Dealing with attrition in longitudinal panel data

[Candice Yandam]

Dear all,

I am working with longitudinal panel data for individuals (years 2007 and 2009) and there is a very big attrition rate between the two waves (64%). For now, I had just been working with the individuals that remained in the sample. However, I think I need to address this attrition issue.

I have never dealt with attrition before and I do not know any methods to deal with it. I know for a fact that the data is MNAR (Missing Not At Random): the people who left the sample are most likely those who were affected by the 2008 economic crisis.

From what I've read online (and understood) so far, there are a few ways of dealing with the attrition issue such as reweighing the sample with IPW (Inverse Probability Weighting), and the use of refreshment panels (which I have with the 2009 wave).

I am wondering if any of you had to deal with attrition in longitudinal panel before, and what was the best and "easiest" technique to use- I am a bit constrained by time. I was thinking perhaps to match some individuals that left the sample with some who "come in" the sample in the 2009 refresher data (with Propensity Score Matching?) and work with that.

Thanks a lot for your help!

[Carlo Lazzaro]

Candice:
you've already preformed the first relevant step: diagnosing the mechanism underlying the missingness of your data.
Unfortunately, 64% of attrition is quite alarming and reinforces the MNAR diagnosis.
The best approach in this cases is setting up different imputation scenarios, assuming a variation in the attrition rate.
For more on this topic, you can take a look at: http://www.stefvanbuuren.nl/FIMD/index.html and http://www.stefvanbuuren.nl/publicat...2014%20HMD.pdf.

[Candice Yandam]

Carlo-

Thank you for your reply. It is normal in my sample to have a large attrition rate because migrant workers tend to move quite a lot, but this is definitely high. More importantly, it's not random... I will explore the option you advise.

[Clyde Schechter]

I have no advice to offer here; this is a difficult situation and there are no simple solutions.

I just wanted to commend both Candice and Carlo for being attentive to the importance of trying to understand the missingness mechanism and tailoring the response accordingly. In my end of the world, magical imputation is running rampant in applications where no thought has been given to the mechanisms generating missing values, and often in cases where it is obvious that the data are MNAR. The problem is exacerbated by journal reviewers who frequently demand the use of MI in these circumstances. This has been a pet peeve of mine for a while, and is rapidly rising to the top of my list of pet peeves. It has made my day to see two examples of people approaching this issue in a thoughtful manner!

[Dick Campbell]

I'm not sure I would jump to the conclusion that this is entirely a MNAR problem. The issue is whether the loss to follow up is conditional on variables which you measured in 2007. You want to know if the respondents you lost in 2009 differ in systematic ways from those who remained in terms of variables you were able to measure in 2007. Given what you said about the effects of the 2008 recession you may have some leverage here. Of course, even if you have variables which are associated with attrition some aspects of attrition could still be MNAR but I wouldn't give up on this quite so easily.

[Candice Yandam]

Clyde-

Thank you! I don't know if I will find a solution but I'll try anyway and keep on trying to understand the underlying mechanisms of this attrition.

Dick-

Thank you for your reply and input. I'm working on Chinese migrant workers and those who left the sample are those who were most affected by the recession and they have certain characteristics such as a lower level of education, more precarious jobs etc.. than those who remained in the sample. But, I guess I have to investigate more.

Thank you all!

[Suhyun Jung]

I wanted to follow up on this issue of attrition with refreshment samples. With refreshment samples, you can estimate the additive non-ignorable (AN) model by Hirano et al. (1998; 2001). Does STATA have a simple procedure to implement it? Or should it be done by using imputation commands to separately estimate all missing values?

[Carlo Lazzaro]

Suhyung:
as contributors to this forum come from many different research fields, please provide full reference of the literature you quoted. Thanks.

[Gaurav Dhamija]

Dear All,

I tried accessing these links mentioned by Carlo. However, they no longer exist.
Is there anyone who can share the updated link?

Candice:
you've already preformed the first relevant step: diagnosing the mechanism underlying the missingness of your data.
Unfortunately, 64% of attrition is quite alarming and reinforces the MNAR diagnosis.
The best approach in this cases is setting up different imputation scenarios, assuming a variation in the attrition rate.
For more on this topic, you can take a look at: http://www.stefvanbuuren.nl/FIMD/index.html and http://www.stefvanbuuren.nl/publicat...2014%20HMD.pdf.

[William Lisowski]

The website https://stefvanbuuren.nl/ now redirects to https://stefvanbuuren.name/ but it seems to be organized differently. Perhaps Carlo Lazzaro can find the equivalent URLs at the new location, or perhaps they will be easy to find for someone more familiar with the subject than I am.

[Carlo Lazzaro]

William:
many thanks for pointing this out. I was not aware that the link was changed.
Stef van Buuren's publications can be now found at https://stefvanbuuren.name/publication/ whereas his textbook on missing values can be purchased (among other options) via https://stefvanbuuren.name/publication/vanbuuren-2018/.

ETA: one of my favourite Stef's publication on how to deal with different types of missing values can be found at https://stefvanbuuren.name/publicati...Med%201999.pdf