Scribd
Upload
41 views

1.017/1.010 Class 14 Estimation: Estimating Distributional Properties

Parametric statistics assumes that the form of the x distribution of x is known. Nonparametric statistics makes no assumptions about distribution. How do we know if this is a "good" estimator?

Uploaded by

Dr. Ir. R. Didin Kusdian, MT.
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
41 views

1.017/1.010 Class 14 Estimation: Estimating Distributional Properties

Parametric statistics assumes that the form of the x distribution of x is known. Nonparametric statistics makes no assumptions about distribution. How do we know if this is a "good" estimator?

Uploaded by

Dr. Ir. R. Didin Kusdian, MT.
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

1.017/1.

010 Class 14
Estimation

Estimating Distributional Properties

We can use random samples to estimatethe unknown properties of


random variable x. These may be:

1. Distributional parameters, such as the lower and upper limits of a


uniform distribution (a and b)
2. Other distributional properties such as the mean, variance, 90th
percentile value, etc.

Parametric statistics assumes that the form of the x distribution of


x is known. Nonparametric statistics makes no assumptions
about distribution of x.

Estimators

An Estimator (or a statistic) aˆ ( x1 , x 2 ,..., x N ) is a function used to derive


an estimate â of the unknown distributional property a from a random
sample x1, x2,..., xN

aˆ = aˆ ( x1 , x 2 ,..., x N )

Example:

Suppose we want to estimate from the random sample x1, x2,..., xN


the mean of Fx(x). A possible choice for the estimator is the sample
mean mx . Then:

a = E[x]
N
1
aˆ = aˆ ( x1 , x 2 ,..., x N ) = m x =
N ∑x
i =1
i

How do we know if this is a “good” estimator?

Properties of Good Estimates

We would like the estimate aˆ to be as “close” as possible to the true value a.

1
Since aˆ is derived from a random sample (a set of random variables) it is also a
random variable, defined by the probability distribution Faˆ (aˆ ) . This distribution
describes how much â might deviate from a.

We can sometimes derive Faˆ (aˆ ) or we may use the Central Limit theorem to
justify the assumption that Faˆ (aˆ ) is a normal distribution (if a is derived from a
large random sample).

If Faˆ (aˆ ) is normal, it is completely defined by the mean E[aˆ ] and variance
Var[aˆ ] of aˆ .

Even if Fâ (aˆ ) is not normal we can use the mean and variance of â to
measure the quality of the estimate. “Good” estimates should be
unbiased and consistent:

1.Unbiased
The expected value of the estimate is the true property value:

E[aˆ ] = a

2. Consistent:
The variability of the estimate goes to zero as the sample size
increases:

Var[aˆ ] → 0 as N →∞

If the estimate is unbiased and consistent it will converge to the true value
as the sample size increases.

In general, the mean and variance of an estimate are most readily


estimated with stochastic simulation (see exercise below). In some
cases they can be derived directly from the estimator function â .

Example – Properties of the Sample mean:

Reconsider the sample mean mx used to estimate the mean E[x] of Fx(x):

N
1
aˆ = aˆ ( x1 , x 2 ,..., x N ) = m x =
N ∑x
i =1
i

The mean and variance of this estimate are:

1 N

E[aˆ ] = E[m x ] = E 
N
∑
i =1
x i  = E[ x ] = a

2
1 N
 1
Var[aˆ ] = Var[m x ] = Var 
N
∑ x  = N Var[ x]
i =1
i

So the sample mean is an unbiased and consistent estimator of the


population mean.

For large samples the Central Limit Theorem implies that mx is normally
distributed. The distribution of mx is centered on E[x] and gradually
converges to E[x] as N increases.

Exercise: Comparing Alternative Estimates of Population Properties

Suppose that you have a sample of 10 observations of a random variable


x that you believe to be exponentially distributed. Your objective is to
estimate the 90 percentile value x90 of this variable. This value of the
solution of the equation Fx(x90) = 0.9 [i.e. Fx-1(0.9) = x90]

Propose at least two different methods for estimating x90 from the 10
observations.

Compare the performance of these alternative estimators with a stochastic


simulation that performs the following steps:

1. Generate many (e.g. 1000) replicates, each consisting of 10 observations


drawn from an exponential distribution. Select your own value of the
distributional parameter a = E[x].
2. For each replicate derive an estimate x̂90 of x90 from each of your two
proposed estimators.
3. For each estimator compute the sample mean and variance of x̂90 over all
replicates. Also, use the replicates to construct a histogram and a CDF
plot (using MATLAB’s normplot function) for each x̂90 .
4. Determine whether your estimators are unbiased and consistent (check
consistency by plotting the rerunning your simulation for a much larger
number of observations).

Which of your estimators is better? Explain your reasoning.

Some relevant MATLAB functions: exprnd, hist, normplot, mean,var

Copyright 2003 Massachusetts Institute of Technology


Last modified Oct. 8, 2003

You might also like