Page:Sm all cc.pdf/33

 misled by using what are called ‘parametric statistics’, i.e., statistics that assume a Gaussian distribution of errors. This section is organized in the same sequence that most data analyses should follow:
 * 1) test the data for normality;
 * 2) if non-normal, can one transform the data to make them normal?
 * 3) if non-normal, should anomalous points be omitted?
 * 4) if still non-normal, use non-parametric statistics.

Normality Tests
Because our statistical conclusions are often somewhat dependent on the assumption of a normal distribution, we would like to undertake a test that permits us to say “I am 95% confident that this distribution is normal.” But such a statement is no more possible than saying that we are 95% certain that a hypothesis is correct; disproof is more feasible and customary than proof. Thus our normality tests may allow us to say that “there is <5% chance that this distribution is normal” or, in statistical jargon, “We reject the null hypothesis of a normal distribution at the 95% confidence level.”

Experienced scientists usually test data for normality subjectively, simply by looking at a histogram and deciding that the data look approximately normally distributed. Yet I, an experienced scientist, would not have correctly interpreted the center histogram of Figure 2 as from a normal distribution. If in doubt, one can apply statistical tests of normality such as Chi-square (χ2) and examine the type of departure from normality with measures such as skewness. Too often, however, even the initial subjective examination is skipped.

We can use a χ2 test to determine whether or not our data distribution departs substantially from normality. A detailed discussion of the many applications of χ2 tests is beyond the scope of this book, but almost all statistics books explain how a χ2 test can be used to compare any data distribution to any theoretical distribution. A χ2 test is most easily understood as a comparison of a data histogram with the theoretical Gaussian distribution. The theoretical distribution predicts how many of our measurements are expected to fall into each histogram bin. Of course, this expected frequency [Nf(n)] for the nth bin (or interval) will differ somewhat from the actual data frequency [F(n)], or number of values observed in that interval. Indeed, we saw in Figure 2 that two groups of 50 normally distributed measurements exhibited surprisingly large differences both from each other and from the Gaussian distribution curve. The key question then is how much of a difference between observed frequency and predicted frequency is chance likely to produce. The variable χ2, which is a measure of the goodness of fit between data and theory, is the sum of squares of the fractional differences between expected and observed frequencies in all of the histogram bins:

$$\chi^2 = \sum_n([Nf(n)-F(n)]^2/Nf(n)) $$ (3)

Comparison of the value of χ2 to a table of predicted values allows one to determine whether statistically significant non-normality has been detected. The table tells us the range of χ2 values that are typically found for normal distributions. We do not expect values very close to zero,