Applied Statistical Methods - Final Project

The final project is meant to encourage you to explore possible research opportunities on campus. It should be related to an existing research project that has a clear and relevant research question with existing data. An ideal example would be a paper that examines a public dataset that could be improved or a methodology that could be applied to a new domain. A bad example would be the diamond dataset on kaggle that does not solve any real problems.

For purely exploratory projects, you should propose what the sufficient dataset, based on your pilot data, is to answer the scientific question (similar to a sample size calculation). Other forms of formal statistical analysis is possible but please discuss with the instructor to make sure your work will suffice as a final project.

You are encouraged to use your existing research, in another class, in your lab, or internship, for this project. The only exception is an interview data challenge.

The final project in this class is open-ended with a few requirements. Key deliverables are

• An oral presentation (8-20 minutes) + Q&A that you sign up here
• This will be done in Rm 614 Watson Hall (612 W 115th St)

Presentation Elements

• The project must identify a research question that data can possibly falsify (10 pts)
  • Articulate the question, e.g. are there more or fewer illegal drug activities in states where neighboring states have legalized it? Or do startup founders who are minorities get asked qualitatively different questions from investors?
    • Feel free to talk to the instructor/TA if you want to identify a potential problem that’s fun to you.
    • If you are improving a paper, articulate why the approach needs improving or why new fields could benefit.
  • Articulate the ideal dataset you would use to answer this question if money and time were not a problem. For example, all illegal activities related to drugs and the dates of the legalization of marijuana for each state.
    • You should not limit your question so that the obtainable dataset is ideal, e.g. I want to know if AirBnB prices have risen more in 2020 than in 2023 should be formulated as have short term rental prices risen more in 2020 than in 2023.
  • Articulate why the ideal dataset and problem are related. Specifically, why the dataset would falsify one hypothesis but not another?
  • Identify an obtainable dataset that somewhat captures this information
    • Describe the data source, e.g. Crime Data Explorer by the Gederal Bureau of Investigation could be a possible data source for the criminal records.
  • Articulate the difference between this dataset and your ideal dataset, e.g. arrests do not capture all illegal activities and arrests are biased against minority neighborhoods.
  • Articulate how this dataset is still relevant to the original question.
• Examine the dataset (10 pts)
  • A summary of the data source: its time period, its coverage, its size, the notable attributes, the context around it, etc. For example, arrests per state should be reported in the context of the state populations.
  • The original intent of the dataset. Many public datasets are simply records with no clear purposes but some datasets like the FICO score was meant to assess someone’s ability to repay their debts (creditworthiness).
  • Report any anomalies you notice in the data source.
    • Please articulate if the anomalies can impact your final analysis or not.
  • The data source should come from the entity that manages the data. Management is defined by the entity that is responsible for the data quality. For example, data aggregators like Kaggle do not manage the data it posts on its challenges but Twitter manages Twitter data. Some exceptions exist like platforms like NYC OpenData allow you to report data quality issues and will be considered a manager of the data.
  • Please describe the critical processes/decisions you made to the dataset. This should not include every possible operation but critical manipulations that could impact the outcome, e.g. aggregation, dealing with missing values, etc.
  • Create at least one graphical summary that informs you about the problem, e.g. arrest rates over time with lines that highlight the date of marijuana legalization of neighboring states. Misleading graphs that do not have a message related to the main question will receive a penalty.
• Analysis (15 pts)
  • Articulate your model. Your model should be a formal definition that will tie your question to the data. E.g. to understand how genetics affects height, Fisher modeled sons’ heights ($Y$) as a linear function of fathers’ heights ($X$) plus some error ($\epsilon$) that accounted for inaccurate measurements and influences outside of the fathers’ genetics, i.e. $Y_{son i’s\ height} = X_{father i’s\ height}\beta + \epsilon$, the error is assumed to be symmetric around 0. The coefficient for fathers’ heights ($\beta$) and the magnitude of the error term ($\SD(\epsilon)$) will tell us how strong genetics are at influencing their sons’ height.
  • What type of analysis would you do to make your case? e.g. do you need to fit a model to a wide enough dataset to convince investors about your superior approach?
    • Your results do not need to be conclusive
  • What are you expect to see from the analysis? e.g. did you expect a rise or drop in arrests from your understanding of the problem? Did you have a guess on the magnitude?
  • Describe how you implemented your analysis, e.g. packages used, tuning parameters, etc
  • If applicable, are you addressing the uncertainty in your analysis?
  • Validate your model or assumptions if possible for your analysis
    • e.g. does your model fit to the data?
    • e.g. can you run a simulation that reproduces similar looking data?
• Putting your model in context, there are several approaches to this given your problem but here are a few options to consider:
  • Comparing your model to a baseline.
    • Your baseline model should be a “reasonable but stupid” model that is possibly wrong.
    • You should compare your model performance relative to this baseline. A common example is comparing deep learning predictions relative to logistic regression predictions.
    • You should verify why your model outperformed or did not outperform the baseline model.
  • Simulating your strong assumption
    • If you have a strong assumption, it’s sometimes important to simulate the cases when the assumption is not met and how your model fitting and results would change. A common one is when we ignore the spatial correlation when doing prediction.
  • Test the robustness of the model.
    • Perturbing the data, e.g. resampling the data with replacement and seeing how the model changes
    • Perturbing the model, e.g. fitting a different ARIMA model or changing tuning parameters of the model.
• Conclusion (5 pts)
  • What was your conclusion given your original question and analysis?
    • e.g. did the results match your expectations?
    • e.g. did you have to change your question given constraints you learned in the data?
    • What else could you do?
• Code on GitHub (5 pts) - graded by completion
  • Code (or steps) used to extract the data
  • Code used to process the data
  • Code used to create all summary statistics and graphs
  • Code used to do the analysis
  • Your code should be readable and avoids lots of “copying/pasting” code