1Biomathematics and Statistics Scotland, The King's Buildings, Edinburgh EH9 3JZ, UK; 2Department of Psychology, University of Edinburgh, Edinburgh EH8 9JZ, UK; 3Department of Mathematics and Statistics, University of Edinburgh, Edinburgh EH9 3JZ, UK

Background and objectives
Methods of visual assessment of plant disease are liable to a variety of recording errors which can bias results and reduce their precision. A better understanding of the sources of these errors and their mode of action can allow remedial action through improved training of recorders, changes in sampling methods, or statistical analysis. One potentially important source of error, which has received little attention to date, is carryover. Carryover arises in sequential sampling when the score given to the current sample is influenced by previous samples. Two mechanisms for carryover in sensory assessment are assimilation [1], where the scorer supplements the current stimulus with the memory of previous stimuli, and contrast [2], where the memory of previous stimuli leads to a shift in the scorer's measurement baseline towards these values. With assimilation, a sample score will tend to have a positive bias when preceded by samples with high scores, and vice versa, whereas when carryover is of the contrast type, the biases are reversed in sign. This study was designed to investigate the bias from carryover which might arise in scoring for a plant disease where severity is assessed by the percentage of leaf infection.

Materials and methods
Circular lesions with random centres and radii were generated by computer to cover a given percentage of a square image. Images were produced, with seven levels of lesion cover, c=2, 5, 10, 17, 26, 37 and 50%, i.e. at unit intervals of sqrt(c-1). A sequence of 100-200 images was projected, each for 6 s, onto a large screen and the percentage cover assessed in three separate experiments by 20, 100 and 31 students, respectively. A training set of six images was shown before each experiment. The first two experiments were designed to investigate carryover from the immediately preceding image. A balanced sequential design was used in which each image type was followed by all seven types an equal number of times.

The third experiment was designed to study whether the carryover effect was increased by repetitive reinforcement of an image type. A balanced design was used which allowed the scores for an image type to be compared when it was preceded by different sequences of high (H=37%) and low (L=2%) cover images, viz HHH, HLH, LHL, LLL. For each experiment, scoring bias was analysed using the model
y=observed score-true level=scorer bias+image bias+scorer x image interaction+carryover bias+image x carryover interaction+residual error.
Scores were transformed to square roots before analysis to stabilize their variances.

Results and conclusions
Although scorers showed considerable variation in their overall bias, there was a general tendency to overestimate image cover and the bias was larger at lower levels of cover. The first two experiments also showed positive carryover bias so that, when an image followed an image with a higher level of cover, a higher score was recorded than when it followed an image with a lower level of cover. An image's score was affected only by the sign of the difference in cover from the preceding image, not by its magnitude. There was a similar pattern of carryover in experiment 3 in relation to the immediately preceding image, but there was no evidence of added carryover effect from repeating high or low cover images. Analysis of residual errors after fitting the model showed no trends with time.

These experimental results suggest that quantitative estimates of disease obtained by scoring plants visually in rapid succession may be biased because of an assimilation effect. This will tend to increase scores for plants with low infection and reduce scores for plants with high infection. A similar form of bias occurs in disease-screening trials due to the spread of pathogens between small plots [3], and together they can lead to a significant underestimate of plant resistance or treatment control.

1. Baird JC, Berglund B, Berglund, U, Lindberg S, 1991. Perceptual and Motor Skills 73, 3-17.
2. Schifferstein HNJ, Oudejans IM, 1996. Perception and Psychophysics 58, 713-724.
3. Kempton RA, Fox PN, 1997. Statistical Methods for Plant Variety Evaluation. Chapman & Hall, London.