There are three kinds of lies: lies, damned lies, and statistics.
Mark Twain
Unfortunately, the radiologic literature is not immune to statistical error. Estimates of upto 50% of articles are statistically flawed. Many problems arise because radiologists and statisticians talk a much different language and have difficulty communicating with one another. This should help bridge that gap. Use the following table when analyzing the literature. The rows contain the type of DATA and the columns address the appropriate ANALYSIS for the data. While DATA seems so elementary as to be not worth mentioning, most problems start because the investigator chooses the wrong analysis for the DATA.
Statistics for Radiologists Lecture pointing out some of the flaws seen in the literature. Examples include the choice of improper tests, importance of sample size, and kappa for agreement.
DATA
Nominal or named data is common in radiology. Various categories of a specific disease, for example, idiopathic pulmonary fibrosis, scleroderma, and BOOP- are all nominal data. In nominal data there is no arithmetic relationship between the categories and there is no order (i.e. one is no better than another) among categories.
Ordinal or ordered data is also common and is probably the most common basis for statistical error in radiology. In ordinal data there is no arithmetic relationship among categories but the data are ordered. For example, the severity classification of normal, mild, moderate, and severe is ordinal data. Other examples include the stages of cancer. One of the most common mistakes is to take ordinal data and convert it to interval data.
Interval data is what we are most familiar with and includes the measurements of length, volume, and weight. In radiology, interval data includes Hounsefield numbers, T1 and T2 times, and optical density. In interval data there is a mathematical relationship between categories (i.e. 10 cm is two times that of a measurement of 5 cm).
When reading the literature
1. First decide what type of data the researcher is using.
2. Next, what is the research design. Is this a study of agreement among observers, the association between a radiographic sign and a disease, or the comparison of different radiographic modalities and the detection of disease?
3. Use the table to determine whether the correct statistical method was applied.
*Post-Hoc comparisons: Tukey's HSD, Scheffe's Method, Bonferroni correction, Dunnett's, Newman-Keuls Table adapted from Glanz SA. Primer of Biostatistics 3rd ed. McGraw-Hill, New York 1992
Statistical References: Books
Glantz SA, Slinker BK. Primer of Applied Regression and Analysis of Variance. McGraw Hill. New York, 1990.
Glantz SA. Primer of Biostatistics. 3rd ed McGraw Hill, New York 1992.
Norman GR, Streiner DL. Biostatistics. The Bare Essentials. Mosby, St. Louis 1994.
Swinscow TDV. Statistics at Square One. British Medical Association, London 1983.
Gardner MJ, Altman DG. Statistics with Confidence--Confidence Intervals and Statistical Guidelines. British Medical Journal, London 1989.
Daly LE, Bourke GJ, McGilvray J. Interpretation and uses of Medical Statistics. 4ed. Blackwell Scientific Publ., Boston 1991.
Sox Jr HC, Blatt MA, Higgins MC, Marton KI. Medical Decision Making. Butterworths, Boston 1988.
Hirsch RP, Riegelman RK. Statistical First Aid--Interpretation of Health Research Data. Blackwell Scientific Publ., Boston 1992.
Fleiss JL. Statistical Methods for Rates and Proportions. 2nd ed. Wiley, New York 1981.
Bailar III JC, Mosteller F. Medical uses of Statistics. 2nd ed. NEJM Books, Boston 1992.
Norman GR, Streiner DL. PDQ Statistics. BC Decker, Toronto 1986.
Hirsch RP, Riegelman RK. Statistical Operations--Analysis of Health Research Data. Blackwell, Boston 1996.
Tufle ER. The Visual Display of Quantitative Information. Graphics Press, Cheshire 1993.
Statistics Education Large list of links related to statistics education: online statistical material, courses, handouts, articles etc.
Statistics Every Writer Should Know Plain english concepts mean, median, percentage, rates, standard deviation, normal distribution, margin of error, confidence interval, data analysis, and sample sizes.
Evidence Based Radiology Detailed in depth analysis of applying evidence based principles to radiology
Statistical programs
Download these programs (Excel 5.0 spreadsheets). Use 'save this link as' and save as 'source' file. References and directions are included
ROC Analysis (19K) calculate the nonparametric area under the curve.
Meta-Analysis ROC (14K) method to produce summary curve from multiple studies.
Power Calculations (9K) sample size calculations for sensitivity and specificity.
Volume of SPN (9K) calculate doubling times for SPN.