It is current practice to record the performance of immunoassays by means of precision profiles (according to Ekins), in which the variation coefficient (relative standard deviation) is plotted against the analyte concentration. On the other hand, precision profiles are only occasionally used for evaluating the performance of conventional clinical-chemical methods. It is relatively uncommon to find bias plotted against analyte concentration, presumably because this type of analysis requires reference specimens, whose true analyte concentrations are known or guaranteed by reference methods. If the relative systematic variations are plotted against the true analyte concentrations, and a confidence interval is added to the resulting regression curve, the result is a "bias profile"; if tolerance limits are added, the result is a "deviation profile". The present work describes the preparation of specimens, which can be used to provide experimental data for the three performance profiles. A computer program is used to construct the precision profile, bias profile and/or deviation profile. The mathematical-statistical basis of the program is described in detail. For evaluation of the statistical procedure, three analytes and six methods were used: determination of sodium activity/concentration with an ion sensitive electrode and by flame photometry; determination of creatinine by a manual enzymic and a mechanized Jaffé method; determination of thyrotropin by radioimmunoassay and by luminescence immunoassay. Different purposes are served by bias and deviation profiles. Thus, bias profiles can be used to compare the bias of two methods, whereas a deviation profile can be used to define the analytical range of a method. If the acceptable limits of deviation are added to the deviation profile, then the useful analytical range of the method is immediately apparent.

The biometrical treatment of laboratory data may require the estimation of a regression line for the transformation of one set of measurements to another. The regression procedure introduced in part I (1) of our work does not always yield unbiased results in such situations, since its estimators are not scale invariant. In part III we present the parameter estimation of a general regression equation which is scale invariant and retains all properties of the method comparison procedure, in particular its robustness. Its application is demonstrated by several examples, and the results are compared with other robust biometrical procedures. The mathematical aspects are explained in the appendix.

In continuation of the three previous workshops on this theme, the Passing-Bablok method was recapitulated and compared with standardized main component analysis. Participants in the workshop recommended both methods in place of the classical regression analysis. An essential advantage of the standardized main component is that it is more easily understood by the clinical chemist and his coworkers, and it involves fewer calculation steps. On the other hand, outliers must be recognized and eliminated, and certain other assumptions must be fulfilled. In comparison, the Passing-Bablok method is much more robust, but requires more calculation. The workshop also discussed unresolved questions concerning bases for the comparison of methods: the necessary number of samples, measures of scatter for description of precision, comparison of discrete results and several independent variances. These will be treated in greater depth in subsequent workshops.

Evaluation of analytical instruments in clinical labora- tories should be carried out in a systematic and uniform manner in order to provide data which can be compared and reproduced by different users, manufacturers and officials.Concepts for such an evaluation of instruments with continuous measurement values are presented in the ECCLS guidelines [1].To quantify the results of an

Zusammenfassung: Es werden die Faktoren, die einen Einfluß

Stimmungen durchgeführt.In derselben Serie werden von einem Aliquot des Serumpools ohne Wirkstoff ebenfalls nochmals n>10 Bestimmungen mitgezogen.Die Aus

Let categorical data coming from a control group and ( r - 1) treated groups be given in an r × c contingency table. A simultaneous test procedure of the ( r - 1) hypotheses that the probabilities of all c categories do not differ between the i -th treated group and the control is derived. For small tables and small cell frequencies it is exactly performed by generation of all tables having the given marginal sums. If 2 categories or 2 groups only are given the asymptotic distribution of the test statistic is known; otherwise its distribution may be simulated if the computational expenditure of performing an exact test is too large. By means of a Monte Carlo study it is shown that this method meets its level more reliably and that it has a better power than others.

In part I of this series (H. Passing & W. Bablok (1983), J. Clin. Chem. Clin. Biochem. 21, 709-720) we described a new biometrical procedure for the evaluation of method comparison studies. In part II we now discuss its properties and compare them with those of other established procedures by means of a simulation study. We demonstrate that the reliability of the results not only depends on the sample size but also on the sampling distribution, the precision of the methods, and the concentration range covered by the samples. Linear regression and principal component procedures are either inadequate or not as reliable as our new procedure. The appropriate sample size is discussed and recommendations are given.

