0
$\begingroup$

I have a dataset with multiple variables having different distributions - some are normal and others are skewed. I want to summarize locations for some identified values on both kinds of distributions. Simple percentiles can work for both normal and skewed distributions. (http://www.dummies.com/education/math/statistics/how-to-calculate-percentiles-in-statistics/)

But what's the difference between using simple percentiles vs. z-scores for a normal distribution?

$\endgroup$
  • $\begingroup$ There's a one-to-one mapping between z-score and percentile, so there's no fundamental difference. $\endgroup$ – Moss Murderer Jan 17 '18 at 4:49
  • 1
    $\begingroup$ @Moss That depends on what you mean. The relationship between the percentiles of empirical data and their Z-scores is not given by the Normal probability function: it's specific to each dataset. $\endgroup$ – whuber Jan 17 '18 at 16:54
  • $\begingroup$ @whuber: Yes, the theoretical relation between z-score and percentile given by the normal pdf need not apply to any given dataset, but the empirical relationship between the two measures is guaranteed to be a monotone function, so the two measures are equivalent in the sense that you can freely go back and forth between the two, no? $\endgroup$ – Moss Murderer Jan 17 '18 at 20:28
  • 1
    $\begingroup$ @Moss That's a correct mathematical statement--but it doesn't sound practically useful. Regardless, it's a definite and important difference, which seems to me like it contradicts your original opinion of "no fundamental difference." $\endgroup$ – whuber Jan 17 '18 at 20:44
1
$\begingroup$

The Z-score is a quantile, and takes values from $-\infty$ to $\infty$. The cumulative percentile is bounded from 0 to 1. When the distribution is known, the percentile can map 1-1 to any observation for any distribution, whereas the Z-score only has this property for normally distributed data; hence these summaries are equivalent when normality is met.

Perhaps an advantage of the Z-score is that it can more accurately summarize the distance between two extreme values. An observation 6 standard deviations above the mean is more extreme than one only 4 standard deviations about the mean, but both are beyond 3 decimal places of accuracy for the percentile, so the Z-score can more accurately summarize their relative extremity.

$\endgroup$
  • 1
    $\begingroup$ What you're calling the cumulative percentile is often called a percentile rank (in practice, often given in percent terms, not as a proportion) and could also be called a cumulative probability. It's not the same as percentiles as defined in the OP's link (and more generally). Such percentiles are values on the scale of the data, such as the 1% point, 50% point or the 99% point for heights or weights, which would be height or weight measurements such those fractions of the data have lower values. $\endgroup$ – Nick Cox Jan 17 '18 at 17:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.