# When can I assume a variable as constant

#### benben93

##### New Member
Hey guys,

I'm new to the forum and have the following setting. I work with panal data and have a variable (self reported risk aversion) in the years 2002, 07, 11 but I need it for the years 2004 and 2009. How do I handle it? I would like to assume it as a constant but what year do I take. Should I create the average risk aversion in between the years? What are the requirements on the variance to assume a variable as constant? Is it reasonable to drop observations with a high variation in between the years? Sorry for the many questions at once but I really donˋt know how to handle this. Most of the papers I read just assumed risk aversion as constant but in my dataset it varies (properbly because of the financial crisis) and I would like to have a reasonable approach on that for my research question very important variable.

I didnˋt even found literature on that question (properbly I googled wrong). So if you guys have a literature advice, I would really appreciate it. I would also appreciate if you guys had similar problems and you can tell me how you solved it.

#### fed2

##### Member
I guess the only thing that comes to mind from biostats is the notion of last observation carried forward that comes out in, for example diet studies. The idea is that the last observed observation shows the least improvement, and is in some sense the most conservative imputation possible given the study. Its like how if you lose a metro ticket, they give you the highest fair, its the biggest possible penalty. I don't know if LOCF cocept exists in your field cuz i don't really know what that is.

#### noetsi

##### Fortran must die
Do you mean time invariant when you say constant.

Multilevel models deal with fixing something to zero based on specific tests. I don't work with panel data but I would guess they do the same.

#### benben93

##### New Member
I guess the only thing that comes to mind from biostats is the notion of last observation carried forward that comes out in, for example diet studies. The idea is that the last observed observation shows the least improvement, and is in some sense the most conservative imputation possible given the study. Its like how if you lose a metro ticket, they give you the highest fair, its the biggest possible penalty. I don't know if LOCF cocept exists in your field cuz i don't really know what that is.
Thanks for your answer I really appreciate is. I had the same thought but the risk aversion don't has a linear trend. But I got an idea because of your answer. Maybe I take always the minimum in between two periods for the most conservitive approach possible. Have a nice day

#### benben93

##### New Member
Do you mean time invariant when you say constant.

Multilevel models deal with fixing something to zero based on specific tests. I don't work with panel data but I would guess they do the same.
Hey, thanks for your answer. It really helps. Time invariance is exactly what I meant. On one hand there are a lot of papers who assume risk aversion as constant but on the other hand I found papers who show that risk aversion has a time variant component. I will take a look at this model class and the tests for time invariance.

#### noetsi

##### Fortran must die
Essentially you are asking, or the research is, if there is a random effect I think. Its nice to have literature to test -I almost never have this in my field.