. unsupervised learning
- Regression vs. classification problems
- Inference vs. prediction
- Sampling vs. specification error
Coping with the fundamental non-identifiability of E(y|x)
- Parametric vs. non-parametric models
- The trade-off between prediction accuracy and model interpretability
- Goodness-of-fit measures
- Measuring the quality of fit: in-sample vs. out-of-sample prediction power
- The bias-variance trade-off and the Mean Square Error (MSE) minimization
- Training vs. test mean square error
- The information criteria approach
Estimating training and test error
- Validation set, K-fold cross-validation, and the Bootstrap
Session 2 (14:00-16:00 GMT): Model Selection and regularization
Model selection as a correct specification procedure
- The information criteria approach
- Subset Selection
- Best subset selection
- Backward stepwise selection
- Forward stepwise Selection
Shrinkage Methods
- Lasso and Ridge, and Elastic regression
- Adaptive Lasso
- Information criteria and cross validation for Lasso
Stata implementation
DAY 2
Session 1 (10:00-12:00 GMT):Discriminant analysis and nearest-neighbor classification
- The classification setting
- Bayes optimal classifier and decision boundary
- Misclassification error rate
Discriminant analysis
- Linear and quadratic discriminant analysis
- Naive Bayes classifier
The K-nearest neighbours classifier
Stata implementation
Session 2 (14:00-16:00 GMT): Neural networks
The neural network model
- Neurons, hidden layers, and multi-outcomes
Training a neural networks
- Back-propagation via gradient descent
- Fitting with high dimensional data
- Fitting remarks
Cross-validating neural network hyperparameters
Stata implementation
Session 3 (16:00-17:00 GMT): Q&A session with the instructor
The neural network model
- Neurons, hidden layers, and multi-outcomes
Training a neural networks
- Back-propagation via gradient descent
- Fitting with high dimensional data
- Fitting remarks
Cross-validating neural network hyperparameters
Stata implementation
Reading List:
- The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Hastie, T., Tibshirani, R., Friedman, J., Springer (2009)
- An Introduction to Statistical Learning, Gareth, J., Witten, D., Hastie, T., Tibshirani, R., Springer (2013)
- Microeconometrics Using Stata, Cameron e Trivedi, Revised Edition, StataPress (2010)
- A Super-Learning Machine for Predicting Economic Outcomes, Giovanni Cerulli
Prerequisites
Knowledge of basic statistics, Stata and econometrics is required, including:
- The notion of conditional expectation and related properties;
- point and interval estimation;
- regression model and related properties;
- probit and logit regression.
Terms & Conditions
- Student registrations: Attendees must provide proof of full time student status at the time of booking to qualify for student registration rate (valid student ID card or authorised letter of enrolment).
- Additional discounts are available for multiple registrations. Contact us for more information.
- Temporary, time limited licences for the software(s) used in the course will be provided. You are required to install the software provided prior to the start of the course.
- Full payment of course fees is required prior to the course start date to guarantee your place.
- Registration closes 1 calendar day prior to the start of the course.
Cancellations or changes to your registration
- 100% fee returned for cancellations made over 28-calendar days prior to start of the course.
- 50% fee returned for cancellations made 14-calendar days prior to the start of the course.
- No fee returned for cancellations made less than 14-calendar days prior to the start of the course.