Examples of response surfaces for pairs of nutrients and results of 41 multi-level experiments with N only were used to compare the goodness-of-fit of polynomial, inverse polynomial, exponential and intersecting-straight-lines models. Whereas no one model fitted best at every site, many results were well represented by two intersecting straight lines and on average, this model had the least residual mean square. Of 17 experiments with spring barley in south western England the few results best represented by smooth curves were from crops much affected by leaf diseases.Fertilizer response was poorly represented by models without a falling asymptote, like the simple exponential and inverse linear. Study of residuals after fitting the quadratic showed that this widely used model consistently over-estimated both the amount of fertilizer needed for maximum yield and the yield loss when too much fertilizer was given.When fitted to the mean yields of each nitrogen treatment, most models had residual mean squares equal to or less than the error mean square, repeating a result obtained at Rothamsted as early as 1927. We question the validity of some well-known evidence for block and treatment additivity.For 12 experiments in 1970, between-site differences in the parameter values of the two straight lines representing grain yield were related to leaf area at ear emergence.TNTTT? onnrTTO'Nr series of experiments, each experiment testing many levels of one or more nutrients. Instead of estimating Most of the experiments discussed in this and the the effects of individual treatments and their following paper result from important develop-significance, the main object is to determine the ments in experimental methods in the past decade, optimal dressing (or, with several nutrients, their In crop nutrition these changes have resulted optimal combination) and to discover how and why mainly from soil scientists' need to make their fertilizer requirements differ from place to place advice to farmers increasingly precise and at the and from year to year. This paper discusses the form same time less empirical as a result of improved of response curve relating crop yield and fertilizer understanding of plant:soil relationships. Greater input, based mainly on results of multi-level nitroattention has been paid to planning, whereby site gen experiments with cereals, selection is related to soil and husbandry factors likely to influence fertilizer requirements, and there The need f or f resh information on fertilizer response has been more thorough and more uniform recording curves of these and other external factors. The isolatedThe average results of experiments with sugar experiment with few treatments and treatment beet, testing different amounts of N fertilizer, were levels but much replication has been replaced by well fitted by a quadratic response curve, but when