Introduction to visualization+modeling+computing (VMC) I

Stat 221, Lecture 2

@snesterko

More on final project

  • Team assignment details.
  • Descriptions released approx. February 2.
Harvard Statistics DepartmentSiemensMITEbayHubwayMAPCSense PlatformIBMdiffeoTREC KBADeloitteStarbucksAthena HealthCaesars EntertainmentRisk Management SolutionsThe Boston Consulting GroupNationwide Insurance

Student questions, administrative

  • Can submit visualizations to class website! Good for participation.
  • What is good style of code?
  • Homework previews.
  • Time commitment to problem sets.
  • d3js support possibilities (online tutorials, workshop, TF)
  • T-shirt competition clarified.

Modeling

  • Models - practical approximations of the real world
  • Important: visual representation vs. model itself

Many areas apply modeling

  • Statistics
  • Computer science
  • Ecomonics
  • Physics
  • Biology
  • others

Statistical models

  • Close to physics, chemistry, biology etc.
  • Defining trait - distributional assumptions on noise.
  • Similar to Computer Science when no clear intuition driving the process can be found.

Main ingredients

Response, data, noise, and parameters:$$\begin{align} \rr{response} &= f(\rr{data}, \rr{noise})\\ Y &= f_{\theta}(X, \epsilon) \end{align}$$

Ways to write it down

Representation:$$\begin{align}Y &= f_{\theta}(X, \epsilon) \rr{ or} \\ Y &\sim f_{\theta}(X, \epsilon)\end{align}$$

Graphical or generative :

Representation examples

Linear regression:$$\vec{Y} \sim \mathbf{X}\vec{\beta} + \vec{\epsilon}, \rr{ } \vec{\epsilon} \sim N\left(0, \sigma^2\mathbf{I}\right)$$

Anchor Process: $$\begin{align} Y_i \given Y_{r(i)} & \sim N \left(\rho Y_{r(i)} + (1 - \rho) \mu, (1 - \rho^2) \sigma^2 \right), \\Y_i & \sim N\left( \mu, \sigma^2 \right) \rr{ if } r(i)\rr{ doesn't exist}\end{align}$$

How do you design these?

  • Translate scientific/expert intuition into mathematical statements.
  • Follow the quantitative analysis workflow.

Models as stories

  • Each model tell a story.
  • Let's tell a story - class activity.

The workflow

Iterate between the stages:

  1. Initialize: talk to collaborator/client, get the data, understand the question of interest.
  2. Design: an analytic method for a solution, evaluate its properties and assumptions.
  3. Implement: a working, practical computing solution to perform the designed procedure.
  4. Communicate: the findings, help understand and interpret them. Advise on the method's applicability in case of repeated use.

Case study: RDS

  • Respondent-Driven Systems/Sampling
  • Respondents follow their social network to recruit peers to achieve a common goal.
  • Public health agencies are interested in estimating population average income, illness status etc.

How RDS looks

The first guess

The first guess for the model is $$Y_i \mathop{\sim}^{\rr{iid}} N\left( \mu, \sigma^2\right)$$

Yields the classic vanilla estimator $$\hat{\mu} = \bar{Y}$$

with liberal (tight) confidence bounds.

Almost no computation is needed for this model.

A problem

  • The model is unbiased, but
  • The naive 95% confidence interval has coverage of around 20% in case of homophily.
  • How do we enhance our model to keep the unbiasedness, but achieve the right coverage rate for uncertainty intervals?

An approach to solve it

  • Understand the system better.
  • Build the intuition about system mechanics into our estimation model.
  • Hope that our intuition and the way we've postulated it are good enough.

An RDS sample

Distribution+network (with homophily)

Two variants of population networks

RDS with homophily

Anchor Process model

Model statement: $$\begin{align} Y_i \given Y_{r(i)} & \sim N \left(\rho Y_{r(i)} + (1 - \rho) \mu, (1 - \rho^2) \sigma^2 \right), \\Y_i & \sim N\left( \mu, \sigma^2 \right) \rr{ if } r(i)\rr{ doesn't exist}\end{align}$$

  • \\( \{ \rho, \mu, \sigma^2 \} \\) is the parameter set
  • \\( r(i) \\) is the index of referrer of respondent \\( i \\).
  • \\( Y_i \\) is the observation of respondent \\( i \\).
  • Needs more involved computation - there is no closed-form solution for \\( \hat{\mu} \\).

AP model performance

AP model simulation

What drives the process?

  • The question to be answered a.k.a. the estimand.
  • Developing the intuition of how things work.
  • Translating the intuition into models, evaluating their properties.
  • Being able to fit/compute with these models.
  • Reporting the results.
  • Looping back.

Announcements

  • Final project proposals available soon - check email and course Twitter!
  • Odyssey access, ask your TF for clarifications.
  • Free T-shirt competitions start now!

Resources

Final slide

  • Next lecture: Introduction to data Visualization + statistical Modeling + Computing (VMC) II
  • Don't hesitate to talk to course staff about the class and your specific needs.