We examined the extent of nonresponse and the need for imputations in the AHEAD. We find that the bracketed information elicited by the unfolding questions can have a great benefit in increasing the information known for wealth. In addition, it appears that the bracketed responses appear to represent real uncertainty concerning the value of the assets rather than providing a convenient substitute. Despite these benefits, a very large portion of wealth is subject to some type of imputation. We examine several alternative imputation models and procedures. These methods differ along the dimentions of the amount and the interactions of covariates, whether the error terms in the model of response and wealth are thought to be dependent, and the method by which random errors, if any, are incorporated into imputed values. We found that different imputation methods may potentially imply different economic behavior. ∗We thank Mike Hurd, Dan McFadden and Hilary Hoynes for their comments. All remaining errors are ours. An early version of this paper was presented in AHEAD Early Results Conference, September, 1994, Survey Research Center, University of Michigan. †The corresponding author. Email: gan@eco.utexas.edu, phone: (512)475-8540.
[1]
R. Little,et al.
A note about models for selectivity bias.
,
1985
.
[2]
D. Rubin,et al.
Statistical Analysis with Missing Data
,
1988
.
[3]
Roderick J. A. Little.
Missing-Data Adjustments in Large Surveys: Reply
,
1988
.
[4]
Finis Welch,et al.
What Do We Really Know about Wages? The Importance of Nonreporting and Census Imputation
,
1986,
Journal of Political Economy.
[5]
Daniel F. Heitjan,et al.
Inference from Grouped Continuous Data: A Review
,
1989
.
[6]
Donald B. Rubin,et al.
Inference from Coarse Data via Multiple Imputation with Application to Age Heaping
,
1990
.
[7]
J. Morgan,et al.
Survey estimates of wealth: An assessment of quality
,
1989
.