N. Scott Cardell

Estimating logistic regression models when the dependent variable has no variance

We show that the binary logistic regression model can often be estimated even when the study sample is confined to observations on only one of the possible outcomes of the dependent variable. Provided that an appropriate supplementary sample can be found, the two samples may be pooled, and a simple method employed to estimate the model with a conventional statistics package. The supplementary sample must contain information on the regressors of the model, but need not contain any information on the dependent variable. Hence supplementary samples can often be found in general purpose public use surveys such as the U.S. Census.

Why some feedforward networks cannot learn some polynomials

It seems natural to test feedforward networks on deterministic functions. Yet, some simple functions, notably polynomials, present some difficult problems for approximation by feedforward networks. The estimated parameters become unbounded and fail to follow any unique pattern. Furthermore, as the fit to the specified functions becomes closer, numerical problems may develop in the algorithm. This paper explains why these problems occur for polynomials of order less than or equal to the number of hidden units of a feedforward network. We show that other examples occur for functions mathematically related to the networks squashing function. These difficulties do not indicate problems with the training algorithm, but occur as an inherent consequence of the role of the connection weights in feedforward networks.

Decisionmaking in the motor carrier industry

Distinct discrete decisions made by the same economic actor are likely to be correlated, particularly those decisions involved in transportation demand. These correlations may be due to economic interactions among the decisions, or to heterogeneity, or both. This paper develops an econometric method that can distinguish among these possibilities and applies this method to analyzing motor carrier deregulation. One overall goal of the Motor Carrier Act of 1980, which substantially reduced federal regulation, was to promote greater competition in the industry. Other things equal, the more competition in an industry, the less concentration. However, concentration has been increasing among general freight carriers since deregulation. The positive correlations found between the decisions to acquire broker authority and to be a general freight carrier could be due to joint economies between broker authority and general freight carriage, in which case the two interact as economic complements, or to heterogeneity among motor carrier licensees. In the first case, but not the second, deregulation of broker authority could lead to increased concentration in the general freight sector. Empirically, we find that the decisions are correlated due to heterogeneity, but do not interact. Thus, increased concentration in the general freight sector must be attributed to other causes.

Coordinate Based Empirical Likelihood-Like Estimation in Ill-Conditioned Inverse Problems

In the context of a semiparametric regression model with underlying probability distribution unspecified, an extremum estimator formulation is proposed that makes use of empirical likelihood and information theoretic estimation and inference concepts to mitigate the problem of an ill-conditioned design matrix. A squared error loss measure is used to assess estimator performance in finite samples. In large samples, the estimator can be designed to be consistent and asymptotically normal, so that limiting chi-squared distributions provide a basis for hypothesis tests and confidence intervals. Empirical risk results based on a large-scale Monte Carlo sampling experiment suggest that the estimator has, relative to traditional competitors, superior finite-sample properties under a squared error loss measure when the design matrix is ill-conditioned.

Modeling heterogeneity in susceptibility and infectivity for HIV infection

Models of the spread of human immunodeficiency virus (HIV) infection must deal with substantial heterogeneity in the populations at risk. The virus is spread by behaviors that are far from uniformly distributed in the population, and substantial variations in biological aspects of susceptibility and infectivity are also likely. How adequately a model represents this heterogeneity will substantially determine its accuracy and usefulness for capturing the dynamics of the epidemic, for making forecasts of future spread, and for answering questions of policy interest.

KDD-Cup 2000: question 2 winner's report Salfor sytems

Responding to the challenge required an extended process of data preparation and exploratory data analysis, combining our data with other information from the US Census, deciding on the specific contrasts to study (high spender vs. low spender, high spender vs. non-spender), and running a series of CART models to separate the high spenders from other groups. This section briefly describes the steps we took to arrive at our final analysis.