Multilevel Regression Methodology

[Fayed Yassine]

Dear All,

I'm studying the effect of geographical diversification of property assets on the standard deviation of real estate investment trusts (REITs) stock returns. I have a sample of 12 REITs from 2010Q1 to 2016Q4. I am studying the effect of their exposure to different U.S. states on the standard deviation of their returns (dependant variable). i have to 2 identification variables for REITs: the reitcode and the total number of state-level real estate markets the reit is exposed to (Total exposure variable in table). I also rank exposure to different states for each period. for instance, if REIT 1 has the largest share of its property portfolio on the 4th quarter of 2016 in NY, NY would be his first market for that date. There are 30 possible market ranks for each period. In a 1st step I want to know which state had the highest effect on the standard deviation of returns. Then for the 2nd step, if that state was the first investment market for the REIT, would that effect be higher than if it was its 2nd,3rd,4th market and so on. For the final step, I want to know whether the effect from the 2nd step is lower when REITs are exposed to a larger number of markets (Total exposure variable in table). From what I understand, a Multilevel linear regression would be a best choice but, how can I fix different observation values for the market rank and total exposure variables and see their effect on the dependant variable?

Thanks in advance for your help,

reitcode	Date	Market Rank	Total exposure	stdevR	NJ	TX	CA	NY
1	2016q4	1	4	0.12	0	0	0	1
1	2016q4	2	4	0.12	1	0	0	0
1	2016q4	3	4	0.12	0	0	1	0
1	2016q4	4	4	0.12	0	1	0	0
1	2016q4	5	4	0.12	0	0	0	0
1	2016q4	6	4	0.12	0	0	0	0
1	2016q4	7	4	0.12	0	0	0	0
1	2016q4	8	4	0.12	0	0	0	0
1	2016q4	9	4	0.12	0	0	0	0
1	2016q4	10	4	0.12	0	0	0	0
1	2016q4	11	4	0.12	0	0	0	0
1	2016q4	12	4	0.12	0	0	0	0
1	2016q4	13	4	0.12	0	0	0	0
1	2016q4	14	4	0.12	0	0	0	0
1	2016q4	15	4	0.12	0	0	0	0
1	2016q4	16	4	0.12	0	0	0	0
1	2016q4	17	4	0.12	0	0	0	0
1	2016q4	18	4	0.12	0	0	0	0
1	2016q4	19	4	0.12	0	0	0	0
1	2016q4	20	4	0.12	0	0	0	0
1	2016q4	21	4	0.12	0	0	0	0
1	2016q4	22	4	0.12	0	0	0	0
1	2016q4	23	4	0.12	0	0	0	0
1	2016q4	24	4	0.12	0	0	0	0
1	2016q4	25	4	0.12	0	0	0	0
1	2016q4	26	4	0.12	0	0	0	0
1	2016q4	27	4	0.12	0	0	0	0
1	2016q4	28	4	0.12	0	0	0	0
1	2016q4	29	4	0.12	0	0	0	0
1	2016q4	30	4	0.12	0	0	0	0
1	2016q3	1	3	0.23	1	0	0	0
1	2016q3	2	3	0.23	0	0	0	1
1	2016q3	3	3	0.23	0	1	0	0
1	2016q3	4	3	0.23	0	0	0	0
1	2016q3	5	3	0.23	0	0	0	0
1	2016q3	6	3	0.23	0	0	0	0
1	2016q3	7	3	0.23	0	0	0	0
1	2016q3	8	3	0.23	0	0	0	0
1	2016q3	9	3	0.23	0	0	0	0
1	2016q3	10	3	0.23	0	0	0	0
1	2016q3	11	3	0.23	0	0	0	0
1	2016q3	12	3	0.23	0	0	0	0
1	2016q3	13	3	0.23	0	0	0	0
1	2016q3	14	3	0.23	0	0	0	0
1	2016q3	15	3	0.23	0	0	0	0
1	2016q3	16	3	0.23	0	0	0	0
1	2016q3	17	3	0.23	0	0	0	0
1	2016q3	18	3	0.23	0	0	0	0
1	2016q3	19	3	0.23	0	0	0	0
1	2016q3	20	3	0.23	0	0	0	0
1	2016q3	21	3	0.23	0	0	0	0
1	2016q3	22	3	0.23	0	0	0	0
1	2016q3	23	3	0.23	0	0	0	0
1	2016q3	24	3	0.23	0	0	0	0
1	2016q3	25	3	0.23	0	0	0	0
1	2016q3	26	3	0.23	0	0	0	0
1	2016q3	27	3	0.23	0	0	0
1	2016q3	28	3	0.23	0	0	0
1	2016q3	29	3	0.23	0	0	0	0

[Phil Bromiley]

You didn't get a quick response. You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output (fixed spacing often helps) and sample data using dataex. Shorten your post to the absolute minimum necessary to generate the problem. Your paragraph stating the problem is very hard for me to understand. I also don't understand what market rank means. I'm not even clear what the observation unit is - it seems to be market rank within quarter, but what is market rank?

I know this isn't what you want to hear, but this kind of complex, multi-stage analysis is problematic. It is hard to have the estimates from one stage not mixed with the later stages in ways that invalidate the later analysis. I suggest you talk over the entire analysis design with someone local who can help you simplify and clarify it.