Twoway spline with multiple variables

[Chinmay Sharma]

Hi All,

I am currently using a twoway connected graph, with multiple variables as follows:

twoway connected x1 x2 x3 year

However, this graph is all over the place, as the values of x1, x2 and x3 change a lot. As such, I wish to smooth it. The twoway mspline command is what I am looking for, but it does not allow me to graph multiple variables. Is there any option that will allow me to do this?

Thanks,
CS

[FernandoRios]

Perhaps you are looking for something like this:

sysuse auto
twoway scatter mpg weight, msize(*.5) || mspline mpg weight || scatter price weight, yaxis(2) || mspline price weight, yaxis(2)

HTH
Fernando

[Bruce Weaver]

Fernando has expanded on one of the examples shown under -help twoway spline-. Just below that example, there is a cautionary note regarding the setting for bands().

Cautions

    The graph shown above illustrates a common problem with this technique:  it tracks
    wiggles that may not be real and can introduce wiggles if too many bands are chosen.  An
    improved version of the graph above would be

        . scatter mpg weight, msize(*.5) || mspline mpg weight, bands(8)

I was curious about the default setting for bands(), and found this:

    bands(#) specifies the number of bands for which cross medians should be calculated.
        The default is max{min(b1,b2),b3}, where b1 is round{10*log10(N)}, b2 is
        round{sqrt(N)}, b3 is min(2,N), and N is the number of observations.

I worked out that bands(9) was the default setting for this example. Then I tried again with bands(8).

clear *
sysuse auto
summarize mpg weight price

twoway scatter mpg weight, msize(*.5) ///
 || mspline mpg weight ///
 || scatter price weight, yaxis(2) ///
 || mspline price weight, yaxis(2)
graph rename g1

display "Bands = " max(min(round(10*log10(_N)),round(sqrt(_N))),min(2,_N))

// Change default for bands() from 9 to 8
twoway scatter mpg weight, msize(*.5) ///
 || mspline mpg weight, bands(8) ///
 || scatter price weight, yaxis(2) ///
 || mspline price weight, yaxis(2) bands(8)
graph rename g2

HTH.

[Nick Cox]

mspline is great for showing as a smooth curve which you know already to be a smooth curve.

As a smoothing method it is idiosyncratic and thus needs to be explained unless reviewers or audience are happy with "I don't know" or "I did it in Stata" as a story. The big picture is slicing into bands, finding (median of y, median of x) in each band and joining those summary points. The result can be quirky and sensitive to band number and boundaries, as I think Bruce Weaver is hinting.

For regularly spaced time series, I would usually start with tssmooth nl -- which allows multiple repetitions of Hanning and thus any binomial filter (which are eminently linear).

For scatter plot smoothing, I strongly recommend lpoly as highly flexible (pun intended) and as easy to explain (in the sense that you can throw literature references back at any questioners).

But if a series really is "all over the place", no smoother can tame it!