[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: event history analysis with years clustered in individuals

From   Steven Samuels <>
Subject   Re: st: event history analysis with years clustered in individuals
Date   Sun, 15 Feb 2009 16:07:09 -0500



You might explain to the professor that, with survival data, the number of years of observation is itself the (posssibly censored) outcome. Therefore "year" cannot be a level 1 effect in a multilevel model;


On Feb 15, 2009, at 3:43 PM, Hilde Karlsen wrote:

Ah. Ok, I see I have to do some serious rethinking when it comes to this essay, then. I guess this to a certain degree explains why I have trouble understanding what sigma_u refers to in this specific analysis. I am wondering if I should forward this e-mail correspondance to the professor who held the course in multilevel techniques, because what I've learned from you today are not in line with what we were told at the course when it comes to this matter. Anyway. Thank you so much for the advice and for answering me.


Quoting Steven Samuels <>:

I agree with Austin. Just to be clear: sigma_u is a parameter that is meaningless for this problem, No interpretation is possible.

On Feb 15, 2009, at 9:22 AM, Austin Nichols wrote:

Hilde Karlsen <>:
If you have to use a mixed model as an exercise, and you have no
compelling reason to choose a particular research question, you should
ask a different research question where a mixed model is a more
appropriate model, not apply it blindly to data you know is better
suited to a survival model. Why not use the attrition dummy you have made as the explanatory variable in a mixed model instead--what other
variables do you have on the data?

On Sun, Feb 15, 2009 at 8:26 AM, Hilde Karlsen <> wrote:
Thank you both for the advice. However, I don't think I can do as you suggest because I have to use a multilevel approach for this essay (it is an essay for a multilevel course I followed a while ago). I should probably have been more clear on this issue, and on what my problem really is. What I am wondering is not which method/command I should use, but how I am going to interprete the sigma_u estimate when my level 1 variable is years and my
level 2 variable is individuals.

As mentioned, I find it more intuitive to grasp the point of separate variance estimates when the levels are schools, classes etc, but for some reason I have a hard time understanding how I should interpreate the variance estimate sigma_u when the years are clustered in individuals. How should I interpreate sigma_u when years are clustered in individuals.

I asked the professor who was leading the course which command I should use, and he told me I should use xtmelogit (my advicor told me the same thing). As he is the one who is going to judge wheter I pass or not on this essay,
it is probably best to follow his advice.

I agree that it is a survival model, and I have designed my data for this type of analysis (i.e. all individuals in the file start out with 0 on the dependent variable, and when/if they drop out of the nursing occupation, they receive 1 on the dependent variable. I have no info on which date/month people drop out; I only have information on which year they drop out).


Quoting Steven Samuels <>:

Hilde, I agree with Austin's approach. Even if you have only months, not days, of starting and quitting, use that time unit in a survival or discrete survival model. I recommend Stephen Jenkins's -hshaz- (get it from SSC); his "model 1" (the "Prentice-Gloeckler model" is the same as that fit by -cloglog-. His model 2 adds unobserved heterogeneity and so may be more
realistic (Heckman and Singer, 1984).

I would not be surprised if prediction equations for of early and later quitting differed. If so, time-dependent covariates or separate models for
early and later quitting, would be informative.


Prentice, R. and Gloeckler L. (1978). Regression analysis of grouped survival data with application to breast cancer data. Biometrics 34 (1):

Heckman, J.J. and Singer, B. (1984). A Method for minimizing the impact of distributional assumptions in econometric models for duration data,
Econometrica,         52 (2): 271-320.

Hilde Karlsen <>:
Attrition from nursing sounds like a survival model, probably in
discrete time, using -logit- or -cloglog- with time dummies, not
-xtmelogit- (see for a textbook and self-guided course on discrete time survival models). If you have T years of data on each individual, all of whom are first-year nurses in period 1, and some of whom quit nursing in each of the subsequent years, with a variable nurse==1 when a nurse (and zero otherwise), an
individual identifier id, a year variable year, and a bunch of
explanatory variables x*, you can just:

tsset id year
bys id (year): g quit=(l.nurse==1 & nurse==0)
by id: replace quit=. if l.quit==1 | (mi(l.quit)&_n>1)
tab year, gen(_t)
drop _t1
logit quit _t* x*

and then work up to more complicated models with heterogeneous
frailty, etc. The main issues are that someone who quit nursing last year cannot quit nursing again this year, and people who never quit nursing might at some future point that you don't observe, which is
why you use survival models...

If you know the day they started work and the day they quit, you might
prefer a continuous-time model (help st).

I've been assuming you had data on people working as nurses, but
rereading your email, maybe you have data on breastfeeding mothers, though I suppose the same considerations apply (though with multiple
years of data on breastfeeding mothers, there is probably no

On Fri, Feb 13, 2009 at 9:19 AM, Hilde Karlsen <>

Dear statalisters,

This is probably a stupid question, but I've been searching around the
and in books and articles, and I've still not grasped the concept: When
performing a multilevel analysis of attrition from nursing using
and time (year) is the level 1 variable and individuals (id) is the
level 2
variable (i.e. years are clustered within individuals; I have a
file), how do I formulate the expectation related to this model? Why is
important to separate between these two levels?

I find it more intuitive to grasp the fact that individuals are
within schools, and that variables on the school level - as well as variables on the individual level - may influence e.g. which grades a
student gets.

I understand (at least I hope I understand) the point that when the same individuals are followed over a period of time, the individual's
are probably highly correlated, and that this implies a violation to
assumption about the heteroskedastic error-terms. As I see it, I could
used the cluster() - command (cluster(id))to 'avoid' this violation; however, I have to write an essay using multilevel analysis, so this is
an option.

I don't know if I'm being clear enough about what my problem is, but any information regarding this topic (how to grasp the concept of years
clustered in individuals) will be greatly appreciated.
I'm really sorry for having to ask you such an infantile question.. My colleagues and friends are not familiar with multilevel analyses, so I
know who to turn to.

Best regards,
*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index