Final Review

[lwoldemariam]

The goal of the project was to understand the damage of earthquakes on buildings in Nepal by predicting the damage grade for buildings. There are 39 features and 260,601 observations. Ordinal regression, multinomial logistic regression, and random forests were used, resulting in the multinomial regression having the lowest misclassification rate.  

Things I liked

1. The exploratory data analysis section was very useful and in depth. The plots, such as in figure 3, were nicely made. This section was useful in gaining background with regards to the problem and data set.
2. The formatting and style of the project (sections, font, and organization) helped make the report clear.
3. A very in-depth discussion of the models. The reasoning behind model choice was described; for example, section 4.3 talked of choosing trees because of previous analysis. These sections also included mathematical expressions, which served as a useful reminder to the reader.  
4. A good discussion of false positive versus false negatives in the ordinal regression section: it is better to predict a higher grade than a lower grade because the prediction of a higher grade will result in more protection. 

Improvement

1. It might have been useful to talk about why the ordinal regression model gave the results that it did. Why were so many buildings being predicted to be damage grade 2 for ordinal regression (although it is better to predict a higher grade)?
2. Although it was not difficult to understand, there was some inconsistency between variable names (lambda in the equation, but ‘c’ in the discussion).
3. A discussion on which the difference between features in predicting (coefficients, etc.) would help understand the importance of each variable.  Overall, the project was well thought-out and very clear!