# Quantile regression revealing different relationships at different quantiles: how?

Quantile regression (QR) is sometimes said to reveal different relationships between variables at different quantiles of the distribution. E.g. Le Cook et al. "Thinking beyond the mean: a practical guide for using quantile regression methods for health services research" imply that QR allows for relationships between the outcomes of interest and the explanatory variables to be nonconstant across different values of the variables.

However, as far as I know, in a standard linear regression model $$y = \beta_0 + \beta X + \varepsilon$$ with $\varepsilon$ being i.i.d. and independent of $X$, the QR estimator for the slope $\beta$ is consistent for the population slope (which is unique and does not anyhow vary across quantiles). That is, the object being estimated is always the same, regardless of the quantile. Admittedly, this is not the case for the intercept, since the QR intercept estimator aims at estimating a particular quantile of the error distribution. Taken together, I do not see how the different relationships between variables are supposed to be revealed at different quantiles via the QR. I guess this is a property of the standard linear regression model rather than a mistake in my understanding, but I am not sure.

I suppose the situation is different when some of the assumptions of the standard linear model are violated, e.g. under certain forms of conditional heteroskedasticity. Then perhaps the QR slope estimators converge to something else than the true slope of the linear model and somehow reveal different relationships at different quantiles.

What am I getting wrong? How should I properly understand/interpret the claim that quantile regression reveals different relationships between variables at different quantiles?

• There are lots of ways to think about QR. One is that it is a type of kernel regression where the kernels are the quantiles. In this way it is nonparametric and robust approach in which linear solutions cannot be assumed. Hyndman, et al, has proposed boosted adaptive quantile regression as a global framework for QR based modeling. Ungated copy here ... robjhyndman.com/papers/sig-alternate.pdf – Mike Hunter Sep 21 '17 at 14:12
• @DJohnson, thanks. I guess I am too influenced by the original paper Koenker and Bassett (1978) where the motivation is solely to find a robust slope estimator in the standard linear model rather than elicit different relationships at different quantiles. – Richard Hardy Sep 21 '17 at 14:14
• There's no question but that papers like Koenker and Bassett's impacts how future analysts frame a question. Another good paper on QR is Le Cook and Manning's 2013, *Thinking Beyond the Mean: a practical guide for using quantile regression methods"... ungated copy here ... dash.harvard.edu/bitstream/handle/1/12406692/… FWIW...but their focus is health care... – Mike Hunter Sep 22 '17 at 11:23

The "true slope" in a normal linear model tells you how much the mean response changes thanks to a one point increase in $x$. By assuming normality and equal variance, all quantiles of the conditional distribution of the response move in line with that. Sometimes, these assumptions are very unrealistic: variance or skewness of the conditional distribution depend on $x$ and thus, its quantiles move at their own speed when increasing $x$. In QR you will immediately see this from very different slope estimates. Since OLS only cares about the mean (i.e. the average quantile), you can't model each quantile separately. There, you are fully relying on the assumption of fixed shape of the conditional distribution when making statements on its quantiles.

EDIT: Embed comment and illustrate

If you are willing to make that strong assumtions, there is not much point in running QR as you can always calculate conditional quantiles via conditional mean and fixed variance. The "true" slopes of all quantiles will be equal to the true slope of the mean. In a specific sample, of course there will be some random variation. Or you might even detect that your strict assumptions were wrong...

Let me illustrate by an example in R. It shows the least squares line (black) and then in red the modelled 20%, 50%, and 80% quantiles of data simulated according to the following linear relationship $$y = x + x \varepsilon, \quad \varepsilon \sim N(0, 1) \ \text{iid},$$ so that not only the conditional mean of $y$ depends on $x$ but also the variance.

• The regression lines of the mean and the median are essentially identical because of the symmetrical conditional distribution. Their slope is 1.
• The regression line of the 80% quantile is much steeper (slope 1.9), while the regression line of the 20% quantile is almost constant (slope 0.3). This suits well to the extremely unequal variance.
• Approximately 60% of all values are within the outer red lines. They form a simple, pointwise 60% forecast interval at each value of $x$.

The code to generate the picture:

library(quantreg)

set.seed(3249)
n <- 1000
x <- seq(0, 1, length.out = n)
y <- rnorm(n, mean = x, sd = x)

plot(y~x)

(fit_lm <- lm(y~x)) # intercept: 0.02445, slope: 1.04858
abline(fit_lm, lwd = 3)

# quantile cuts
taus <- c(0.2, 0.5, 0.8)

(fit_rq <- rq(y~x, tau = taus))
#               tau= 0.2      tau= 0.5    tau= 0.8
# (Intercept) 0.00108228 -0.0005110046 0.001089583
# x           0.29960652  1.0954521888 1.918622442

lapply(seq_along(taus), function(i) abline(coef(fit_rq)[, i], lwd = 2, lty = 2, col = "red"))

• +1. I think the crucial part is in variance or skewness of the error depending on $x$, which is what I tried ruling out by saying "standard linear regression model". I have edited my post accordingly to make it clearer. Regarding By assuming normality and equal variance, all quantiles of the conditional distribution of the response move in line with that, I guess the normality assumption is redundant. – Richard Hardy Sep 20 '17 at 16:34
• Exactly. If you are willing to make that strong assumtions, there is not much point in running QR as you can always calculate conditional quantiles via conditional mean and fixed variance. The "true" slopes of all quantiles will be equal to the true slope of the mean. In the sample, there will be some random variation. Or you might even detect that your strict assumptions were wrong... ;-) – Michael M Sep 20 '17 at 17:05
• That makes sense. In sample, I think the QR slope estimates for different quantiles will likely be somewhat spread out in line with the quantiles. This is because the loss function being minimized will drag the estimator asymmetrically to one side (the direction and the magnitude of the drag depending on the quantile), although asymptotically this effect will get ever smaller. – Richard Hardy Sep 20 '17 at 18:40
• It is a good answer, and thank you for it, but I wonder if you could illustrate with a simple example how the QR reveals different relationships at different quantiles when some of the standard assumptions (e.g. homoskedasticity) do not hold. – Richard Hardy Sep 21 '17 at 12:07
• So the data generating process is a linear model but with the standard deviation (variance?) of $x$ directly proportional to the mean of $x$, right? I.e. $y=x+x\varepsilon$ where $\varepsilon \sim i.i.N(0,1)$? Spelling the model out explicitly in the answer would be very helpful, IMHO. – Richard Hardy Sep 21 '17 at 13:48