Hi Scott, If you are speaking about modern imputation techniques, such as Bayesian multiple imputation, then I am afraid your initial assumption is wrong. The particular value imputed has no predictive value for how the person would have responded, and indeed, should never be interpreted as such (that's Nostradamian imputation, if I am right...). The reason for imputing is to allow for all of the observed data to be analyzed. Any particular value that is imputed is only meant to preserve the characteristics of the variance-covariance matrix and mean vector, not to be interpreted. That's why a critical aspect of Bayesian multiple imputation is the third phase wherein we estimate the within and between database error provided by the imputed values, then decrease the significance based on treating our imputed values as though they were real in the analysis phase.
Jason ____________________________________ Jason C. Cole, PhD Senior Research Scientist & President Consulting Measurement Group, Inc. Tel:?? 866 STATS 99 (ex. 5) Fax: 310 539 1983 2390 Crenshaw Blvd., #110 Torrance, CA 90501 E-mail: [email protected] web: http://www.webcmg.com ____________________________________ -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Scott Ferson Sent: Thursday, February 15, 2007 3:34 PM To: [email protected] Subject: [Impute] interval statistics (the un-imputation) If an imputation is an intelligent guess about the value of a missing piece of information, this list might be interested in related methods that refrain from making any guesses at all about the missing information. A draft report on "interval statistics" is available at http://www.ramas.com/intstats.pdf that reviews basic descriptive statistics for data sets that contain intervals (rather than exclusively point values). It reviews methods to compute basic univariate descriptive statistics, including various means, the median and percentiles, variance, interquartile range, moments, confidence limits, and introduces the prospects for analyzing such data sets with the methods of inferential statistics such as outlier detection and regressions. The report also explores the trade-off between measurement precision and sampling effort in statistical results that are sensitive to both, and considers the use of interval statistics as an alternative approach for the field of metrology. I'd be very interested to hear your thoughts about this topic, including arguments that imputation procedures that generate specific values are better than interval statistics methods that don't. Best regards, Scott Scott Ferson [email protected] Applied Biomathematics 1-631-751-4350 _______________________________________________ Impute mailing list [email protected] http://lists.utsouthwestern.edu/mailman/listinfo/impute
