Inaccurate regression coefficients in Microsoft Excel 2003: an investigation of Volpi’s 'zero bug'
Autor: | Bruce D. McCullough, H.-J. Sun, Kaoru Fukuda |
---|---|
Rok vydání: | 2017 |
Předmět: |
Statistics and Probability
Numeric precision in Microsoft Excel Computer science Numerical analysis 05 social sciences Zero (complex analysis) Microsoft excel Linear regression function 01 natural sciences QR decomposition 010104 statistics & probability Computational Mathematics 0502 economics and business Linear regression 0101 mathematics Statistics Probability and Uncertainty Algorithm 050205 econometrics |
Zdroj: | Computational Statistics. 32:1411-1421 |
ISSN: | 1613-9658 0943-4062 |
Popis: | Leonardo Volpi found that Excel 2003, rather than report correct coefficients, would sometimes change them to zero. We have investigated this so-called “zero bug” of the linear regression function LINEST(), and have found that the inaccuracy is caused by a non-standard modified back-substitution procedure. The modification, for which we can find no justification in the numerical analysis or statistical literature, uses a logic to control the bug: when certain conditions are met, accurate coefficients are replaced with inaccurate coefficients that may be zeros or nonzeros. Although Excel 2003 is now out of support, it is still in use. We do not know whether the modification is limited to Excel 2003, or whether Microsoft has programmed similar inaccuracies into other functions or other versions of Excel. |
Databáze: | OpenAIRE |
Externí odkaz: |