Graphing quintiles and medians by state using "twoway connected"

[Karen M Brummond]

Hi Statalist,

I'm using Stata 12 and trying to make a graph that shows the variation in wage gaps within and across states. I want to show how many sub-state geographic regions are in each quintile (Not quartile) within each state as well as the state's median wage gap. I want each state on the y-axis and the wage gap on the x-axis, and the states to be listed from largest median wage gap to smallest median wage gap. For each state, I want a horizontal line the length of the range of wage gaps within the state, and that line should change colors for each quartile. Note: some states have less than 5 substate geographic regions, and it is okay if those states show less than 5 colors.

The following is a sample of my data (in long format; strank shows the labels of the ranking. NM=5, CA=6, HI=7):

strank level femwgcoefst

NM amin -0.247207195
NM bp20 -0.1900343
NM cp40 -0.153927803
NM dmedian -0.145714894
NM ep60 -0.145714894
NM fp80 -0.145714894
NM gmax -0.122379899
CA amin -0.264407486
CA bp20 -0.160001606
CA cp40 -0.149955705
CA dmedian -0.146284893
CA ep60 -0.134702295
CA fp80 -0.134702295
CA gmax -0.035036702
HI amin -0.152224705
HI bp20 -0.147228003
HI cp40 -0.147228003
HI dmedian -0.147228003
HI ep60 -0.147228003
HI fp80 -0.147228003
HI gmax -0.092804603
____________________________


Here is the code I have been using:

sort strank levels

twoway connected strank femwgcoefst if (level=="amin" | level=="bp20"), ///
    msize(0) c(L) lcolor(purple) lwidth(thick) ///
    ylab(1(1)51,valuelabel labsize(tiny) alternate tposition(inside)) ///
    ytitle("State", size(vsmall) margin(right)) ysize(9) ///
    xlab(,labsize(tiny)) xsize(6.5)///
    xtitle("Wage Gap (Negative numbers indicate lower than expected wages)",size(vsmall)) ///
    legend(size(tiny) order(1 "0th to 20th Percentile" ///
    2 "20th to 40th Percentile" 3 "40th to 60th Percentile" ///
    4 "60th to 80th Percentile" 5 "80th to 99th Percentile" 6 "Median" ) cols(3) rows(2)) ///    
    title("Female Wage Gap Variation by State")     ///    
    || connected strank femwgcoefst if (level=="bp20" | level=="cp40"), ///
        msize(0) c(L) lcolor(blue) lwidth(thick)  ///
    || connected strank femwgcoefst if (level=="cp40" | level=="ep60"), ///
        msize(0) c(L) lcolor(green) lwidth(thick)  ///
    || connected strank femwgcoefst if (level=="ep60" | level=="fp80"), ///
        msize(0) c(L) lcolor(orange) lwidth(thick)  ///    
    || connected strank femwgcoefst if (level=="fp80" | level=="gmax"), ///
        msize(0) c(L) lcolor(red) lwidth(thick)  ///
    || scatter strank femwgcoefst if level=="dmedian", msize(small) ///
        mcolor(black)

_______________________

This is very close to working, but lines are sometimes drawn diagonally between adjacent states (see attachment). I have found that this only occurs when the next observation to be plotted has a more positive value for femwgcoefst. For example, looking at the first plot [twoway connected strank femwgcoefst if (level=="amin" | level=="bp20")], the line for NM does not connect to the line for CA because the row for CA-amin has a more negative value for femwgcoefst than NM-bp20 has for femwgcoefst (-0.264407486 < -0.1900343). The diagonal line problem does occur between CA and HI: the row for HI-amin has a more positive value for femwgcoefst than CA-bp20 (-0.152224705 > -0.160001606). I can't think of an if statement to eliminate this problem. Does anyone have a solution or alternatives I could use?

Best,

Karen M. Brummond
Doctoral Student and Research Assistant
University of Massachusetts - Amherst
[email protected]

[Nick Cox]

You're already using the main trick, c(L) rather than c(l). You just need to work harder at the sort order to do this way.

There is another way to do it, using graph bar and some invisible bars, and to work over(state). That will stop all connections between states.

On a different note, recall that many readers have difficulty in distinguishing between red and green.

[Karen M Brummond]

I'll try the graph bar method, and if that doesn't work, I'll get back to you. Thank you!

[Karen M Brummond]

I'm seeing a few issues with the graph bar method.
1) I don't see an option for overlaying another plot so I can mark the state medians with dots. I tried "||" and addplot() and the error message "option not allowed" came back.
2) I have to make the same style of graph for some race data, and there the range of the wage gaps crosses the zero line. I can't think of how a bar graph can represent that.
3) I'm trying to get all of the quintiles on a single line. The only option I see that can make the bars all overlap is stack, but that of course adds up the values, which is not what I want. I might be able to do this with some newly constructed variables and the stack option. I'm yet to try that.

Any thoughts?

[Nick Cox]

1) I don't see an option for overlaying another plot so I can mark the state medians with dots. I tried "||" and addplot() and the error message "option not allowed" came back.

NJC: Correct. Medians as dots are a casualty of this approach. addplot() is only an option for twoway. Conversely, graph dot could be a serious competitor for showing your quintiles directly.

2) I have to make the same style of graph for some race data, and there the range of the wage gaps crosses the zero line. I can't think of how a bar graph can represent that.

NJC: graph bar has no problem with negative values.

Dopey illustration:

sysuse auto
gen foo = runiform() - 0.5
graph bar (asis) foo if foreign, over(make)
graph hbar (asis) foo if foreign, over(make)

3) I'm trying to get all of the quintiles on a single line. The only option I see that can make the bars all overlap is stack, but that of course adds up the values, which is not what I want. I might be able to do this with some newly constructed variables and the stack option. I'm yet to try that.

NJC: Yes, that's right. If you choose bars, what is to be added has to be reverse engineered from what stack will produce. But see my first comment again.