Hi
I am a PhD student and a statistics novice. I would very much
appreciate your advice regarding my approach to using Amelia. I have
administered 4 research instruments to 218 study participants. The
instruments and number of items in each is as follows:
-
CDI – 36 items
-
SDQII - 66 items
-
Parental Bonding Instrument (PBI) – 25 items for Female
Parent form and 25 items for Male Parent Form
-
“I am … Test” – yields a single score for each subject.
Out of the total sample of 218, only 140
(64.22%) of the total sample provided complete information across all
research questionnaires. The proportion of cases with missing data is consistent with King, Honaker, Joseph, &
Scheve’s (2001) estimate that on average one third of cases in a data
set will have missing values. To preserve as many cases as possible I
want to use Amelia II to multiply impute missing values.
Initially this is what I did:
1.
Data were not MCAR, but MAR assumption appeared
to hold, therefore Amelia II seemed promising.
2.
When all items of the CDI, PBI, SDQ, and “I am
…” Test were combined into one large data set, the number of cases or
observations (n = 218) was only slightly greater than the
number of variables in the combined data set. Even
a ridge prior of 11 (5.05% of 218) was found to be insufficient to
estimate the imputations. Following the failure Amelia II to estimate
the imputations with the implementation of a moderate ridge prior, the
CDI data set, the PBI data set, and the SDQII data set were submitted
to Amelia separately. This ensured that the number of observations was
adequate for the number of imputations to be estimated by Amelia II.
For each of the CDI, SDQ, PBI Mother Form and PBI Father Form, this
preserved all cases except where the participant did not provide any
data whatsoever for the questionnaire concerned. The resulting 5 data
sets for each instrument were produced. The “I
am Test” could not be submitted to AmeIia II for multiple imputation of
missing values because the test yields only one score (the %S score
which is the % of Collectivist responses based on content analysis of
answers provided by the participant).
3.
I calculated scale scores for each of
the instruments, involving each of the 5 datasets produced by Amelia II
for each instrument. The scale scores were to be used for
multiple regression analysis.
My new problem is that where study
participants did not complete any of the items in an instrument, Amelia
II inserted NA in the entire row. This means that there are still
missing values in the data set comprised of scale scores that I wish to
analyse using multiple regression.
if the observations are independent and you know nothing about some
people, then only magic would enable you to impute anything other than
NAs. so if you know something about these folks, code them up as
variables.
Another approach I had in mind is to:
1. Before submitting the instruments to Amelia
II, calculate scale scores for all instruments, except for cases that
have skipped items and therefore have missing values on one or more
items in a scale.
2. Submit the scale scores for all instruments
to Amelia II to produce 5 data sets with imputed missing values.
3. Run multiple regression analysis with each of
the 5 datasets comprised of the same obseved and different estimates of
missing values for scale scores.
This approach would mean that Amelia II would be
imputing values for missing scale scores even when a study participant
did not complete any items whatsoever in the scale concerned, or
indeed, the entire instrument. This is like "getting something out of
nothing". Would this be an appropriate or inappropriate use of Amelia
II?