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.

632 Chapter 20<br />

.26. This means that <strong>in</strong> fewer than five tests <strong>in</strong> a hundred would we expect to<br />

f<strong>in</strong>d the correlation smaller than .14 or larger than .26. In other words, we are<br />

95% confident that the true r for the population (written , which is the Greek<br />

letter rho) is somewhere between .14 and .26.<br />

By contrast, the 95% confidence limits for an r of .30 <strong>in</strong> a random sample<br />

of 30 is not significant at all; the true correlation could be 0, and our sample<br />

statistic of .30 could be the result of sampl<strong>in</strong>g error.<br />

The 95% confidence limits for an r of .40 <strong>in</strong> a random sample of 30 is statistically<br />

significant. We can be 95% certa<strong>in</strong> that the true correlation <strong>in</strong> the population<br />

() is no less than .05 and no larger than .67. This is a statistically significant<br />

f<strong>in</strong>d<strong>in</strong>g, but not much to go on <strong>in</strong>sofar as external validity is<br />

concerned. You’ll notice that with large samples (like 1,000), even very small<br />

correlations are significant at the .01 level.<br />

On the other hand, just because a statistical value is significant doesn’t<br />

mean that it’s important or useful <strong>in</strong> understand<strong>in</strong>g how the world works.<br />

Look<strong>in</strong>g at the lower half of table 20.18, we see that even an r value of .40 is<br />

statistically <strong>in</strong>significant when the sample is as small as 30. If you look at the<br />

spread <strong>in</strong> the confidence limits for both halves of table 20.18, you will notice<br />

someth<strong>in</strong>g very <strong>in</strong>terest<strong>in</strong>g: A sample of 1,000 offers some advantage over a<br />

sample of 400 for bivariate tests, but the difference is small and the costs of<br />

the larger sample could be very high, especially if you’re collect<strong>in</strong>g all your<br />

own data.<br />

Recall from chapter 7, on sampl<strong>in</strong>g, that <strong>in</strong> order to halve the confidence<br />

<strong>in</strong>terval you have to quadruple the sample size. Where the unit cost of data is<br />

high—as <strong>in</strong> research based on direct observation of behavior or on face-toface<br />

<strong>in</strong>terviews—the po<strong>in</strong>t of dim<strong>in</strong>ish<strong>in</strong>g returns on sample size is reached<br />

quickly. Where the unit cost of data is low—as it is with mailed questionnaires<br />

or with telephone surveys—a larger sample is worth try<strong>in</strong>g for.<br />

Nonl<strong>in</strong>ear Relations<br />

And now for someth<strong>in</strong>g different. All the examples I’ve used so far have<br />

been for l<strong>in</strong>ear relations where the best-fitt<strong>in</strong>g ‘‘curve’’ on a bivariate scatterplot<br />

is a straight l<strong>in</strong>e. A lot of really <strong>in</strong>terest<strong>in</strong>g relations, however, are nonl<strong>in</strong>ear.<br />

Smits et al. (1998) measured the strength of association between the educational<br />

level of spouses <strong>in</strong> 65 countries and how that relates to <strong>in</strong>dustrialization.<br />

Figure 20.5 shows what they found. It’s the relation between per capita energy<br />

consumption (a measure of <strong>in</strong>dustrialization and hence of economic development)<br />

and the amount of educational homogamy <strong>in</strong> those 65 countries.

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

Saved successfully!

Ooh no, something went wrong!