This research project explores the possibility of using small area statistical techniques to generate information for areas not covered by the American Housing Survey (AHS). This information is only available in the AHS, and could prove useful in preparing for or responding to a disaster.

The study experienced difficulty in identifying variables that would be of obvious help in disaster situations. It looks at nine conditions that might be useful to measure:

• The percent of households with needs—households that (a) are poor, (b) have an elderly householder or spouse, (c) are single-parent households, or (d) have a householder or spouse who is a recent immigrant;
• The proportion of occupied units with severe physical problems;
• The proportion of occupied units with either severe or moderate physical problems;
• The proportion of occupied units that have severe physical problems and are occupied by households with needs;
• The proportion of occupied units that have severe or moderate physical problems and are occupied by households with needs;
• The proportion of occupied units built prior to 1940 and are occupied by households in need;
• The proportion of occupied units that are mobile homes and are occupied by households in need;
• The proportion of owner-occupied units with homeowners' insurance; and
• The proportion of renter-occupied units with renters' insurance.

The research chose to use fractional logit to estimate these conditions, using variables measured for all metropolitan areas by the annual American Community Survey (ACS). Three types of independent variables were used: ACS approximations of the AHS-measured conditions, covariates to distinguish among metropolitan areas on non-housing factors, and covariates related to local housing markets and economic conditions.

The results are only mildly encouraging. Of the nine equations estimated, only two were statistically meaningful using the Chi-squared test. However, the R2s for these two equations indicate that using the predicted values may be a worthwhile improvement over assuming that all metropolitan areas have the same values, that is, using the sample mean. Out-of-sample predictions using these two equations suggest that conditions may vary substantially across metropolitan areas.

Equations with better statistical properties might result from using household-level data rather than metropolitan-level data. This would involve using microdata from the AHS to estimate the equations and microdata from the ACS public use files to apply the equations to predictive purposes. The next major disaster may very well indicate what AHS information is crucial. If such an event were to occur, there may not be time to carry out the desired microdata analysis. If so, the techniques explored in this paper might serve as a valuable second-best approach.