Testing for statistical significance of a trend line or pattern of means in stata

[Tinofa]

I have panel data by year covering all years between 2003-2013, and in each year, I calculate mean age at first sex. I then graph these means which reveal a pattern of the age at first sex over time, separately for females and males. I now want to assess if the trend pattern I am observing is statistically different. For variables that are proportions like proportion using condoms, I use ptrend command in stata. But I don't know what to do if its a mean?

Regards,

Tinofa

[Joseph Coveney]

The conventional test of differences between means is ANOVA. So, you could try something like

anova age year##sex

If the residuals are right skew, then you might wish to consider various generalized linear models, such as

glm age i.year##i.sex, family(gaussian) link(log) nolog

or

glm age i.year##i.sex, family(poisson) link(log) robust nolog

or

glm age i.year##i.sex, family(gamma) link(log) nolog

If you wish to perform a nonparametric test for trend (separately for each sex), then there is the official Stata

nptrend age if sex == "M", by(year)
nptrend age if sex == "F", by(year)

and the user-written

jonter age if sex == "M", by(year)
jonter age if sex == "F", by(year)

[Tinofa]

Dear Joseph,
after running the command jonter AgeFirstSex if Sex == "MAL", by(SurveillanceYear) I got:
Jonckheere-Terpstra Test for Ordered Alternatives

J = 21564583
J* = 15.766 (corrected for ties)


Pr(|Z| > |J*|) = 0.0000 (ordered alternative in either direction)
Pr(Z > J*) = 1.0000 (descending ordered alternative)
Pr(Z < J*) = 0.0000 (ascending ordered alternative)

But I neither sure if the command is correct nor how to interpret the results; could you help please?

[Federico Tedeschi]

Question asked a long time ago, but it seems to me the command is still an option for investigating non-parametric time trends. In this case, I'd say that, for males, there's overwhelming significance of an increasing time trend of age at first sex.