Computing confidence intervals.

[Jean Jacques]

Hi Statausers,

I'm struggling with computing confidence intervals. I have a dataset where each observation is a talk where it is recorded 1) the topic of the talk (dominant_topic), 2) if the speaker is male or female (gender_main_speaker), 3) and the number of times that speaker was interrupted (inter). I want to produce a chart where in the x-axis i have the topics of the talk and in the y-axis the average difference in interruptions for each topic registered for males and females.

This is the code that I'm running to compute the intervals

levelsof dominant_topic, local(hh7s)
gen lowerbound = .
gen upperbound = .
foreach h of local hh7s {
    ci means inter if dominant_topic == `h' & gender_main_speaker == 1
    replace lowerbound = r(lb) if dominant_topic == `h' & gender_main_speaker == 1
    replace upperbound = r(ub) if dominant_topic == `h' & gender_main_speaker == 1
    ci means inter if dominant_topic == `h' & gender_main_speaker == 0
    replace lowerbound = r(lb) if dominant_topic == `h' & gender_main_speaker == 0
    replace upperbound = r(ub) if dominant_topic == `h' & gender_main_speaker == 0
    }

collapse inter lowerbound upperbound, by(dominant_topic gender_main_speaker jel_combined)
bysort dominant_topic(gender_main_speaker): gen diff = inter[_n] - inter[_n-1]
bysort dominant_topic(gender_main_speaker): gen lb = lowerbound[_n] - lowerbound[_n-1]
bysort dominant_topic(gender_main_speaker): gen ub = upperbound[_n] - upperbound[_n-1]

And here the code in which i create the graph

gen newvar = _n
labmask newvar, values(jel_combined)

twoway rcap ub lb newvar, lstyle(ci) ||   ///
       scatter diff newvar, scale(0.5)  xlabel( ,valuelabel angle(45)) xlabel(#20) yline(0,lstyle(foreground) lcolor(red)) legend(off) xtitle("Identified Seminar Topics", size(medium))  ytitle("Difference in Average Interruptions (Females - Males)", size(medium)) msymbol(S) mcolor(dknavy) graphregion(color(white)) bgcolor(white)

I'm unsure about the way in which the confidence intervals were constructed. Note that jel_combined are just the string labels to use instead of the numeric dominant_topic.

Here you have data as example.

* Example generated by -dataex-. For more info, type help dataex
clear
input byte(dominant_topic inter gender_main_speaker)
13 17 1
10  6 0
 6 17 1
13 30 1
13 13 1
13 20 0
 6  8 1
13  5 0
13 18 0
10  6 1
 6  7 0
13 11 0
 6 18 0
10  6 0
13  6 0
13 27 0
 1  7 0
 9 10 1
13 27 1
 0  4 1
13 29 1
 1  7 1
13 19 0
13  6 0
 2  7 1
13 21 1
12  8 0
 3 12 0
14 13 0
 3 19 0
 6 23 1
 3 12 1
 3 13 1
 3  8 1
 7 13 0
14  8 0
11 15 0
 3 11 0
13  2 1
13 32 0
13  7 0
13 23 1
14 21 1
 8 14 1
 0  6 1
 8 12 0
 5 13 0
 8  6 1
 8  8 0
 6  0 0
 4  0 0
13  0 0
11  0 0
 4  0 0
12  0 1
13 16 1
 0 10 0
10  7 1
13 10 0
10 12 0
 2 13 1
13  9 0
13 17 0
13 11 0
13  6 0
13  8 0
13 23 1
13 15 0
10  4 0
13 19 1
13 14 1
13 30 0
14 21 0
10 15 0
13 17 0
 1  6 0
13 30 0
13 11 0
13 29 1
 6  0 1
13 18 0
13 14 0
13 22 1
12  1 1
 3  1 0
 4  0 1
12  0 0
12  1 0
13 13 0
13  6 0
 5 30 0
13 12 0
13 11 1
 0 38 0
13 13 1
13 16 0
 9  9 1
12  9 1
 2  1 0
11  3 0
end

If anybody can help me by reassuring me that the way that I followed is correct I would appreciate it a lot. Thanks!

[Maarten Buis]

I'm unsure about the way in which the confidence intervals were constructed.

To find out you read the pdf manual, this is much more detailed than the help-file. So type help ci and at the top of the help file you see a link to the pdf manual. Click there, and look for the link to "Methods and Formulas", click there and see that it is just a regular t-test.