I am trying to understand “higher order asymptotics”. I find several texts on Likelihood asymptotics, nothing’s easy to read... if you have any nice pointers on this direction, I’ll be interested; however my main question follows. The following ”roadmap“ to higher order asymptotics seems natural to me, and seems something to do and understand before looking at likelihood theory, but I don’t find anything on these lines:

Consider $T = f( X )$ where $X$ is a random vector of known distribution; let's say, for simplicity, that $X$ is normal, or multinomial (or the concatenation of several independent multinomials). The Delta method tells me that asymptotically, $T$ is normal, and how to compute its mean and variance (using linear approximation of $f$). I think one can find a better approximation of the distribution of $T$ by

  • computing the first moments of $T$ (using linear, or quadratic, or higher order approximations of $f$)
  • ”finding“ a distribution with the same first moments plus some other “parsimony” criteria to ensure uniqueness (”finding” is not well-defined: at least, being able to evaluate it numerically)

Is that possible? Do you know any textbooks/lecture/article going in that direction?

Edit fg nu gave me some pointers for the second step, that lead me to Edgeworth series. A few references:

The first step is rather elementary, however any good pointer is still appreciated.

  • 4
    $\begingroup$ A highly readable introduction to this, from an econometric perspective is given in Rothenberg (1984). $\endgroup$ – tchakravarty Dec 18 '12 at 13:17
  • $\begingroup$ @fgnu Please make an answer from your comment, I’ll accept it if nobody comes with a transcendent link! (I am currently reading Rothenberg and I enjoy it) $\endgroup$ – Elvis Dec 18 '12 at 20:32

The higher order asymptotics books that I have on my shelf are Barndord-Nielsen and Cox, Brazzale, Davison and Reid, and Young and Smith (the very latter was my dissertation adviser, and I think he has a great ability to explain very complex concepts in a reasonably understandable way; his review of non-standard problems in likelihood inference is a must second reading in asymptotics, although it is nearly impossible to get short of asking him directly). The ultimate reference on Edgeworth expansions is probably Peter Hall's bootstrap book (1995). I would have to say that Rothenberg's chapter recommended by fg nu may be beating any of them in terms of clarity. Some books, like Rencher's Multivariate Analysis, just put Bartlett corrections everywhere without explaining much of them, but they motivate it as a small sample correction.

  • 1
    $\begingroup$ Many thanks for these references. I think I’ll check the Young and Smith — I am not proud of this, but I have a copy of this book that I didn’t read entirely. The first chapters were a delight, though. It seems that I needed your advice to go back to this... shame on me. $\endgroup$ – Elvis Dec 19 '12 at 14:33
  • $\begingroup$ That's fine, I read at most one chapter from it [blushing] $\endgroup$ – StasK Dec 19 '12 at 15:30
  • 1
    $\begingroup$ Here is a link to Smith’s Survey of nonregular problems, I didnt find it immediately. I add this here because it can interest others. $\endgroup$ – Elvis Dec 19 '12 at 15:35
  • 1
    $\begingroup$ It's on his webpage, see stat.unc.edu/postscript/rs/ISI89.pdf $\endgroup$ – StasK Dec 19 '12 at 15:39
  • $\begingroup$ I edited my comment while you were typing yours... :) $\endgroup$ – Elvis Dec 19 '12 at 15:41

In addition to the other answer, a simpler book to read about this (and newer) is maybe Ronald Butler: "Saddlepoint approximations with applications". I guess it would be difficult to find a more elementary treatment than this!

As others mentioned Edgeworth approximations: Saddlepoint approximations can be seen as a tilted version of the Edgeworth approximation. The edgeworth approximations is more exact at the mean of the distribution we want to approximate. If we want to approximate the distribution at some point $x$ other than the mean, we can "tilt" the distribution by multiplying it with some exponential factor, so moving the mean to $x$, and using the Edgeworth approximation there. That is the saddlepoint approximation.

For more about the saddlepoint approximation see this post: How does saddlepoint approximation work?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.