Chapter 16 Answers to discussion questions

Data analysis in cytopathology

16.1 Reflect on Chapters 4, 5, 8, 9, 10, and 12 with regard to the relative importance of sensitivity and specificity in each branch of screening and diagnostic cytology.

When cytology is used as a screening test (e.g. cervical cytology), sensitivity takes precedence over specificity. High sensitivity is critically important in screening because individuals with negative test results are generally not followed up with further diagnostic tests or clinical evaluation. In this scenario it is important to minimise false negative reporting. However, achieving high sensitivity is not so critical when cytology is used as a diagnostic test (e.g. urine, serous fluid, respiratory tract and FNA cytology) because symptomatic patients, even those with negative cytology, generally receive immediate follow-up tests and further clinical management. In diagnostic cytology it is more important to maintain high specificity, which minimises the unnecessary and possibly harmful further investigations that inevitably follow false positive cytology results.


16.2 Explain why the calculated values of sensitivity, specificity, negative predictive value, and positive predictive value depend critically on the diagnostic threshold adopted by the cytologist.

A low diagnostic threshold can be thought of as an increased tendency to report a positive test result, which generally tends to occur when the consequences of false negative reporting are dire (e.g. in screening tests). A low threshold results in relatively high sensitivity and high negative predictive value but poor specificity and low positive predictive value. On the other hand, a high diagnostic threshold implies a greater tendency to report test results as negative, which is likely to occur when the consequences of a false negative report are relatively unimportant (i.e. when further investigations are planned regardless of the cytology result) or when false positive reporting is likely to result in harmful and/or expensive follow up tests. When the diagnostic thresholds are high, specificity and positive predictive value are also high, but sensitivity and negative predictive value are relatively low.


16.3 Elaborate on the reasons for not including error bars when considering trends in disease mortality rates or incidence rates.

Remember that error bars are only appropriate when we want to make inferences about a population from a representative sample. Since samples are never exact replicas of the population (there will always be an element of chance in the sampling process), there will always be uncertainty (i.e. variability) in the data. Error bars attempt to quantify this uncertainty. Mortality rates and incidence rates are generally reported for whole populations, not samples from the population, so there is no uncertainty in the data and error bars are therefore not appropriate.


16.4 Outline the design of a suitable experiment to test the hypothesis that cell morphology is affected by the type of cytological fixative used.

When designing an experiment it is important to state the hypothesis, to define the independent and dependent variables and to decide in advance the sampling method to be used (two or more groups? related or independent groups?) and the type of analysis to be undertaken. In the given scenario the hypothesis is that cell morphology is affected by the type of fixative used. The independent variable is clearly nominal (i.e. two or more named types of fixative) and the dependent variable should be chosen so that it is an accurate and precise measure of cell morphology. Related samples (i.e. the same samples are subjected to all types of fixative) are ideal for this kind of experiment because this reduces the data variability that would otherwise occur if different samples were used across the different fixative types. Morphometric continuous variables such as nuclear shape, size, density are ideal for this kind of experiment, but alternative ordinal variables (e.g. small, medium sized or large nuclei; low, moderate or high nucleocytoplasmic ratio; pale vs dark nuclei, etc, etc) may have some validity. The type of statistical analysis will depend on the number of groups being compared, whether the groups are related or independent, and on the distribution of data collected for the dependent variable. If the data are normally distributed then choose a suitable parametric test of significance (Paired t-test for two related groups, independent t-test for two unrelated groups, ANOVA for more than two groups). Non-Gaussian data should be analysed by the appropriate nonparametric test (Wicoxon’s signed rank test for two related groups, Mann-Whitney U test for two unrelated groups, and Kruskal-Wallace test for more than two groups).