# What jitter value should I use for this scatter plot?

I plotted a scatterplot between humidity and temperature (air) in centigrade. I got the following graph;

It is evident that the points fall in discrete columns. This might be because of the rounding of the humidity estimates to one decimal point. To break up these columns, I think some random noise (jitter) could be added to the humidity estimates. I want to know what value of jitter should I use? The new variable (jittered humidity) shall be of the form; humidity + jitter.

• What is the problem that you’re trying to solve with jitter? The data being arranged in those vertical lines is not inherently a problem, because it reflects exactly how the data are recorded.
– Sycorax
Commented May 23, 2022 at 17:03
• @Sycorax ; Though it is not inherently a problem, it is probably not how the original data had been collected. The variable has been rounded to 2 decimal points so perhaps adding some random noise will make it look more empirical. Commented May 23, 2022 at 17:19
• @Sycorax I would maintain it is inherently a problem because the discretization potentially (and almost certainly in this case) creates a great deal of overplotting and thereby can cause the visual impression of regions of greater or less density to be deceptive or even horribly incorrect. See stats.stackexchange.com/a/506611/919 for an illustration.
– whuber
Commented May 23, 2022 at 17:32
• If distance between vertical 'lines' is $w$ then maybe jitter with $\mathsf{Unif}(-w/3,w/3)$ to (mostly) avoid overplotting while still making it clear you're using rounded data. // Similar to suggesting in @whuber's Comment. I have used this and was happy with results. Commented May 23, 2022 at 21:29
• In this case, $0.01.$ So, add iid uniform noise supported on $[-0.005,0.005]$ (to the horizontal coordinate) as one of your initial attempts at visualization and go from there.
– whuber
Commented May 24, 2022 at 20:05

In your case, there appears to be even spacing between columns of points. For a point from a column of points at $$x=a$$ you could uniformly sample a new position of that point from $$[a - \frac{w}{2}, a + \frac{w}{2}]$$ where $$w$$ is the width between columns of points. This should almost-surely protect the order of points along the horizontal axis between columns while spreading out the values.