Imagine I want to study career longevity of scientists. I define a career based on publication records. So a (publishing) career starts with the year of first publication and ends in the year of the last paper. It is a retrospective analysis with a time-frame from 1960-2010. I define career exit as not publishing any papers 10 years prior to study ending. I only include scientists who haven't published prior to 1960 and the first publication needs to be before 2000 to ensure a minimum of 10-year follow-up. Delayed entry is therefore allowed. My main interest are gender differences in career longevity. Additional to gender, I want to model some social network metrics (from co-author networks) as predictors. I assume for example that publishing with high status coauthors throughout my career will be beneficial for my career longevity.

My question is: Am I running into some kind of bias here? Do I have to control for the different starting times (cohorts?), for example cohort1 for everyone starting in 1960-1969, because they will be more likely to have an event since they started so early?

  • $\begingroup$ Is this just supposed to be a descriptive study, or is there some particular hypothesis about career longevity that you wish to test? Please add information about that, in particular details of any hypothesis of interest, by editing the question, as comments are easily overlooked and can even be lost. $\endgroup$
    – EdM
    Feb 3, 2021 at 13:22
  • $\begingroup$ Thank you @EdM. I edited my question! $\endgroup$
    – MaiMai
    Feb 3, 2021 at 18:48

1 Answer 1


Do I have to control for the different starting times (cohorts?), for example cohort1 for everyone starting in 1960-1969, because they will be more likely to have an event since they started so early?

You should certainly control for the actual starting date, but not because those individuals are more likely to have an event. The increasing probability of an event with time since the start of career is handled by the survival modeling itself.

Most important, based on my experience covering most of the time frame in question, the ability of women to have a long, successful academic career has changed dramatically over those decades. Thus you should consider an interaction between starting date and gender. Model starting date as continuous, with a flexible spline-based form, rather than breaking it up into decades. That will provide better modeling and should require fitting fewer parameters.

There also might be some changes in the structures of social networks over the years, depending on how those are defined. At least in my field, large numbers of authors per paper are much more likely in recent years than they were at the start of my career. Also, expectations for publications per year seem to have increased over calendar time, which might affect how you define "high status coauthors" or how you evaluate things like publication rates per year. Those sorts of things will require control for the actual starting date.

Am I running into some kind of bias here?

One thing that comes to mind is a significant competing risk that can lead to an end of publications: death. It's not clear how your model will take mid-career death into account. Also, you don't say how you will handle those who change gender. The case of Ben Barres illustrates both those problems. On the other hand, death doesn't immediately stop publication output; Ben Barres has been dead for over 3 years yet still has a publication that came out this year.

Another potential source of bias is differences among academic fields, and how those might have changed over time. Women have been more generally accepted into some fields than others, with a pattern that seems to have changed over the decades. Different fields have different possibilities and expectations about publication productivity, which could affect covariates you associate with individuals and who counts as a "high status coauthor" depending on the field.

If you are including time-dependent covariates in your model (which seems likely), you can also run into problems with survivorship bias. For example, my H-index probably depends as much on my having lived long enough to have nearly a 50-year publication history as it does on the quality of my publications.

These aren't easy issues to deal with. I'd suggest working with a local statistical expert from the start to make sure that the design and conduct of your study are solid.

  • $\begingroup$ Thank you so much for your helpful input. I really have to consider those points for my study. Another question came up (let me know if i should post it as a seperate question): The procedure that people define career longevity as the duration between first publication and last publication is something that I have seen a lot. I only wonder how they treat careers with for example two publications but with a time span of 40 years. Doesn't it distort my results for some reason that I am not aware of? $\endgroup$
    – MaiMai
    Feb 4, 2021 at 12:35
  • $\begingroup$ @MaiMai the issue is whether the definition of "career longevity" that you use distorts what one might consider to be "real carer longevity." So: do you think that someone with only 2 publications 40 years apart should be considered to have a 40-year-long career, not distinguished from someone with 4 papers per year over 40 years? That's an issue for your understanding and interpretation of the subject matter, not for statistics per se. $\endgroup$
    – EdM
    Feb 4, 2021 at 13:02
  • $\begingroup$ I see. A solution in my head, that would make the question a statistic topic again is: Instead of modeling survival as a function of time, can I model survival as a function of active years (years of active publishing)? So the career of the scientist with only 2 publications has a survival time of 2 active years. $\endgroup$
    – MaiMai
    Feb 4, 2021 at 13:40
  • $\begingroup$ @MaiMai that might be tricky, as it runs a risk of survivorship bias; see the link on that in the answer. You might consider modeling publications themselves as a function of time, with each publication considered an "event" in a repated-events model, then evaluating things like number of publications, publications per year, etc. See the R survival vignette for an introduction to how to model repeated events. $\endgroup$
    – EdM
    Feb 4, 2021 at 13:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.