Advice for assigning student projects in masters data science stats course

Question

I'd like some advice on assigning final projects in an applied stats course that I'm developing as part of a master's in data science program.

I don't have a lot of experience assigning student projects, and have never done it at the master's level.

One particular piece of advice I'd like is some guidelines for students to select a dataset for a final project. In said final project, I want students to ask at least one (is that enough?) interesting research question. Answering that question would ideally require them to employ many of the techniques learned in the course.

Here are some examples of the kind of advice I have in mind. Feel free to let me know why any example I've suggested is a bad idea.

• require at least _ continuous numerical variables, _ of which are continuous and _ categorical
• avoid datasets that are of the _ type
• require _ visualizations (where blank could be a number or a type)

Here are the topics of the course:

• Frequentist unit topics: central limit theorem, hypothesis testing, confidence intervals, ANOVA, linear regression, logistic regression.

• Bayesian unit topics: hypothesis testing, credible intervals, ANOVA, linear regression, logistic regression.

• Machine learning unit topics: KNN, regularized regression, linear & quadratic discriminant analysis, support vector machines, decision trees (w/ boosting & bagging).

Answer 1

I assign projects in my graduate courses. I've had good luck with the following:

I require students to submit a title, abstract, and references for approval. This ensures that the proposed project is on topic for the course and that it isn't too small (e.g. something like a homework exercise) or too big (e.g. something like a PhD disseratation.)

I encourage students to select projects related to or part of their MS thesis or PhD dissertation, provided that this is acceptable to their supervisor. Students in non-thesis master's programs (or who haven't started on a thesis yet) have a harder time finding topics- I usually have to supply a project topic in those cases.

I require students to submit a complete draft of the project well in advance of the due date for the final project. I grade this on completeness (i.e. no penalty for mathematical/statistical errors, but penalties for having a report missing important sections.) I provide a lot of feedback on these drafts and make it clear to students that they have to address any issues raised in my review of the draft if they want a good grade on the final version.

In addition to the final report, I require an in-class presentation.

I require the students to go beyond methods presented in class and draw on the research literature for more advanced methods. Students are expected to explain the methods they've used in their project report and presentation.

This scaffolding helps to prevent "buy a paper" cheating or other kinds of plagiarism. It also gives the student plenty of feedback to improve the project if it goes off the rails.

Answer 2

I have done similar things (more Six Sigma DOE oriented than stats only). My advice is to make a list and let the students pick from it. Asking them to pick new problems on their own is likely to come a cropper unless you have very strong students. It can be difficult to pick tractable yet interesting questions, even for Ph.D. students. Heck even for PIs. And it's a lot to expect that students will do this well when it is just a class exercise.

At a minimum, even if you disagree with me and want to force the students to do this, please force YOURSELF to make such a list. You will get some insights from manually doing this, that will help the students. And you'll get some sympathy for them.

On the practical side, I think the biggest issue is data availability. If you expect them to actually answer the question they pose.

Consider having, I donno "themes" that they can use for prodding their thinking. I sort of think about it in terms of subject matter (which implicitly means data sets). So e.g. financial/business topics (stock prices, YF data, etc.). Energy (e.g. EIA data reporting). Sports statistics. Political polling. Etc.

I would also consider "small" type insights. Not the next Nash Nobel Prize. But some kind of "brick" in the knowledge edifice. There are probably a lot of these in business/finance. (E.g. looking at a specific industry sector and running some hypothesis tests.)

P.s. I think you'd also get some good insights from the Stats SE. Maybe even better than here. Not sure the etiquette on cross posting. To Cross Validated. Haha. [Edit: Crap! Just saw that they closed your question. Grrr.]