Resampling, the jackknife, and pseudo-observations

Resampling methods approximate the sampling distribution of a statistic or estimator. In essence, a sample taken from the population is treated as a population itself. A large number of new samples, or resamples, are taken from this “new population”, commonly with replacement, and within each of these resamples, the estimate of interest is re-obtained. A large number of these estimate replicates can then be used to construct the empirical sampling distribution from which confidence intervals, bias, and variance may be estimated. These methods are particularly advantageous for statistics or estimators for which no standard methods apply or are difficult to derive.

The jackknife is a popular resampling method, first introduced by Quenouille in 1949 as a method of bias estimation. In 1958, jackknifing was both named by Tukey and expanded to include variance estimation. A jackknife is a multipurpose tool, similar to a swiss army knife, that can get its user out of tricky situations. Efron later developed the arguably most popular resampling method, the bootstrap, in 1979 after being inspired by the jackknife.

In Efron’s (1982) book The jackknife, the bootstrap, and other resampling plans, he states,

Good simple ideas, of which the jackknife is a prime example, are our most precious intellectual commodity, so there is no need to apologize for the easy mathematical level.

Despite existing since the 1940’s, resampling methods were infeasible due to the computational power required to perform resampling and recalculate estimates many times. With today’s computing power, the uncomplicated yet powerful jackknife, and resampling methods more generally, should be a tool in every analyst’s toolbox.

Continue reading Resampling, the jackknife, and pseudo-observations

Parametric vs. Nonparametric Approach to Estimations

Parametric statistics assume that the unknown CDF F belongs to a family of CDFs characterized by a parameter (vector) \theta. As the form of F is assumed, the target of estimation is its parameters \theta. Thus, all uncertainty about F is comprised of uncertainty about its parameters. Parameters are estimated by \hat{\theta}, and estimates are be substituted into the assumed distribution to conduct inference for the quantities of interest. If the assumed distribution F is incorrect, inference may also be inaccurate, or trends in the data may be missed.

To demonstrate the parametric approach, consider n = 100 independent and identically distributed random variables X_1, …, X_n generated from an exponential distribution with rate \lambda = 2. Investigators wish to estimate the 75^{th} percentile and erroneously assume that their data is normally distributed. Thus, F is assumed to be the Normal CDF but \mu and \sigma^2 are unknown. The parameters \mu and \sigma are estimated in their typical way by \bar{x} and \sigma^2, respectively. Since the normal distribution belongs to the location-scale family, an estimate of the p^{th} percentile is provided by,

    \[x_p = \bar{x} + s\Phi^{-1}(p)\]

where \Phi^{-1} is the standard normal quantile function, also known as the probit.

library(tidyverse, quietly = T)
# Generate data from Exp(2)
x <- rexp(n = 100, rate = 2)

# True value of 75th percentile with rate = 2
true <- qexp(p = 0.75, rate = 2) 
## [1] 0.6931472
# Estimate mu and sigma
xbar <- mean(x)
s    <- sd(x)

# Estimate 75th percentile assuming mu = xbar and sigma = s
param_est <- xbar + s * qnorm(p = 0.75)
## [1] 0.8792925

The true value of the 75^{th} percentile of \text{Exp}(2) is 0.69 while the parametric estimate is 0.88.

Nonparametric statistics make fewer distributions about the unknown distribution F, requiring only mild assumptions such as continuity or the existence of specific moments. Instead of estimating parameters of F, F itself is the target of estimation. F is commonly estimated by the empirical cumulative distribution function (ECDF) \hat{F},

    \[\hat{F}(x) = \frac{1}{n} \sum_{i=1}^{n} \mathbb{I}(X_i \leq x).\]

Any statistic that can be expressed as a function of the CDF, known as a statistical functional and denoted \theta = T(F), can be estimated by substituting \hat{F} for F. That is, plug-in estimators can be obtained as \hat{\theta} = T(\hat{F}).

Continue reading Parametric vs. Nonparametric Approach to Estimations

Kernel Regression


For observed pairs (x_i, y_i), i = 1, …, n, the relationship between x and y can be defined generally as

    \[ y_i = m(x_i) + \varepsilon_i \]

where f(x_i) = E[y_i | x = x_i] and \varepsilon_i \stackrel{iid}{\sim} (0, \sigma^2). If we are unsure about the form of m(\cdot), our objective may be to estimate m(\cdot) without making too many assumptions about its shape. In other words, we aim to “let the data speak for itself”.

Simulated scatterplot of y = f(x) + \epsilon. Here, x \sim Uniform(0, 10) and \epsilon \sim N(0, 1). The true function f(x) = sin(x) is displayed in green.

Non-parametric approaches require only that m(\cdot) be smooth and continuous. These assumptions are far less restrictive than alternative parametric approaches, thereby increasing the number of potential fits and providing additional flexibility. This makes non-parametric models particularly appealing when prior knowledge about m(\cdot)‘s functional form is limited.

Estimating the Regression Function

If multiple values of y were observed at each x, f(x) could be estimated by averaging the value of the response at each x. However, since x is often continuous, it can take on a wide range of values making this quite rare. Instead, a neighbourhood of x is considered.

Result of averaging y_i at each x_i. The fit is extremely rough due to gaps in x and low y frequency at each x.

Define the neighbourhood around x as x \pm \lambda for some bandwidth \lambda > 0. Then, a simple non-parametric estimate of m(x) can be constructed as average of the y_i‘s corresponding to the x_i within this neighbourhood. That is,

(1)   \begin{equation*} \hat{f}_{\lambda}(x) = \frac{\sum_{n} \mathbb{I}(|x - x_i| \leq \lambda)~ y_i}{\sum_{n} \mathbb{I}(|x - x_i| \leq \lambda)} = \frac{\sum_n K\left( \frac{x - x_i}{\lambda} \right) y_i}{\sum_n K\left( \frac{x - x_i}{\lambda} \right) } \end{equation*}


    \[ K(u) = \begin{cases} \frac{1}{2} & |u| \leq 1 \\ 0 & \text{o.w.} \end{cases} \]

is the uniform kernel. This estimator, referred to as the Nadaraya-Watson estimator, can be generalized to any kernel function K(u) (see my previous blog bost). It is, however, convention to use kernel functions of degree \nu = 2 (e.g. the Gaussian and Epanechnikov kernels).

The red line is the result of estimating f(x) with a Gaussian kernel and arbitrarily selected bandwidth of \lambda = 1.25. The green line represents the true function sin(x).

Continue reading Kernel Regression

Kernel Density Estimation


It is important to have an understanding of some of the more traditional approaches to function estimation and classification before delving into the trendier topics of neural networks and decision trees. Many of these methods build on an understanding of each other and thus to truly be a MACHINE LEARNING MASTER, we’ve got to pay our dues. We will therefore start with the slightly less sexy topic of kernel density estimation.

Let X be a random variable with a continuous distribution function (CDF) F(x) = Pr(X \leq x) and probability density function (PDF)

    \[f(x) = \frac{d}{dx} F(x)\]

Our goal is to estimate f(x) from a random sample \lbrace X_1, …, X_n \rbrace. Estimation of f(x) has a number of applications including construction of the popular Naive Bayes classifier,

    \[ \hat{Pr}(C = c | X = x_0) = \frac{\hat{\pi}_c \hat{f}_{c}(x_0)}{\sum_{k=1}^{C} \hat{\pi}_{k} \hat{f}_{k}(x_0)} \]

Continue reading Kernel Density Estimation