- Review of Sierra Plots
- Updated pictures, makes better sense, more interpretable?
- Case when plateau: SE show better?
- Letter
- Alternative ways to visualize confidence intervals
- ex1
- ex2
- Distributions:
- Dividing two: ratio distribution (Cauchy Distribution special case)
- Found that always slightly above expectation?
- Multiplying two: product distribution
- Found that always slightly below expectation (but less so)
- https://en.wikipedia.org/wiki/Distribution_of_the_product_of_two_random_variables
- "extra stuff"
Meet:
- Sierra Plots
- Go back to pure coloring/scale
- 3x as much
- Smooth scaling
- Write in Python+one more
- Letter: 600 words max, AJE and Epidemiology
- Epidemiology: most recent issue, look for Kaplan Meier curves, Risk Difference function by time,
- Letter say: descriptions: turn, cite Paul, degradation of random error, generalize P-value plot to time.
- Next things:
- Attributable Fractions? Exploration/writeup: how to make inference - Bootstrap, Delta Method (data fusion),
- Lead: data repository, datasets+code book (short piece to advertise space), count as practicum
- Build contributory trials
- Daniel: Causal Consistency assumption, (710, 715),
- Baric:
- Herd immunity + interference
- Difficult without potential outcomes
- Herd immunity + interference of COVID19 (Halloran's work)
- Modern take on Herd Immunity, Interference
Main ideas:
- Therapeutic antibodies: role in controlling COVID19 pathogenesis.
- Paramyxovirus genus Henipavirus: Emerging Disease Potential.
- Follow structure pretty straightforward
-
- Review of Sierra Plots
- Questions from last time:
- Always normally distributed? NO
- Generalizable when given function, take in mean+sd+point query (aka could give uniform distribution = Twister plot)
- Not 100% vectorized
- TODO: documentation, more generalizable
- Questions:
- Confirm: risk difference has normal distribution, risk ratio has lognormal distribution/pseudo lognormal?
- What other distributions are possible/wanted?
- Next steps:
- Data Scrapping/Intro to Internet:
- HTTP requests/how is data sent over websites?
- Scrape data based on that?
- How to find elgible websites/examples? (aka not Tableau IIRC)
- Focus on survival functions!/figures/back data out from there
- Epi focus on data when deciding to collect/how to collect it (wrong timeframe type deal)
- Classes:
- Baric: material super advanced
- CS Minor?
- Pros: combine with research, capability to do so, crossover knowledge, expect to use later, fellowship opportunities?
- Cons: extended classes into 4th year realistically (1/semester * 5 semesters), one course subset not perfect
- Letter: summarizing of alternative ways to express confidence intervals, break away from interval+point paradigm/understanding?
- Expected pushback: how to decide shaded at what point?
-
- Descent of gradient sharper?
- Triple standard error for working for example
- So maybe in cases with low SE, this plot not best option?
- Make graph --> go SUPER far out (.9999 -> 4 SE), add lines for 95% confidence interval
- Increase SE for example? (Another dataset/false dataset?)
- Check another colormap
- Output as .eps file for AJE (does not support transparencies) more NOTE
- Colormap to get color specifically
- Case: plateau --> set of answers given data/design complex. If case, have black line if ridge, plateau = no line
- Attributable Fraction --> 2003 Greenland+ ---> Simulation to get CLimits. What happens with varying values of 'a'. Sharpness of probability known/unknown
- Delta Method (Taylor Series expansion other ways)
- Closed form analytic, simulation approach
- Before: assumed constant across W (same population), but what if population changes, W different in two settings (transportability/generalizability problem)
- Ultimate purpose:
- SER workshop? Deadline September 10th? NOT FOR ME.
- What are next steps needed?
- Adding more?
- Going forwards:
If time: - BIOS662 - Not getting much out of it, recorded lectures - Worth coming to meetings for an hour? - BIOS minor common --> What about CS minor? Would it be worth it?
- Trial 1: failure (misunderstood)
- Trial 2: overlapping step functions
- Alright, but relies on a couple of assumptions
- Even, normal distribution around true value and CIs
- Is this always the case? NO???
- Trial 3: shading rectangles
- Allows finer control
- Not straightforward/reproducible way to do this (imshow seems most promising)
- Trial 4: will have to be bit mapping
- Although very computationally intensive, only way to more or less guarantee
- Also allows for stronger control of shading
Next step: Seaborn heatmaps as way to go forward.