A couple of observations which may help:
1. The AMstart option I was suggesting is 2 which assumes--as you point
out--zero covariances for starting values. This is necessary for some
applications because the starting values created by the listwise
deleted dataset (which is how the default starting values are
calculated) have problems. I would try this.
The third option of defining your own matrix for the starting values is
almost never necessary.
2. If you have observation identifiers or any text variables in your
dataset, you have to define them with IDvar global. They can't be
included in the imputation model. Of course, if you don't have such
variables in your dataset because you already eliminated them, you are
fine.
Ken
Quoting Jose Antonio Pena Merino <jpm283(a)nyu.edu>du>:
Dear Prof. Scheve,
Thanks a lot for replying to my message.
I got rid of those variables for which there was perfect linear
dependence (there were 2 of them), and run Amelia but it did not
work, the same message kept coming up:
Iteration Number:
sweep: elements of m cannot be zero
Currently active call: SWEEP [79]
When I tried to change the default start values it gives me either
the option of a default with covariances zero, or file with starting
values (I wanted to write 2 as you suggested me), however it asks me
to enter the name of a file which contains the starting values and
hit return, am I supposed to create such a file? is it a matrix?
WHat I did was:
I set AMpri equal to 2 as it is advised for high missingness data
I described my fully variables
I left AMmtho under the default (should I change it to 2?)
I did not define any idvar (I have tried with and without and it does
not make any difference)
I did not define any time of sectional variable in the AMts and AMcs
(or the AMusets, AMusecs and AMtstep)
I did not ask for any nominal variable to be returned as such
I asked for 1 ordinal variable to be returned as such
I got the same message again...
Sorry for taking so much time out of you and again, thank you for
helping me put here,
Yours
Jose Merino
----- Original Message -----
From: Kenneth Scheve <kenneth.scheve(a)yale.edu>
Date: Saturday, February 28, 2004 4:03 pm
Subject: Re: [amelia] run but
Actually this message is not about which variables are fully
observed. It
usually indicates there is something wrong with the data or at
least that
portion of the dataset being used in the initial calculations.
Did you try setting the AMstart global to "2" rather than the
default?
Also check and make sure that you have not included data for which
there is
perfect linear dependence between some group of variables.
Ken
At 03:53 PM 2/28/2004, Jose Antonio Pena Merino wrote:
>OK, the program is running now, I think I was asking way too much
by
including too many variables and asking it to keep
all the
ordinal and
nominal as such.
However, even if it is running it keeps returning the following
message:>
>Iteration Number:
>sweep: elements of m cannot be zero
>Currently active call: SWEEP [79]
>
>I define those variables which contain full information already,
but it
looks like Amelia keeps detecting one variable
with full
information. In
order to detect which one was it I kept adding
each variable to
the list
of fullys, but all of them were detected as
non-fully. So I
really do not
know why I keep getting the same message. In sum,
there is no
other
variable besides those already defined as fully
for which m=0
(that is,
with no missing values)
Thanks for clarifying this for me
Yours
Jose Merino