Assignment

1 Lalonde NSW Data

A. Load the Lalonde experimental dataset with the lalonde data method from the module causalinference.utils. The outcome variable is earnings in 1978, and the co- variates are, in order:

Black       Indicator variable; 1 if Black, 0 otherwise.
Hispanic   Indicator variable; 1 if Hispanic, 0 otherwise.
Age         Age in years.
Married    Marital status; 1 if married, 0 otherwise. Nodegree Indicator variable; 1 if no degree, 0 otherwise. Education Years of education.
E74         Earnings in 1974.
U74         Unemployment status in 1974; 1 if unemployed, 0 otherwise.
E75         Earnings in 1975.
U75         Unemployment status in 1975; 1 if unemployed, 0 otherwise.

Using CausalModel from the module causalinference, provide summary statistics for the outcome variable and the covariates. Which covariate has the largest normalized difference?

B. Estimate the propensity score using the selection algorithm est propensity s. In se- lecting the basic covariates set, specify E74, U74, E75, and U75. What are the additional linear terms and second-order terms that were selected by the algorithm?

C. Trim the sample using trim s to get rid of observations with extreme propensity score values. What is the cut-off that is selected? How many observations are dropped as a result?

D. Stratify the sample using stratify s. How many propensity bins are created? Report the summary statistics for each bin.

E. Estimate the average treatment effect using OLS, blocking, and matching. For matching, set the number of matches to 2 and adjust for bias. How much do the estimates differ?

2 Document Classification

A. From the module sklearn.datasets, load the training data set using the method fetch 20newsgroups. This dataset comprises around 18000 newsgroups posts on 20 topics. Print out a couple sample posts and list out all the topic names.

B. Convert the posts (blobs of texts) into bag-of-word vectors. What is the dimensionality of these vectors? That is, what is the number of words that have appeared in this data set?

C. Use your favorite dimensionality reduction technique to compress these vectors into ones of K = 30 dimensions.

D. Use your favorite supervised learning model to train a model that tries to predict the topic of a post from the vectorized representation of the post you obtained in the previous step.

E. Use the test data to tune your model. Make sure to include K as a hyperparameter as well. Use accuracy score from sklearn.metrics as your evaluation metric. What is the highest accuracy you are able to achieve?