Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Bias: Monte Carlo

From   Stas Kolenikov <>
To   "" <>
Subject   Re: st: Bias: Monte Carlo
Date   Mon, 6 May 2013 10:04:02 -0500

Continuing on Maarten's note, my -bsweights- package produces
first-order balanced bootstrap weights that remove the random
simulation error from the bootstrap results for the point estimates.
See and
references on the balanced bootstrap therein.

I think the bias in parameter estimates should be put into the context
of MSE: if the bias component is greater than the variance component,
then the estimator is relatively useless. If the magnitude of bias is
say 1/2 that of the standard deviation of the sampling distribution of
your estimator, so that the contribution of bias to the total MSE is
20%, I would personally be able to live with that. With bias this big,
though, your estimation procedure should report the standard errors
that are based on MSE, not on the variance alone. For lots of
"regularly behaving" estimators, the sampling standard deviation
(estimated by the standard error) is O(n^{-1/2}), and bias is
O(n^{-1}), going to zero faster with the sample size than the standard
deviation, so in large samples, the bias asymptotically disappears.
Then the question of "how large bias is tolerable" becomes the
question of "how large my sample size should be" (for the normal
approximations to make sense).

Finally, the % bias is a awful measure when the true value of the
parameter is zero (and the whole situation is shift invariant, as in
the mean of the normal distribution, so the initial point on the scale
is arbitrary). It's probably OK in the constant CV situation of
heavily skewed distributions, but may not make sense in other

-- Stas Kolenikov, PhD, PStat (SSC)
-- Senior Survey Statistician, Abt SRBI
-- Opinions stated in this email are mine only, and do not reflect the
position of my employer

On Mon, May 6, 2013 at 3:25 AM, Maarten Buis <> wrote:
> On Mon, May 6, 2013 at 9:49 AM, John Antonakis wrote:
>> I am running some Monte Carlos where I am interested in observing the bias
>> in parameter estimates across manipulated conditions. By bias I mean the
>> absolute percentage difference of the simulated value from the true value.
>> I was wondering whether there has been another written about how much bias
>> is "acceptable"--I know that this is like asking how long is a piece of
>> string and that there is no statistical fiat that can give a definitive
>> answer, because it also is a very field specific issue.
> It is probably not quite the answer you are looking for (and I think
> you are right by wondering whether such an answer can exist), but one
> thing you can do is take into account that a Monte Carlo experiment
> contains a random component, so if you repeat the experiment (with a
> different seed) you will get a slightly different estimate of your
> bias. The logic behind this variation between Monte Carlo experiments
> is pretty much the same as the logic behind statistical testing: so
> you can compute standard errors and confidence intervals. This is the
> idea behind: Ian R. White (2010) "simsum: Analyses of simulation
> studies including Monte Carlo error" The Stata Journal, 10(3):369--385
> and <>. It is not very
> useful as a definition of what amount of bias is "acceptable" as you
> can arbitrarily make the bounds around your estimate of the bias
> smaller by increasing the number of iterations, but at least this type
> of bounds prevents you from over-interpretting the result from your
> simulation, as happend here:
> <>.
> Hope this helps,
> Maarten
> ---------------------------------
> Maarten L. Buis
> Reichpietschufer 50
> 10785 Berlin
> Germany
> ---------------------------------
> *
> *   For searches and help try:
> *
> *
> *
*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index