• Keine Ergebnisse gefunden

Statistical evaluation of quantitative real-time PCR data

2 Material and methods

2.2 Methods

2.2.12 Statistical evaluation of quantitative real-time PCR data

All statistical evaluations of quantitative real-time PCR data were performed us-ing SPSS 16.0 software.

2.2.12.1 Comparison of independent groups

In order to evaluate whether relative levels of mRNA are significantly differen-tially distributed between groups (patients/tumor samples), non-parametric statis-tical tests were performed. Depending on whether two or more groups were com-pared to one another, Mann-Whitney-U-tests or Kruskal-Wallis-H-tests were conducted. These non-parametric tests for comparison of independent data sets make no assumption about the distribution of data (e.g. normality), unlike the pa-rametric t-test. Thus, they represent alternatives to the independent group t-test, when the assumption of normality or equality of variance is not met.

The p-value of asymptotic significance was chosen as significance estimate, with two-tailed significance being estimated in Mann-Whitney-U-tests.

2.2.12.2 Estimation of accuracy and performance of diagnostic tests

The diagnostic performance of a test or the accuracy of a test to discriminate cer-tain cases from other ones can be evaluated using receiver operating characteristic (ROC) curve analysis (Metz 1978, Zweig and Campbell 1993). The result of any particular test system in two populations will rarely separate the two populations perfectly (to 100%). Instead, the distribution of the test results will overlap to a certain extent. Thus, for every possible threshold point (classifier boundary) value that can be selected to discriminate between the two populations, there will be some cases correctly and some cases incorrectly assigned to their respective group. For binary classification systems (two-class prediction problem), in which the test outcomes are labeled either as positive or negative, four types of outcome are possible (Fig. 12): Correctly classified positives (true-positives) or negatives (true-negatives) and incorrectly classified positives (false-positives) or negatives (false-negatives). Parameters giving information about the ratio of true-positives and false-positives at a certain threshold point are the true-positive rate (TPR) and the false-positive rate (FPR). The TPR determines the ratio of positive cases cor-rectly classified among all positive cases available during the test, whereas the FPR defines how many incorrect positive results occur among all negative cases available during the test.

In a ROC curve the TPR (sensitivity) is plotted as a function of the FPR (1-specificity) for different cutoff points (Fig. 13). Each point on the ROC plot repre-sents a sensitivity-specificity pair corresponding to a particular decision threshold.

Hence, the ROC curve is a visual index of the accuracy of a test. A test with per-fect discrimination (no overlap in the two distributions) has a ROC curve passing through the upper left corner with 100% sensitivity (no false-negatives) and 100%

specificity (no false-positives). Thus, the closer the curve passes by this upper left corner, the higher the overall accuracy of the test (Zweig and Campbell 1993).

The calculated area under the ROC curve gives an estimate about how close the

actual curve is to the one of an ideal test, with a value of 1 being equal to the ideal curve.

Abbildung 12:

Figure 13: Schematic overview of possible ROC curves.

ROC curves from tests with different accuracies are depicted.

Figure 12: Schematic distribution of test results of binary classifica-tion systems in two populaclassifica-tions.

True-positives (TP), false-positives (FP), true-negatives (TN) and false-negatives (FN) are depicted according to selected threshold value.

w/o ‒ without; e.g. ‒ for example

Abbildung 13:

In this work, receiver operating characteristic was used to evaluate the perform-ance of a gene’s relative mRNA expression level to potentially serve as diagnostic tests distinguishing patients with recurrence or formation of postoperative metas-tases from relapse-free patients. According to ROC curve the most “appropriate”

cutoff value of expressions was selected with aiming for high sensitivity and low specificity of the test, respectively.

2.2.12.3 Survival analysis

Survival analyses were performed using the Kaplan-Meier method (Kaplan and Meier 1958). This procedure is a common way to summarize survival data via the estimation of survival probabilities at certain time points. The Kaplan-Meier sur-vival curve plots the proportion of sursur-vival/probability of sursur-vival as a function of

time, with each death being represented by a downward step in the curve. Thus, the curve shows, for each time point, the proportion of subjects that survive at least this length of time.

An important advantage of the Kaplan-Meier method is that it can take "censored"

data (losses of samples before the final outcome is observed) into account. In clinical applications such censored data can occur when a patient withdraws from a study, for instance. In the curve, small vertical ticks indicate such “losses”.

The time until postoperative relapse/progression, the so-called disease-free sur-vival (DFS) or the time until formation of postoperative metastases, the so-called metastasis-free survival (MFS) were chosen for survival analyses.

Differences in survival times between groups were compared using the non-parametric logrank-test. Significance in survival difference was accepted at p<0.05.

2.2.12.4 Multivariate analysis

Multivariate analyses are techniques to analyze more than one variable at a time.

Binary logistic regression is the method of choice if your dependent variable (so-called criterion variable) is binary (dichotomous) and you wish to explore the rela-tive influence of continuous and/or categorical independent variables (so-called covariates) on your dependent variable, and to assess interaction effects between the independent variables (Spicer 2004).

In this thesis, binary logistic regression was used to estimate the independence of the information offered by genes of interest (on transcriptional level) from the information provided by other clinical parameters.