s werden die Faktoren, die einen Einfluß auf die zu ermittelnden Präzisionsdaten haben können, in proben-, reagenzien-, gerate-, personen-und auswertebezogene Faktoren eingeteilt.Es wird unterschieden zwischen Präzisionsangaben bei Einsatz von gleichen und verschiedenen Probenmaterialien.Die verschiedenen Arten der Präzisionen werden je nach den gewählten Randbedingungen kurz beschrieben und die jeweils konstant gehaltenen Faktoren festgelegt.Auf die Planung,

Procedures for the statistical evaluation of method comparisons and instrument tests often have a requirement for distributional properties of the experimental data, but this requirement is frequently not met. In our paper we propose a new linear regression procedure with no special assumptions regarding the distribution of the samples and the measurement errors. The result does not depend on the assignment of the methods (instruments) to X and Y. After testing a linear relationship between X and Y confidence limits are given for the slope beta and the intercept alpha; they are used to determine whether there is only a chance difference between beta and 1 and between alpha and 0. The mathematical background is amplified separately in an appendix.

Distribution-free statistical procedures should be applied to the establishment of assigned values and uncertainty intervals in a control serum. The two problems, how to find an appropriate statistical evaluation procedure and how to find an optimized experimental design, are simultaneously dealt with here: Three distribution-free procedures are presented, each based on elimination of extreme values, and 60 designs are considered differing with respect to the number of reference laboratories, of independent series, and single or double determinations. Using the data of the study described in part 1 of this series (Passing, H. et al. (1981) this j. 19, 1137-1144) we simulated these designs and the pertaining assigned values and uncertainty intervals given by each evaluation procedure. From this study one evaluation procedure is shown to be superior to others. This optimized procedure had the following characteristics: Extreme values are eliminated so that the width of the uncertainty interval is as small as possible. The median of the remaining values is the assigned value. Moreover, 6 reference laboratories are shown to be appropriate.

We compare 4 statistical models for the establishment of assigned values in a control serum which are based on the assumption of a normal distribution. The first model results in mean +/- 2s, whereas each of the following 3 models are based on a special analysis of variance. We studied by means of appropriate statistical tests the distributional properties of the data of the study described in part 1 of this series (Passing, H. et al. (1981) this j. 19, 1137-1144). Many model assumptions are violated: The totality of data of each method was never normally distributed, normal distribution within laboratories was not given in 27 out of 67 cases, and precision and accuracy varied from reference laboratory to reference laboratory. Moreover, assigned values and uncertainty intervals calculated by means of these methods can be misleading to the customer. Therefore, these models cannot be applied, and a distribution-free procedure has to be used instead.

We describe a study by which the establishment of assigned values of a control serum was simulated. The study covered two controls: The internal known control and a blind control. Seven constituents were analyzed in 10 or 11 laboratories, respectively, yielding a total of 72 sets of analytical values. Each set covered double determinations within approximately 18 series for each sample. The course of a blind control correlates better with the unknown sample for which assigned values are to be determined than does a known control. Out of 72 sets 5 sets were found incorrect. Out of these, 2 sets could be recognized exclusively by means of the blind control, and 2 others primarily by means of the known control; one of these could be detected by means of the double determinations. Consequently, a blind control has a greater control efficiency than double determinations.

The dry weight of the ratio dry/moist weight of BCG harvested from surface cultures is essentially affected by the following variables (observed variations of the resulting dry weight in parentheses): the temperature of the bacterial mass when harvested (1:1·36), the pressure exerted on the harvest (1:1·47), the technique of dry weight determination (1:1·20), and the grinding of the bacterial mass (1:1·34). Minor effects are contributed by the amount of BCG harvested per litre of medium and the type of harvest apparatus used. A multiplication of the effects of the different variables must be assumed. Sixty-four independent batches of BCG vaccine produced under standardized conditions of harvesting and processing and adjusted to a constant bacterial concentration by moist weight exhibited the following coefficients of variation for several dependent variables: dry weight 4·6%, extinction value 16%, number of culturable particles (based on a log normal distribution) −27%/+39%.

From the study described in part 1 of this series (Passing, H. et al. (1981), this j. 19, 1137-1144) it has been derived in part 3 (Passing, H. (1981), this j. 19, 1153--1166) that 6 reference laboratories are appropriate for the establishment of assigned values and their uncertainty intervals in a control serum. From the same study the two outstanding characteristics of an optimized design are found: It is sufficient that each reference laboratory performs single determinations within independent series. The number of series however has no relevant influence on accuracy and precision of the assigned values. The following design is obtained: Six reference laboratories perform single determinations in five independent series each. Additionally, the complete concept of establishment of assigned values is summarized. Up to now, the model has met the expectations in practice.