1. Comprehensive introductions to the theory and practice of linear regression modelling are given by Draper and Smith (1981), and Montgomery and Peck (1982). Methods of estimation are described in many text books on statistics, such as Cox and Hinkley (1974), and the different approaches are reviewed and contrasted in Barnett (1982).
2. The logistic regression model was first suggested by Berkson (1944), who showed how the model could be fitted using iteratively weighted least squares. Comparisons between the logistic and probit transformations are made by Chambers and Cox (1967), and Berkson (1951) explains why he prefers logits to probits. The equivalence of iterative weighted least squares and the method of maximum likelihood is discussed by Thisted (1988; Section 4.5.6). The computation of maximum likelihood estimates using the method of scoring is due to Fisher (1925), but a more readable summary of the method is included in Everitt (1987). The algorithm used to obtain the maximum likelihood estimates of the parameters of a linear logistic model was given in a general form by Neider and Wedderburn (1972) and is reproduced in McCullagh and Neider (1989) and Aitken et al. (1989).
3. The use of the deviance as a measure of goodness of fit was first proposed by Neider and Wedderburn (1972) and fuller details on the adequacy of the chisquared approximation to its null distribution are contained in McCullagh and Neider (1989). The argument against using the deviance as a measure of goodness of fit for binary data was given by Williams (1983). The empirical logistic transform and the relationship between logistic regression and discriminant analysis are described in greater detail by Cox and Snell (1989). Krzanowski (1988) includes a clear account of methods of classification in his general text on multivariate methods.
4. Issues to be faced before an attempt is made to model an observed response variable are discussed by Chatfield (1988). Cox and Snell (1989; Appendix 2) provide an excellent summary of how to choose a subset of explanatory variables for use in logistic regression modelling.