Dear,
I am working on a longitudinal data set with an unbalanced panel survey and includes 6 survey years. One of my committee members pointed out the potential timing match problem between the dependent variable (PL take-up, which is binary) and independent variables; also, he asked results as a marginal effect outcome. Moreover, another committee member was curious about adding an interaction term for education and wage(ln). Therefore, as following of logistic retrogression without interaction term (lr1) and with the interaction term (lr2), which was eduwage variable:
1. I used the random effect model without interaction term (re1) and with the interaction term (re2). [xtlogit PL ...., re]
2. I used the mixed effect model by grouping the survey years without interaction term (me1) and with the interaction term (me2). [melogit PL ...., || Wyear: ]
re1 findings:
Random-effects logistic regression Number of obs = 560
Group variable: OPID Number of groups = 516
Random effects u_i ~ Gaussian Obs per group:
min = 1
avg = 1.1
max = 3
Integration method: mvaghermite Integration pts. = 12
Wald chi2(20) = 48.33
Log likelihood = -233.16566 Prob > chi2 = 0.0004
LR test of rho=0: chibar2(01) = 0.44 Prob >= chibar2 = 0.254
re2 findings:
Random-effects logistic regression Number of obs = 560
Group variable: OPID Number of groups = 516
Random effects u_i ~ Gaussian Obs per group:
min = 1
avg = 1.1
max = 3
Integration method: mvaghermite Integration pts. = 12
Wald chi2(21) = 48.05
Log likelihood = -232.20716 Prob > chi2 = 0.0007
LR test of rho=0: chibar2(01) = 0.34 Prob >= chibar2 = 0.281
me1 findings:
Mixed-effects logistic regression Number of obs = 560
Group variable: Wyear Number of groups = 6
Obs per group:
min = 28
avg = 93.3
max = 316
Integration method: mvaghermite Integration pts. = 7
Wald chi2(20) = 94.49
Log likelihood = -231.83373 Prob > chi2 = 0.0000
LR test vs. logistic model: chibar2(01) = 3.10 Prob >= chibar2 = 0.0391
me2 findings:
Mixed-effects logistic regression Number of obs = 560
Group variable: Wyear Number of groups = 6
Obs per group:
min = 28
avg = 93.3
max = 316
Integration method: mvaghermite Integration pts. = 7
Wald chi2(21) = 92.89
Log likelihood = -231.02178 Prob > chi2 = 0.0000
LR test vs. logistic model: chibar2(01) = 2.71 Prob >= chibar2 = 0.0500
3. I have tried the conditional model for the year grouping as well, but marginal effect calculation was problematic, so I decided to try these two. Literature suggested random effect, but I have been considering mixed effect could be an answer for time matching problem. However, when I checked the Hausman test, the results were so much confusing.
hausman re1 lr1
* chi(19) = 0.42
* Prob>chi2 = 1.0000
hausman me1 lr1
* chi(19) = 2.22
* Prob>chi2 = 1.0000
hausman me1 re1
* 5.84
* Prob>chi2 = 0.9983
hausman lr2 lr1
* 1.99
* Prob>chi2 = 1.0000
hausman re2 re1
* = 2.20
* Prob>chi2 = 1.0000
hausman me2 me1
* 1.78
* Prob>chi2 = 1.0000
I have been working on and looking at the forums for weeks to figure out and eliminate potential robustness issues, but I could find the solution; besides, the deadline for submission is very soon.
Please help me to understand and choose a suitable model for my thesis. (This is my first post; hopefully, I've explained clearly)
Best,
I am working on a longitudinal data set with an unbalanced panel survey and includes 6 survey years. One of my committee members pointed out the potential timing match problem between the dependent variable (PL take-up, which is binary) and independent variables; also, he asked results as a marginal effect outcome. Moreover, another committee member was curious about adding an interaction term for education and wage(ln). Therefore, as following of logistic retrogression without interaction term (lr1) and with the interaction term (lr2), which was eduwage variable:
1. I used the random effect model without interaction term (re1) and with the interaction term (re2). [xtlogit PL ...., re]
2. I used the mixed effect model by grouping the survey years without interaction term (me1) and with the interaction term (me2). [melogit PL ...., || Wyear: ]
re1 findings:
Random-effects logistic regression Number of obs = 560
Group variable: OPID Number of groups = 516
Random effects u_i ~ Gaussian Obs per group:
min = 1
avg = 1.1
max = 3
Integration method: mvaghermite Integration pts. = 12
Wald chi2(20) = 48.33
Log likelihood = -233.16566 Prob > chi2 = 0.0004
LR test of rho=0: chibar2(01) = 0.44 Prob >= chibar2 = 0.254
re2 findings:
Random-effects logistic regression Number of obs = 560
Group variable: OPID Number of groups = 516
Random effects u_i ~ Gaussian Obs per group:
min = 1
avg = 1.1
max = 3
Integration method: mvaghermite Integration pts. = 12
Wald chi2(21) = 48.05
Log likelihood = -232.20716 Prob > chi2 = 0.0007
LR test of rho=0: chibar2(01) = 0.34 Prob >= chibar2 = 0.281
me1 findings:
Mixed-effects logistic regression Number of obs = 560
Group variable: Wyear Number of groups = 6
Obs per group:
min = 28
avg = 93.3
max = 316
Integration method: mvaghermite Integration pts. = 7
Wald chi2(20) = 94.49
Log likelihood = -231.83373 Prob > chi2 = 0.0000
LR test vs. logistic model: chibar2(01) = 3.10 Prob >= chibar2 = 0.0391
me2 findings:
Mixed-effects logistic regression Number of obs = 560
Group variable: Wyear Number of groups = 6
Obs per group:
min = 28
avg = 93.3
max = 316
Integration method: mvaghermite Integration pts. = 7
Wald chi2(21) = 92.89
Log likelihood = -231.02178 Prob > chi2 = 0.0000
LR test vs. logistic model: chibar2(01) = 2.71 Prob >= chibar2 = 0.0500
3. I have tried the conditional model for the year grouping as well, but marginal effect calculation was problematic, so I decided to try these two. Literature suggested random effect, but I have been considering mixed effect could be an answer for time matching problem. However, when I checked the Hausman test, the results were so much confusing.
hausman re1 lr1
* chi(19) = 0.42
* Prob>chi2 = 1.0000
hausman me1 lr1
* chi(19) = 2.22
* Prob>chi2 = 1.0000
hausman me1 re1
* 5.84
* Prob>chi2 = 0.9983
hausman lr2 lr1
* 1.99
* Prob>chi2 = 1.0000
hausman re2 re1
* = 2.20
* Prob>chi2 = 1.0000
hausman me2 me1
* 1.78
* Prob>chi2 = 1.0000
I have been working on and looking at the forums for weeks to figure out and eliminate potential robustness issues, but I could find the solution; besides, the deadline for submission is very soon.
Please help me to understand and choose a suitable model for my thesis. (This is my first post; hopefully, I've explained clearly)
Best,
Comment