MATH5855: Multivariate Analysis

During COVID -
- 3 x 25% assignments
- 1 x 25% exam

No formal prerequisites stated (quite common for postgraduate courses). Assumed knowledge is pretty much any general level of statistics (for UNSW undergraduates, definitely MATH2801/2901). Highly advised is some knowledge of linear algebra (MATH2501/MATH2601 is great to have, but also some knowledge of matrix calculus could be useful). To make life easier, I recommend also knowing linear models (for example, equivalent of MATH2831/2931) for the sake of multivariate linear models. I also recommend some in depth statistical inference (for example, equivalent of MATH3811/MATH3911), since the likelihood ratio test features a lot.

This is one of many postgraduate statistics courses. Recently, it has remained on a yearly offering.

Multivariate analysis is basically the answer to the undergraduate problem: why have we never been working with multivariate distributions? This course literally deals with them, but one should note that the multivariate normal distribution takes up about 60% of the work (it is the most important multivariate distribution). Much of the remainder is just on multivariate data in general (with a special topic in network analysis). Examples of multivariate data could be performances in all 10 sports of a decathlon, and say [height, weight, age]. The variables look univariate, but they come multivariate when you consider them together, and that is the main point.

The assessments were somewhat stranger this term. Moodle quizzes were used to input your final answer, and working was uploaded in a separate submission box (to possibly earn partial marks for mistakes). The assignments also had CodeRunner available eventually, which let you copy past R code (or just numeric answers) directly into a console instead. You had a choice on whether you wanted to use R or SAS for the coding components; there was no need to learn both. They honestly weren't hard, and most of my lost marks were (agonisingly) due to carelessness. I feel the weird assessment format was mostly because Pavel never had to deal with a cohort of like 180 students (more than DOUBLE that of last year), and Moodle's automated marking helped alleviate some of the work. It was awkward, but the system was still alright.

A fair bit of code in the lectures could be adapted to the assignments. This was very handy.

Course assessment could be argued to be a 2/5 difficulty by many people. However, Pavel also presented many theoretical derivations in the lectures. The harder ones weren't examinable, but they could've easily skyrocketed the difficulty to 4.5/5, hence the above rating I gave.

Originally, the final exam was going to be substituted by a fourth assignment, but then it was disallowed (I don't know why; he wasn't allowed to disclose). Hence its unusually low weighting. The weighting also resulted in a final exam that really wasn't too difficult, in my opinion. However, it was a bit annoying that we didn't find out about the negative marking until the day before the exam. (The scheme was very similar to that of COMP3231's.)

There's one thing I absolutely have to comment. All the theory in this course made me EXTREMELY thankful that I paid attention in linear algebra. That level of linear algebra can be overwhelming for so many students (understandably, not surprisingly). Very little of it was needed for the assessments, but arguably a lot was for understanding lecture theory.

Pavel tried to appeal to both theoretical minded audiences, as well as those that just wanted to know how to do the job (i.e. mostly applied statistics only). I think it worked; it just was hard.

During COVID - 1 x 2hr lecture, 2 x 1hr lecture (total: 4hrs). (Outside of COVID, I believe one of the lecture hours is switched for a lab/tute instead.)

3/5

Yes during COVID period. Likely yes otherwise, since Pavel uses lecture slides.

Dr. Pavel Krivitsky

Simple lecture slides and more prosaic lecture notes provided. (Exact same content; one is just easier to read because it has more full sentences.) Lecture scribbles usually weren't uploaded, but could be reviewed in the recording. A select few exercises appeared along the way. Support for R and SAS coding both given in this course. Lecture examples provided, and some challenge exercises available to try yourself.

3.5/5

Johnson RA and Wichern DW, Applied Multivariate Analysis, 6ED. Roughly 80% of the course is taken from here, and the textbook was liked by many students. I was too lazy to use it, as usual.

20T3

93 HD