Help in dealing with messy longitudinal line graph

[Zach Goldberg]

Greetings,

I'm running Stata 15.1 on OSX. Using data from the general social survey, I'm attempting to graph longitudinal trends in responses to a binary dependent variable (yes/no) by age group.
I run the following syntax:

regress SPKRAC i.year##i.age_cat4 if white==1 & ideo3==1 [pweight=wtssall]

which produces a graph that looks like this:



As you can see, it's a bit of a mess. I'd like to make it (or, more specifically, the trend lines) more presentable. I was doing a little searching around and read up a bit on 'smoothing'. I have no experience in using this feature, though, and (if this is the appropriate route to take) would like a little assistance in determining which procedure to use.

Here is some sample data:

* Example generated by -dataex-. To install: ssc install dataex
clear
input float(age_cat4 SPKRAC) int year
 1   . 1972
 4   . 1973
 1   . 1974
 2   . 1975
 4 100 1976
 4   0 1977
 4   . 1978
 4 100 1980
 3   0 1982
 1   . 1983
.n 100 1984
 2 100 1985
 1   . 1986
 2 100 1987
 4   . 1988
 1   . 1989
 4   . 1990
 4 100 1991
 3 100 1993
 2   . 1994
 4   0 1996
 4 100 1998
 1 100 2000
 1 100 2002
 4   . 2004
 4   0 2006
 3   . 2008
 2   . 2010
 1   0 2012
 4 100 2014
 3 100 2016
 4   . 1972
 4   . 1973
 3   . 1974
 1   . 1975
 4 100 1976
 2 100 1977
 4   . 1978
 4 100 1980
 3   0 1982
 4   . 1983
 2 100 1984
 3   0 1985
 4   . 1986
 2 100 1987
 3 100 1988
 3   . 1989
 3   . 1990
 2 100 1991
 3 100 1993
 4   . 1994
 2   . 1996
 1   . 1998
 3   0 2000
 3 100 2002
 3   . 2004
 1   . 2006
 3 100 2008
 1   . 2010
 1 100 2012
 1   0 2014
 4   . 2016
 3   . 1972
 2   . 1973
 4   . 1974
 4   . 1975
 3 100 1976
 4 100 1977
 2   . 1978
 4 100 1980
 1   0 1982
 4   . 1983
 3 100 1984
 1   0 1985
 3   . 1986
 4   0 1987
 4   . 1988
 2   0 1989
 1 100 1990
 2   . 1991
 3 100 1993
.n   0 1994
 4 100 1996
 2   . 1998
 4   . 2000
 2   . 2002
 4   . 2004
 4   . 2006
 3   . 2008
 4   0 2010
 3 100 2012
 4 100 2014
 4 100 2016
 1   . 1972
 2   . 1973
 4   . 1974
 1   . 1975
 4 100 1976
 1 100 1977
 4   . 1978
end
format %ty year

Thanks in advance!

[Nick Cox]

The results are what you asked by insisting on so many parameters to be fitted. Smoothing would just make the results into something quite different. It can't help anyone understand what your model is doing.

[Zach Goldberg]

Hey Nick,

I see what you saying. To simplify things, I display only 2 age categories in the graph below


There's a negative trend I'm trying to highlight for AGE_CAT1=1. It's noisy, though. Is smoothing appropriate here? And, if so, how do I determine what bandwidth to use?

Thanks!

[Nick Cox]

As I understand it, you're showing predictions for groups of respondents. As in #2 I have no idea why you think smoothing those predictions will help. What you might do is plot the original data by age categories and ponder smooth summaries of those original data.

The substance is pertinent here too. Is your response likely to fluctuate substantially from year to year or is a smoother long-term pattern expected. That substance surely should guide your exploratory data analysis and modelling.