fixedpoint.jp




2018/06/23

How to estimate a location parameter when average fails to converge

It is sometimes overlooked that increasing the sample size will not always reduce the uncertainty in the estimate of a parameter. [1] illustrates this fact as "Misconception 4" with an example considering a random variable having Cauchy distribution. Its probability density function \(f\) has two parameters: \begin{align}f(x;x_0, \gamma) = \frac{1}{\pi \gamma \big[1 + (\frac{x - x_0}{\gamma})^2\big]}\end{align} where \(x_0\) is a real number called the location parameter and \(\gamma (> 0)\) is its scale parameter.

Since the distribution has no (finite) mean, the law of large numbers cannot assure that taking its sample average converges to \(x_0\). Fortunately it is still possible to estimate \(x_0\) by taking sample median. In fact, the following computational experiment in R code finds sample mean converging to the location parameter quickly:

x <- replicate(16, {
    s <- rcauchy(10000, 42)
    h <- head(s, 10)
    c(median(h), median(s), mean(h), mean(s))
})
boxplot(t(x), names = c("median\n(10 samples)", "median\n(10^4 samples)", "mean\n(10 samples)", "mean\n(10^4 samples)"))

The resulting figure (obtained with seeding set.seed(1)) shows that the error of sample median from its true value \(42\) is small, while average of 10 samples is more diverse and averaging 10,000 samples often fails in even bigger errors.

Figure to show the sample median/mean of a random variable whose distribution is Cauchy

Reference: [1] John Kermond, Counterexamples in Probability and Statistics, MAV Annual Conference 2009.

#permalink

Archives

2018: May / Apr / Mar / Feb / Jan

2017: Dec / Nov / Oct / Sep / Aug / Jul / Jun / May / Apr / Mar / Feb / Jan

2016: Dec / Nov / Sep / Aug / Jul / Mar / Feb / Jan

2015: Dec / Oct / Sep / Jul / May / Mar / Feb

2014: Dec / Nov / Oct / Aug / Apr / Mar / Feb / Jan

2013: Dec / Nov / Oct / Sep / Aug / Jun / May / Apr / Mar

2012: Nov / Oct / Sep / Jul / Jun / May / Mar / Feb / Jan

2011: Dec / Nov / Oct / Sep / Jul / Jun / Apr / Mar / Feb / Jan

2010: Dec / Nov / Oct / Sep / Aug / Jul / Jun / May / Apr / Mar / Feb / Jan

2009: Dec / Nov / Oct / Sep / Aug / Jul / Jun / May / Apr / Mar / Feb / Jan

2008: Dec / Nov / Oct / Sep / Aug / Jul / Jun / May / Apr / Mar / Feb / Jan

2007: Dec / Nov / Oct / Sep / Aug / Jul / Jun / May / Apr / Mar / Feb / Jan

2006: Dec / Nov / Oct / Sep


© 2006-2018 fixedpoint.jp