Course site: Applied Nonparametric and Modern Statistics

## 1. Introduction

A common scenario in applied statistics is that one has an independent variable or outcome $Y$ and various dependent variable or covariates $X_1,\dots,X_p$. One usually observes these variables for various “subjects”.

Statisticians usually assume that $Y$ and the $X$’s are random variables. Then one can summarize a lot of questions by asking: what is $\mathrm{E}[Y \vert X_1, \dots, X_p]$? We usually call $f(X_1, \dots, X_p) = \mathrm{E}[Y \vert X_1, \dots, X_p]$ the * regression function*.

It should be noted that for some designed experiments it does not make sense to assume the $X$’s are random variables. In this case we usually assume we have “design points” $x_{1i}, \dots, x_{pi}, i=1,\dots,n$ and non-IID observations $Y_1, \dots, Y_n$ for each design point. In most cases, the theory for both these cases is very similar if not the same. These are called the *random design model* and *fixed design model* respectively.

How do we learn about $\mathrm{E}[Y \vert X_1, \dots, X_p]$?

- Linear Regression: $\mathrm{E}[Y \vert X_1, \dots, X_p] = \sum_{j=1}^{p} X_j \beta_j$
- Generalized Linear Model (GLM): $g(\mathrm{E}[Y \vert X_1, \dots, X_p]) = \sum_{j=1}^{p} X_j \beta_j$
- $g$ is called a
. We can also write $\mathrm{E}[Y \vert X_1, \dots, X_p] = g^{-1}(\sum_{j=1}^{p} X_j \beta_j)$**link function** - It is typical to assume the conditional distribution of $Y$ is part of an exponential family, e.g. binomial, Poisson, gamma, etc.
- Many times the link function is chosen for mathematical convenience.

- $g$ is called a

Linear Models Pros:

- Having the convenience that the parameters $\beta$ usually have direct interpretation with scientific meaning.
- Once an appropriate model is in place, the estimates have many desirable properties.

Linear Models Cons:

- These models are quite restrictive. Linearity and additivity are two very strong assumptions. This may have practical consequences.
- E.g., by assuming linearity one may never notice that a covariate has an effect that increases and then decreases.

- By relaxing assumptions we loose some of the nice properties of estimates. There is an on going debate about specification vs. estimation.

In this class we will:

- Start by introducing various smoothers useful for smoothing scatter plots $\lbrace (X_i, Y_i), i = 1,\dots, n\rbrace$ where both $X$ and $Y$ are continuous variables.
- Set down precise models and outline the proofs of asymptotic results.
- Introduce local regression (
).**loess** - Examine spline models and some of the theory behind splines.
- Some smoothers are more flexible than others. However with flexibility comes variance. We will talk about the bias-variance trade-off and how one can use resampling methods to estimate bias and variance.
- After explaining all these smoothers we will make a connection between them. We will also make connections to other statistical procedures.
- We will examine the case were one has many covariates. One can relax the linearity assumption, assume additivity and use additive models. One can also forget the additivity assumption and use regression trees.
- After all this we will be ready to consider the case where is not necessarily continuous. We will generalize to this case and look at
and**Generalized Additive Models**.**Local Likelihood** - While examining all these subjects we will be considering various models for one data set. We will briefly discuss techniques that can be used to aid in the choice of such models.
- Finally we will look at a brief introduction of times series analysis.

We will begin the class talking about the case were the regression function $f$ will depend on a single, real-valued predictor $X$ ranging over some possibly infinite interval of the real line, $I \subseteq \mathbb{R}$. Therefore, the (mean) dependence of $Y$ on $X$ is given by

The data to support such investigations are typically a set of $n$ paired observations $(X_1, Y_1), \dots, (X_n,Y_n)$. These can be either a random sample of the joint distribution of $(X,Y)$ as is the case for *observational studies*, or fixed input values $\lbrace x_i \rbrace$, arising perhaps from a *designed experiment*.

So once we have the data what do we do?

If we are going to “model” $(\ref{eq1.1})$, we gain insight into the important features of the relationship between $Y$ and $X$ by entertaining various descriptions of or models for $f$. Through this exercise we might:

- identify the width and height of peaks
- explore the overall shape of $f$ in some neighborhood
- find areas of sharp increase or regions exhibiting little curvature.

We will then move on to the case where we have many covariates, then cases where the expectation needs to be transformed, and various other generalization.