Propensity Score Matching and Panel Data

[Marie Blaise]

Dear Statalisters,

I am currently working on the life satisfaction of potential caregivers. More precisely, I analyse whether providing care would increase or decrease their level of life satisfaction.
As shown in the care literature, it exists a selection bias as people who turn caregivers and keep on providing support over time are usually women, poorer and have lower opportunity costs. To what I know, there is no clear consensus on the impact of life satisfation in selection into caregiving. In other words, I wonder whether more satisfied (or less) have higher (or lower) propensities to provide care.

I use panel data (LISS) from 2008 to 2018 (with a gap in 2016) gathering information on socio-demographic characteristics, life satisfaction and care patterns of Dutch potential cargeivers. I have estimated the effect of providing care (or not: dummy) on the life satisfaction level (from 0 to 10) using an OLS with pooled data, then a OLS with FE and then a2 step system GMM approach (introducing lags of the independent variable and of the endogenous variables) to deal with endogeneity issues, namely simultaneity, dynamic endogeneity and unobserved heterogeneity.
As I am also interested in a potential selection of caregivers in my sample, I wanted to perform a match between those who provide help (the treated) and those who do not (untreated) and then compare the differences. However, my difficulty comes from the application of such a method. As I have panel data on 9 years, I have first reshape my sample from long to wide but still stuck here... Indeed, an individual might decide to provide care in 2008 and 2009 then stop during few years and care again in 2017 and I am not sure how to deal with it.
I wonder whether I can match individuals per year ? If yes, I do not know how to aggreagte them then.

I hope that my explanation is clear enough.
Any kind of help would be appreciated, thanks a lot,

Bests
Marie

[Tom Scott]

This article just came out that seems important: https://link.springer.com/article/10...40-020-09455-9

In terms of adjusting for a time-varying treatment, I know there is inverse probability of treatment weighting as an option and I think also marginal structural models: https://static1.squarespace.com/stat...s.Ong.2016.pdf

[Marie Blaise]

Dear Tom Scott,

Thank you for the references, I have read these papers. Actually, for now, I do not want to estimate the propensity scores (I have to change the title of my post!). I want to know whether there is selection into caregiving through life satisfaction. I have thought about using a nearest neighbour approach, including times dummies and the lag of the life satisfaction (the intuition is that I want to compare individuals with the same baseline level of life satisfaction before the treatment in a given year).
More precisely, I have written the following command:

teffects nnmatch (Lifesat lagLS Gender i.Agecat Obj_health Educ i.Occupation_cat i.Marital Children Whours ln_st_Hninc i.Urban i.year) (Informal) if gmm_sample==1 (for now the default is one neighbour)

The goal is to estimate the differences between those who provide care and those who do not, matching covariates including the year and others (lagLS Gender i.Agecat Obj_health Educ i.Occupation_cat i.Marital Children Whours ln_st_Hninc i.Urban i.year).
For the total sample, I find that the difference between the treated and the untreated is negative and significant, which is in line with the main results that I get from OLS and GMM, leading to the conclusion that there is no clear selection in caregiving due to life satisfaction.. however, I am not sure of my interpretation. What do you think ?

Thanks again

Marie