Is there a theoretically-sound way to perform propagation of errors with robust statistics? I am trying to characterize the errors inherent in a measurement and propagate the uncertainty through calculations involving that estimate.

In the past, I have assumed that my errors are normally distributed and then used uncertainty propagation. However, this is clearly a bad assumption (my data is symmetric but guaranteed to have significant outliers), and I am intrigued by using a measure like Median Absolute Deviation.

Is it valid to treat MAD like a standard deviation? Is there a different measure I should be using for this application?

  • $\begingroup$ @Glen_b has answered the question I asked, but I've realized that I didn't frame the question in a way that would get at what I'm really trying to figure out. I will try again in a separate question, unless somebody tells me that goes against site norms. $\endgroup$ – user670416 Mar 26 '13 at 2:10

Beware your acronyms; MAD has been used for more than one thing.

I assume you mean 'median absolute deviation from the median'; speak up if you meant one of the other MADs.

Variances (and hence s.d.'s) have some nice properties. e.g. sums of r.v.s have nice simple variance results.

MAD doesn't have a result like that.

Consider $X_1, X_2$ both take the value $-1$ with probability $1/3$ and $1$ with probabiity $2/3$, while $Y_1$ and $Y_2$ take the values of $-1$, $0$ and $1$ with probability $1/3$, and $Z_1$ and $Z_2$ take the values of $-1$ and $1$ with probability 0.1 each and $0$ with the remaining probability (0.8). All variables are independent.

The $X$ and $Z$ variables have a MAD of 0. The $Y$'s have a MAD of 1.

$X_1 + X_2$ has a MAD of 2. (0,0) -> 2

$Z_1 + Z_2$ has a MAD of 0. (0,0) -> 0

$Y_1 + Y_2$ has a MAD of 1. (1,1) -> 1

There's clearly no such neat relationship.

MAD's of linear rescalings will scale - MAD(2X+3) = 2 MAD(X), but most other manipulations don't behave in neat ways.

  • $\begingroup$ Yes - I had been referring to "median absolute deviation from the median". $\endgroup$ – user670416 Mar 26 '13 at 0:03
  • $\begingroup$ Great, no problem then. $\endgroup$ – Glen_b Mar 26 '13 at 0:04
  • $\begingroup$ (sorry - accidentally posted incomplete comment) You've shown that it will not work for my case - is there another statistic that I should use that will allow me to produce error estimates on a linear combination of my variables? $\endgroup$ – user670416 Mar 26 '13 at 0:06
  • $\begingroup$ 'Robust' implies 'nonlinear' ... so I wouldn't generally expect linear combinations to have robust-scale-measures that are neat functions of the scale-measures of the components. Quantiles don't work that way for example, so doing complicated things with functions of quantiles certainly wouldn't be expected to. $\endgroup$ – Glen_b Mar 26 '13 at 0:09
  • $\begingroup$ OK, I think in my ignorance of statistics I'm asking the wrong question. Maybe a better phrasing would be: "Given that I need to provide error estimates of some form for a linear combination of these variables, and I know that they are not normally distributed, what is my best option?". I can give more information about the application if that would be helpful. The standard approach in my field's literature seems to be the head-in-sand, pretend that everything's gaussian and just report standard deviation. $\endgroup$ – user670416 Mar 26 '13 at 0:15

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.