MCMC Estimation
MCMC = Markov chain Monte Carlo
an alternative approach to estimating models
1
What is the big deal about Markov chain Monte Carlo methods?
While MCMC methods are not new, recent advances in algorithms using these
methods have led to a bit of a revolution in statistics. This revolution is typically
seen as a "Bayesian" revolution because of the fact that the MCMC methods
have been put to work by relying on Bayes theorem. It turns out that
combining the use of Bayes theorem with MCMC sampling permits an
extremely flexible framework for data analysis. However, it is important for us
to keep in mind that a Bayesian approach to statistics and MCMC are separate
things, not one and the same.
For the past several years, MCMC estimation has given those adopting the
Bayesian philosophical perspective a large advantage in modeling flexibility
over those using other statistical approaches (frequentists and likelihoodists).
Very recently, it has become clear that the MCMC-Bayesian machinery can be
used to obtain likelihood estimates, essentially meaning that one doesn’t have
to adopt a Bayesian philosophical perspective to use MCMC methods. The good
news is that there are a lot of new capabilities for analyzing data. The bad
news is that there are a lot of disparate perspectives in the literature (e.g.,
people attributing the merits of MCMC as being inherent advantages of the
2
Bayesian perspective, plus different schools of Bayesian analysis).
Bayesian Fundamentals
1. Bayes Theorem
P(D|M) P(M)
P(M|D) =
P(D)
where:
P(M|D) = the probability of a model/parameter value given the data
P(D|M) = the probability (likelihood) of the data given the model
P(M)
= the prior probability of the model/parameter value given
previous information
P(D)
= the probability of observing these data given the
data-generating mechanism
3
Bayesian Fundamentals (cont.)
2. The context of the Bayesian approach is to reduce
uncertainty through the acquisition of new data.
3. What about that prior?
a. When our interest is predicting the next event,
the prior information may be very helpful.
b. When our interest is in analyzing data, we usually
try to use uninformative priors.
c. When we know something, like percentage data
don't go beyond values of 0 and 100, that can be
useful prior information to include.
d. The biggest worry about priors is that they may
have unknown influences in some cases.
4
Bayesian Fundamentals (cont.)
4. Bayesian estimation is now frequently conducted using
Markov Chain Monte Carlo (MCMC) methods. Such
methods are like a kind of bootstrapping that estimates
the shape of the posterior distribution.
5. MCMC methods can also be used to obtain likelihoods;
remember,
posterior = likelihood * prior.
By data cloning as described in Lele* et al. (2007), it is
possible to obtain pure likelihood estimates using MCMC.
*Lele, Dennis, and Lutscher (2007) Data cloning: easy maximum likelihood
estimation for complex ecological models using Bayesian Markov chain
Monte Carlo methods. Ecology Letters 10:551-563.
5
MCMC Estimation in Amos
The next few slides give a few screen shots of
Bayesian/MCMC estimation in Amos. I highly recommend
the brief video developed by Jim Arbuckle that can be
found at www.amosdevelopment.com/site_map.htm. Just
go to this site and look under “videos” for Bayesian
Estimation: Intro.
6
Illustration of Bayesian Estimation in Amos
icon to initiate MCMC
7
Illustration (cont.)
frown means not yet
converged
8
Illustration (cont.)
point estimates
smile means program
converged; once you
have converted, you
can pause simulation
none of the 95% credible
intervals include the value
of 0. This indicates that we
are 95% sure that the true
values of the parameters
fall within the CIs and are
nonzero.
9
Illustration (cont.)
Some measures of model fit.
Posterior predictive p values
provide some information on overall
model fit to data, with values closer
to 0.50 being better than ones
larger or smaller.
DIC values for different models can
be compared in a fashion similar to
the use of AIC or BIC.
Discussions of model comparison
for models using MCMC will be
discussed in a separate module.
10
Illustration (cont.)
shape of the prior for the parameter for the
path from cover to richness.
right-click on
parameter row
to select either
prior or posterior
for viewing
11
Illustration (cont.)
shape of the posterior for the parameter for the
path from cover to richness.
S.D. is the standard deviation
of the parameter.
S.E. is the precision of the
MCMC estimate determined
by how long you let the
process run, not the std. error!
there are important options
you can select down here,
like viewing the trace,
autocorrelation, or the 12
first
and last half estimates.
Illustration (cont.)
shape of the trace for the parameter for the
path from cover to richness.
The trace is our evaluation of
how stable our estimate was
during the analysis. Believe it
or not, this is how you want
the trace to look! It should
not be making long-term
directional changes.
13
Illustration (cont.)
shape of the autocorrelation curve for the parameter for the
path from cover to richness.
The autocorrelation curve
measures the asymptotic
decline to independence for
the solution values. You want
it to level off, as it has done
here.
14
Standardized Coefficients
To get standardized coefficients, including a full set of moments plus their
posteriors, you need to select "Analysis Properties", the "Output" tab, and then
place a check mark in front of both "Standardized Estimates" and "Indirect,
direct & total effects". If you don't ask for "Indirect, direct, and totol effects",
you will not actually get the standardized estimates.
Then, when you have convergence from your MCMC run, go the the "View"
dropdown and select, "Additional Estimands". You will probably have to grab
and drag the upper boundary of the subwindows on the left to get to see
everything produced, but there should be a column of choices for you to view
(shown on next slide).
For more information about standardized coefficients in SEM, see, for example,
Grace, J.B. and K.A. Bollen. 2005. Interpreting the results from multiple
regression and structural equation models. Bulletin of the Ecological Society of
America. 86:283-295.
15
Standardized Coefficients (cont.)
here you can see the results
here you can choose various options
16
Calculating R2 Values
Amos does not give the R2 values for response variables when the MCMC
method is used for estimation. Some statisticians tend to shy away from
making a big deal about R2 values because they are properties of the sample
rather than of the population. However, other statisticians and most subjectarea scientists are usually quite interested in standardized parameters such as
standardized coefficients and R2 values, which measure the "strength of
relationships". On the next slide I show one way to calculate R2 from the MCMC
output. The reader should note that R2 values from MCMC analyses are (in my
personal view) sometimes problematic in that they are noticably lower than a
likelihood estimation process would produce. I intend a module on this
advanced topic at some point.
17
Calculating R2 Values
error variance for response variable
R2 = 1 – (e1/variance of salt_log)
We need the implied variance of response variables to calculate R2. To get implied
variances in Amos, you can select that choice in the Output tab of the Analysis
Properties window. With the MCMC procedure, you have to request Additional
Estimands from the View dropdown after the solution has converged. For this example,
we get an estimate of the implied covariance of salt_log of 0.119. So, R2 = 118
(0.034/0.119) = 0.714. This compares to the ML estimated R2 of 0.728. Again, I will
have more to say in a later module about variances and errors estimated using MCMC.
Final Bit
Amos makes Bayesian estimation (very!) easy. Amos can do a
great deal more than what I have illustrated, like estimate
custom parameters (like the differences between values).
Unfortunately, Amos cannot do all the kinds of things that can
be done in lower level languages like winBUGS or R. This may
change before too long (James Arbuckle, developer of Amos,
is not saying at the moment). For now, tapping the full
potential of MCMC methods requires the use of another
software package, winBUGS (or some other package like R). I
will be developing separate modules on SEM using winBUGS in
the near future for those who want to use more complex
models (and are willing to invest considerably more time).
19
Descargar

Intro to Amos Bayesian SEM and MCMC