Estimation with Censored Regressors: 
Basic Issues
 

“Estimation with Censored Regressors: Basic Issues” with Tom Stoker


In this paper, we study the bias that arises from using censored regressors in regression analysis. Firts, we show that, in general, when there are censored regressors in linear models, the estimates are biased. Second, when we concentrate on a prevalent form of censoring - i.e. on censoring to bounds (top-coding or bottom-coding) - it gives rise to expansion bias, or coefficient estimates that are proportionally too large. Third,.we show that there are no informational gains from using the censored data. Therefore, the common practice of introducing a dummy variable to deal with the censoring aspect of the regressors either will not help in the estimation, or continue to bias the coefficients. Finally, we propose a simple semi-parametric procedure to estimate using censored data and applied to the estimation of the marginal propensity of consumption out of wealth.

One focus is on censoring to bounds (top-coding or bottom-coding) which give rise to expansion bias, or coefficient estimates that are proportionally too large. .... We discuss several aspects of consistent estimation with censored regressors, including showing the necessity of certain restrictions for the censored data to offer any information at all on the regression coefficients. We propose a model of mixed censoring, which captures both censoring to bounds and random censoring, The concepts are illustrated by showing how censored regressors arise in the analysis of wealth effects on consumption, including application to household consumption data.


Download recent version of the paper:

ebcr_McFadden_files/ebcr_McFadden.pdf


Download the paper from other sources:

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1071239
http://econpapers.repec.org/article/ieriecrev/v_3a48_3ay_3a2007_3ai_3a4_3ap_3a1441-1467.htm
 
HomeMain.html
BackResearch.html