I will make it four responses. Your trek could easily become a primrose path
with terrible cliffs on each side. I magine a survey in which you have say 100
questions and switch only one, 50/50. It would be great. You have all the other
data to play a game of "Who's got the Association."
But then imagine 20 questions being switched. Think of what you must look at to
assure yourself that the switched questions are not impacting on the survey as a
whole. And think about your reduced sample sizes, with concomitant inflated
errors, and then ask just how sure you are that your imputations are not
capitalizing on the data to erroneously decrease your error. Your inferences are
*not quite* what they would be if all the kids took the entire test, unless your
other data do what YOU want them to.
It's the kind of thing that needs strict limits. Nothing at all wrong with split
surveys, but just how much information do you need? :)
John French
----- Original Message -----
From: <[EMAIL PROTECTED]>
To: "Imputations in Data Analysis" <[EMAIL PROTECTED]>
Sent: Monday, February 05, 2001 4:07 PM
Subject: IMPUTE: Re: planned missingness
| Several people noticed that I thanked three people, but only two responses
| were sent to the impute list, and asked for the third response. Below is
| the response I received from Donald Rubin.
|
| Based on his response, I have tracked down two references so far based on
| the NAEP:
|
| Johnson, E.G. (1992). The design of the National Assessment of Educational
| Progress. Journal of Educational Measurement, 29, 95-110.
|
| Mislevy, R.J., Beaton, A.E., Kaplan, B., & Sheehan, K.M. (1992).
| Estimating population characteristics from sparse matrix samples of item
| responses. Journal of Educational Measurement, 29, 133-161.
|
|
| Mike Frone
|
| --------------------------------
|
| Donald Rubin wrote:
|
| There was(is?) a very major project using this idea years ago done by ETS
| for NAEP (Nat Assessment of Educ Progress). Had many articles in journals
| too using multiple imputation to get valid inferences. I think some of the
| articles were in J of Educ Statistics, in the late 80's or in the 90's,
| with Robt Mislevy as one of the authors. The idea was to limit the size
| of the test that each kid had to take, but getting inferences as if all
| the kids took the entire test. Know there as "matrix sampling" (of kids
| and test items).
|
| Neal Thomas, then of ETS now at B-M-S in Conn, has done much work on it
| too. Also, check out a JASA article in the late 90's by Raghunathan and
| Grizzle on "split questionnaire" designs in health-care research -- exact
| same idea: create systematic missing data (which you know are missing at
| random by design) and multiply-impute to get correct inferences.
|
| Hope this helps.
|
|
|
|
