Rapid Responses to:

PRACTICE: J Martin Bland and Douglas G Altman Analysis of continuous data from small samples BMJ 2009; 338: a3166 [Full text]

Rapid Responses: Submit a response to this article

Rapid Responses published:

Importance of "eye test" and null hypothesis

Arun S. Nanivadekar   (7 April 2009)

Parametric versus non-parametric: the parametric horse may have pulled ahead

Tony H. Reinhardt-Rutland   (9 April 2009)

Importance of "eye test" and null hypothesis 7 April 2009
Arun S. Nanivadekar, Medical Research Consultant C-2. Flushel Apts, 21 Road, Bandra (W), Mumbai 400050, India Send response to journal: Re: Importance of "eye test" and null hypothesis	I liked to read this article because it, especially the example in Table 2, illustrates the importance of "eye test" or looking at the data meaningfully, and choosing a sensible null hypothesis to test. The researcher must never lose sight of the biological meaning he or she is trying to extract from the data, and of how it can be revealed in a manner that the reader will understand and appreciate. This is especially important when the computing power available to researchers is likely to divert their attention from what they are computing, how, and for what purpose. Competing interests: None declared
Parametric versus non-parametric: the parametric horse may have pulled ahead 9 April 2009
Tony H. Reinhardt-Rutland, Reader in Psychology University of Ulster, BT52 1SA Send response to journal: Re: Parametric versus non-parametric: the parametric horse may have pulled ahead	Decisions on the choice of parametric or non-parametric inferential statistics for continuous data are by no means hard and fast. This is reflected in the believed robustness of parametric techniques, which is now recognised generally as much less of an issue than once believed. W. L. Hays (1) who wrote a sequence of editions of his standard statistical text-book over many years became notably iconoclastic in this regard. While earlier editions included non-parametric techniques, the final edition eschewed them entirely. Such a decision was partly pragmatic: the range of non-parametric techniques for dealing with multivariate continuous data is meagre. However, Hay's take on robustness implied that the issues were largely misconstrued. For example, transformations can indeed deal with skewed data. However, the effect is to reduce the influence of extreme data points on the error variance as a proportion of the total variance. The outcome of this is often an increase in the number of significant F ratios, rather than - as might be supposed - to make the criteria for significance more stringent. A major problem continues regarding floor and ceiling effects: no transformation may be feasible or desirable in such circumstances. Arguably, such cases represent methodological failure, which may be retrieved by re-design of the experiment. Alternatively, a non-parametric approach might provide a solution. References (1) Hays, W. L. (1994). Statistics (Fifth ed.). New York: Harcourt-Brace. Competing interests: None declared