Analysis

[fabiangunzinger]

Steps to final first draft

• replicate fig1 in diclimente2018sequences

• If we use tag_spend as main var, use PCA as category controls

• Understand entropy differences

  • Typical behaviour of high/low entropy cases for smooth and unsmooth

  • Correlation graphs

    • Grocery entropy

    • Age

    • Month income

    • Month spend

    • Number of children (boxplot)

    • Income variability

• Calculate additional outcomes and entropy vars (see methods)

• Produce additional results

• Estimation results

  • [ ] If time, use Bayesian stats, in any case, show plot rather than table for main result (use modelsummary)

  • [ ] Use PCA for spend tags as robustness

[fabiangunzinger]

Choice of dependent variable - savings only or financial outcomes (saving, spending, od-fees)

Issue:

• My instinct is to focus only on savings. Reasons: it makes for a clear and simple story, and we can sell the paper as (one of?) the first to focus on short-term savings and as thus making a clear contribution to new knowledge. While we can still claim that if we also look at other outcomes, the paper would loose focus.
• One reason to look at other outcomes is that it would be good to have more than only one dependent variable. But we can look at different savings variables: For user-month-level analysis: binary indicator for savings in given month, amount of gross and net savings. For user-level analysis (done later), average monthly amount saved, whether user makes regular savings, number of months with savings per year (count outcome).
• Another reason to also look at overdraft fees is that we might want to discuss the implications of smoothing with reference to muggleton2020evidence, who study negative financial outcomes including od fees. Yet if we do that, then I think we should do that separately. Could be a short note or comment somewhere (equivalent of AER P&P).

Suggestion:

• Make focus on paper effect of spend profile on savings behaviour.
• Discuss details of smoothed and unsmoothed entropy with regard to muggleton2020evidence separately (does Neil have suggestion for suitable outlet?)

Decision:

• Focus on savings.
• Write up findings about entropy if it seems worth it once we understand the difference between smoothed and unsmoothed one.

[fabiangunzinger]

Scaling of entropy variables - standardising vs normalising

Issue:

• Social scientist often standardise variables so they can interpret regression coefficients as standard deviation changes. This is what muggleton2020evidence did.
• Econophysicists (guidotti2015behavioural, krumme2013predictability) normalise entropy - dividing it by log(num_cats), which is the max entropy score, so scores are expressed as proportion of max entropy.
• I find the latter very appealing and more useful for interpreting plots and raw scores, but worry that social science editors might not like it.

Suggestion:

• Normalise and just change in case requested to do so.

Decision:

• tbd with Neil

[fabiangunzinger]

Entropy scores to calculate

Choice dimensions:

• Category {tag, auto tag, merchant}
• Statistic {count, value}
• Time component (wknd vs weekdays) {yes, no}
• Smooth {yes, no}

Choices: