27.10.2014 Views

Russel-Research-Method-in-Anthropology

Russel-Research-Method-in-Anthropology

Russel-Research-Method-in-Anthropology

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

640 Chapter 20<br />

What to do with outliers? Try us<strong>in</strong>g the median to describe the central tendency<br />

rather than elim<strong>in</strong>at<strong>in</strong>g cases. Or report the results of your analysis with<br />

and without outliers, as I just did with figures 20.9a and b.<br />

Tests of Significance<br />

This is a hot topic <strong>in</strong> social science. Some years ago, researchers began<br />

argu<strong>in</strong>g that statistical tests of significance are virtually useless (Rozeboom<br />

1960; Labovitz 1971b; Carver 1978), and the drumbeat has cont<strong>in</strong>ued ever<br />

s<strong>in</strong>ce (Carver 1993; Cohen 1994).<br />

I wouldn’t go that far. Consider the hypothesis that the universe is expand<strong>in</strong>g.<br />

As Wa<strong>in</strong>er (1999) po<strong>in</strong>ts out, be<strong>in</strong>g able to reject the null hypothesis at<br />

p .05 would be quite a contribution. Will Dr. X be denied tenure a year from<br />

now? Lots of people, says Wa<strong>in</strong>er, would be happy to know that they could<br />

reject the null hypothesis at, say, p .001.<br />

It’s true that if you don’t have a representative sample, then a test of statistical<br />

significance doesn’t allow you to generalize beyond your particular sample<br />

of data. On the other hand, if you get significant results on a nonrandom sample,<br />

at least you can rule out the operation of random properties <strong>in</strong> your sample<br />

(Blalock 1979:239ff).<br />

Use tests of significance but remember that they aren’t magical and that the<br />

.01 and .05 levels of significance, while widely accepted tribal customs, are<br />

not sacred. They are simply conventions that have developed for convenience<br />

over the years. Greenwald et al. (1996:181–182) offer some useful advice<br />

about report<strong>in</strong>g p values.<br />

1. In many situations, it’s enough to use simple asterisks <strong>in</strong> your prose to <strong>in</strong>dicate<br />

statistical significance. But if you need to report p values, then do so with an <br />

sign, not with a or sign. If a p value is .042, don’t report it as p .05 (‘‘the<br />

probability is less than .05’’). Just report it as p .042 and be done with it.<br />

Probability, just like the confidence we have <strong>in</strong> probabilistic results, is a cont<strong>in</strong>uous<br />

variable. Why cut out all that <strong>in</strong>formation with arbitrary cutoffs?<br />

2. A s<strong>in</strong>gle p value of .05 is an <strong>in</strong>dicator of a relation, but is not conv<strong>in</strong>c<strong>in</strong>g support<br />

for a hypothesis. By tradition, researchers almost never report probabilities that<br />

are greater than .05. Five repeated results of p .06, or even .10, are more<br />

conv<strong>in</strong>c<strong>in</strong>g than a s<strong>in</strong>gle result of p .05 that someth<strong>in</strong>g’s really go<strong>in</strong>g on.<br />

The Bonferroni Correction<br />

If you want to be especially cautious <strong>in</strong> report<strong>in</strong>g correlations, you can<br />

apply a test known as the Bonferroni correction. There are two k<strong>in</strong>ds of spu-

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!