Explore the vast range of titles available.

Life expectancy at birth, estimated from United States period life tables, has been shown to vary systematically and widely by region and race. We use the same tables to estimate the probability of survival from birth to age 70 (S(70)), a measure of mortality more sensitive to disparities and more reliably calculated for small populations, to describe the variation and identify its sources in greater detail to assess the patterns of this variation. Examination of the unadjusted probability of S(70) for each US county with a sufficient population of whites and blacks reveals large geographic differences for each race-sex group. For example, white males born in the ten percent healthiest counties have a 77 percent probability of survival to age 70, but only a 61 percent chance if born in the ten percent least healthy counties. Similar geographical disparities face white women and blacks of each sex. Moreover, within each county, large differences in S(70) prevail between blacks and whites, on average 17 percentage points for men and 12 percentage points for women. In linear regressions for each race-sex group, nearly all of the geographic variation is accounted for by a common set of 22 socio-economic and environmental variables, selected for previously suspected impact on mortality; R(2) ranges from 0.86 for white males to 0.72 for black females. Analysis of black-white survival chances within each county reveals that the same variables account for most of the race gap in S(70) as well. When actual white male values for each explanatory variable are substituted for black in the black male prediction equation to assess the role explanatory variables play in the black-white survival difference, residual black-white differences at the county level shrink markedly to a mean of -2.4% (+/-2.4); for women the mean difference is -3.7% (+/-2.3).

Life expectancy at birth, estimated from United States period life tables, has been shown to vary systematically and widely by region and race. We use the same tables to estimate the probability of survival from birth to age 70 (S70), a measure of mortality more sensitive to disparities and more reliably calculated for small populations, to describe the variation and identify its sources in greater detail to assess the patterns of this variation. Examination of the unadjusted probability of S70 for each US county with a sufficient population of whites and blacks reveals large geographic differences for each race-sex group. For example, white males born in the ten percent healthiest counties have a 77% probability of survival to age 70, but only a 61% chance if born in the 10% least healthy counties. Similar geographical disparities face white women and blacks of each sex. Moreover, within each county, large differences in S70 prevail between blacks and whites, on average 17 percentage points for men and 12 percentage points for women. In linear regressions for each race-sex group, nearly all of the geographic variation is accounted for by a common set of 22 socio-economic and environmental variables, selected for previously suspected impact on mortality; R2 ranges from 0.86 for white males to 0.72 for black females. Analysis of black-white survival chances within each county reveals that the same variables account for most of the race gap in S70 as well. When actual white male values for each explanatory variable are substituted for black in the black male prediction equation to assess the role explanatory variables play in the black-white survival difference, residual black-white differences at the county level shrink markedly to a mean of -2.4% (+/-2.4); for women the mean difference is -3.7 % (+/-2.3).

Several methods have been proposed for the treatment of missing variables in the context of linear regression estimation. This thesis surveys these methods and finds most of them inadequate due to their complexity and inadequate benefits. The most powerful methods (Minimum Distance and Maximum Likelihood) are potentially useful to the applied researcher.If the true model is known, ML is more efficient than all other methods, but in practice it has a greater computational cost and involves a greater risk of specification error, (new) denotes results which have not previously been published.1. OLSC (regression with the complete data only) is the standard, with low computational cost, no specification of auxilliary models, and it works just as easily for several missing variables.2. OLSI (missing variables replaced by imputed values) can potentially be more efficient than OLSC, but it has several complications which make it not work the effort:a. Difficulty in computing the correct coefficient standard errors.b. Naive standard errors are too large (new) or too small.c. Multiple imputation, while promising simpler computation of standard errors, is inconsistent (new).d. Possible specification error in the auxilliary imputation model.e. Often less efficient, or the efficiency gain is negligible.f. GLS is not worth the additional complications.3. PD (pairwise deletion estimation of moment matrices) is roughly the same as OLSI--it is potentially (but not necessarily) better, and it has impractical complications. The formula for its correct coefficient standard errors, previously derived asymptotically, is shown here to be exact for small samples (new).4. MD (minimum distance) offers efficiency gains with some robustness to heteroscedasticity, and is easier to compute than GLSI and ML (new).5. ML entails simultaneous estimation of the regression and imputation models for full efficiency. A specification test is derived (new). Examples with artifical data show a standard error reduction from OLSC to ML of from 2% to 43% for complete sample percentages of from 90% to 10%. Examples with real data show a larger standard error reduction--from 18% to 33% for complete percentages of from 70% to 50%.

Working Paper No. 908 This paper describes the construction of a large panel data set covering about 2600 firms in the U.S. manufacturing sector for up to twenty years which contains annual data on financial variables, employment, research and development expenditures, and aggregate patent applications. This data set is to be used in a larger study of R&D, inventive output and technological change. In the present paper we present preliminary results on the R&D and patenting behavior of the 1976 cross section of these firms. We find an elasticity of R&D with respect to sales of close to unity, with both very small and very large firms being slightly more R&D intensive than average. Because only 60% of the firms report R&D expenditures, we attempt to correct for selectivity bias and find that though the correction is small, it increases the estimated complementarity between capital intensity and R&D intensity. In exploring the relationship of the patenting activity of these firms to their contemporaneous R&D expenditures, we look with some care at the choice of econometric specifications since the discrete nature of the patents variable for our smaller firms may cause difficulties with the conventional log linear model. The choice of specification does indeed make a difference, and the negative binomial model, which is a Poisson-type model with a disturbance, is preferred. Substantively, we find a much larger output of patents per R&D dollar for the small firms, with a decreasing propensity to patent with size of R&D programs throughout the sample. However, this conclusion is highly tentative both because of its sensitivity to specification and choice of sample and also because we expect that errors in variables bias due to our focus on R&D and patent applications in a single year is far worse for the small firms.