Suppose we have two real-valued random variables $X,Y$. Let $cdf_X$ and $cdf_Y$ be the corresponding cumulative distribution functions. We are interested in graphically comparing the distributions of $X$ and $Y$.

If we plot the set of points $$(cdf_X^{-1}(z),cdf_Y^{-1}(z))$$ for some $z\in[0,1]$, the resulting graph is called a Q-Q plot. If $cdf_X=cdf_Y$, then the Q-Q plot lies along the $\textbf{x=y line}$ on the graph.

The Q-Q plot is very useful, but if $X$ or $Y$ have a few extremal values that differ, the plot can be somewhat visually misleading. For example, suppose $X$ is a uniform distribution over 1000 samples drawn from a standard normal distribution. $Y$ is generated the same way, with independent samples. Here is a corresponding QQ-plot; note that the points in the upper right and lower left corners wander off the dotted $\textbf{x=y line}$. enter image description here Although the extremal points diverge, there aren't many of them. In order to display the alignment of the majority of the points, we could instead plot $$(z,cdf_Y(cdf_X^{-1}(z)))$$ Here is the corresponding "inverse Q-Q plot"; because the majority of points align well, it is more visually obvious (to me, anyway) that the distributions are similar. enter image description here

I haven't run across the "inverse Q-Q plot" before, but it's sufficiently natural that it's probably a standard tool. Does this plot have a name?

  • 1
    $\begingroup$ Isn't this the P-P plot? en.wikipedia.org/wiki/P%E2%80%93P_plot $\endgroup$ – Nick Cox Jan 6 '16 at 19:49
  • $\begingroup$ Ah! Thank you, that looks right. If you post that as an answer, I'll accept below. $\endgroup$ – Bill Bradley Jan 6 '16 at 20:13
  • $\begingroup$ @NickCox is right. To understand the relationship b/t pp-plots & qq-plots, it may help to read my answer here: PP-plots vs. QQ-plots. Regarding your concern that some points may diverge by chance alone, you just need to incorporate the idea of a sampling distribution & perhaps plot confidence bands. Many implementations of qq-plots will do that for you. $\endgroup$ – gung Jan 6 '16 at 20:17

You've re-discovered the P-P plot. For an introduction, see here.

I'll add a slightly droll comment from one text, to the effect that if you want to be, or to appear, optimistic about fit, you use a P-P plot, whereas if you want to be (appear) pessimistic, you use a Q-Q plot.

Your example is a case in point. The P-P plot is necessarily anchored in principle at [0, 0] and [1, 1], but come even slightly waggly tails, the Q-Q plot shows them quite explicitly. Come a lousy fit, whether through outliers, curvature or grouping, and the Q-Q plot tells the bad news without restraint.

Despite that, the lesser use of P-P plots I guess arises because you have to do more work to relate them to the original data.

EDIT The quotation I had in mind:

Exaggerating a bit, one may say that one should apply the sample df $F_n$ (or, likewise, the survivor function $1 - F_n$) and the P-P plot if one wants to justify a hypothesis visually. The other tools are preferable whenever a critical attitude towards the modeling is adopted.

Reiss, R.-D. and Thomas, M. 2007. Statistical Analysis of Extreme Values: With Applications to Insurance, Finance, Hydrology and Other Fields. Basel: Birkhäuser, p.63. (nearly identical wording in 2nd edition 2001 p.67 and 1st edition 1997 p.57)

  • $\begingroup$ @Gilles Thanks; I'll add the reference next time I am near that text. $\endgroup$ – Nick Cox Jan 6 '16 at 20:22
  • $\begingroup$ That will be greatly appreciated ! $\endgroup$ – Gilles Jan 6 '16 at 20:24
  • 1
    $\begingroup$ @Gilles Now added. $\endgroup$ – Nick Cox Jan 9 '16 at 16:26

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.