Dear all,
I have questionnaire data across three waves, year 0, year 5, and year 10. At max this data had 1124 mothers responding to a questionnaire on their health. I have harmonized a separate dataset on the local area unemployment of these women, by manually entering each womans local unemployment into the excel file that this questionnaire data comes from. I import this into Stata and would like to analyze this as panel data, so I do the following:
I have each womans id, their county id (geographic area) that these women are living in, they are also nested in family groups for which I have a family group id, however, as I drop anyone else from the family group who isn’t a mother from the sample, each family group now only contains the mother.
In my analysis I tested for attrition by creating a variable equal to one if mothers had left the sample, based on having filled a questionnaire in wave 1 but not in wave 2 and wave 3:
Following this I look at the differences between the sample stayers and the sample leavers, and whether this difference is significant:
Results suggest that health differs for leavers and stayers in the sample, and that there is a significant relationship between leaving the sample and health.
I obviously wanted to do something to deal with this attrition bias.
Searching the forums I followed the advice from this post to consider inverse probability of attrition weighting: https://www.statalist.org/forums/for...istrative-data
And followed the steps linked to here:
http://www.chronicpoverty.org/upload...N-revfinal.pdf
I cloned the health variable from earlier as cbinary_health and created a variable A (for attrition) that was equal to 1 if binary health in waves 2 and 3 was missing and 0 otherwise. I also generated a lagged health value, although I don’t know if I did this right as this is a study measured at years 0, 5 and 10, so maybe it needs to be lagged differently.
As guided by the attached document, I calculate a probit of those variables that I think may lead to attrition in health, these include education, bmi, medical card holding (a form of social health insurance), employment data, marital status, age, and the local area unemployment rate and lagged binary health.
.
Then I employ a Wald test for whether attrition is random on those variables that were significant in this probit
Results suggest that i.cown_education_y0 i.cemployment_y0 cage_y0 and lcbinary_health_y0 are significant predictors of attrition.
So when I calculate inverse probability weights below, I exclude the above as causing attrition.
When I initially did my analysis I used random effects models, clustering at the county level and in a linear probability model.
Following creating the I weights I would like to apply my weights to my random effects regressions in this panel data as follows:
I regress percentage unemployed and other variables on binary_health_y, which is the health across all waves of this panel data, i.e. the long health. The other variables included in this model are similarly those which have been changed from age_y0 age_y5 age_y10 to age_y as the data was changed from wide to long.
My analysis without the weights is fine, as you can see.
But when I apply the weights I have following problem:
My question is basically, what can I do? I would really like to stick to the random effects model with linear probability models in panel data, as this is what I’ve been working hard on for the past number of months, but is there another approach I should be taking here? Or another way I can make this work? GLLAMM had popped into my head but I don’t really know what implication this might hold here or how to apply it. I could really do with some advice.
I have questionnaire data across three waves, year 0, year 5, and year 10. At max this data had 1124 mothers responding to a questionnaire on their health. I have harmonized a separate dataset on the local area unemployment of these women, by manually entering each womans local unemployment into the excel file that this questionnaire data comes from. I import this into Stata and would like to analyze this as panel data, so I do the following:
Code:
reshape long health_y current_county_y psum_unemployed_total_cont_y i.own_educatin_y i.binmartatus_y i.medical_card_y, i(id) j(year)
. reshape long health_y current_county_y binary_health_y /*has_questionnaire_y*/ bmi_y binbmi_overweight_y binbmi_underweight_y binbmi_obese_y ord_bmi_y own_education_
> y medical_card_y employment_y binary_employment_y maritalstatus_y binmartatus_y age_y ord_age_y psum_unemployed_total_cont_y, i(id) j(year)
(note: j = 0 5 10)
Data wide -> long
-----------------------------------------------------------------------------
Number of obs. 1787 -> 5361
Number of variables 1181 -> 1148
j variable (3 values) -> year
xij variables:
health_y0 health_y5 health_y10 -> health_y
current_county_y0 current_county_y5 current_county_y10->current_county_y
binary_health_y0 binary_health_y5 binary_health_y10->binary_health_y
bmi_y0 bmi_y5 bmi_y10 -> bmi_y
binbmi_overweight_y0 binbmi_overweight_y5 binbmi_overweight_y10->binbmi_overweight_y
binbmi_underweight_y0 binbmi_underweight_y5 binbmi_underweight_y10->binbmi_underweight_y
binbmi_obese_y0 binbmi_obese_y5 binbmi_obese_y10->binbmi_obese_y
ord_bmi_y0 ord_bmi_y5 ord_bmi_y10 -> ord_bmi_y
own_education_y0 own_education_y5 own_education_y10->own_education_y
medical_card_y0 medical_card_y5 medical_card_y10->medical_card_y
employment_y0 employment_y5 employment_y10-> employment_y
binary_employment_y0 binary_employment_y5 binary_employment_y10->binary_employment_y
maritalstatus_y0 maritalstatus_y5 maritalstatus_y10->maritalstatus_y
binmartatus_y0 binmartatus_y5 binmartatus_y10->binmartatus_y
age_y0 age_y5 age_y10 -> age_y
ord_age_y0 ord_age_y5 ord_age_y10 -> ord_age_y
psum_unemployed_total_cont_y0 psum_unemployed_total_cont_y5 psum_unemployed_total_cont_y10->psum_unemployed_total_cont_y
-----------------------------------------------------------------------------
.
. xtset id year
panel variable: id (strongly balanced)
time variable: year, 0 to 10, but with gaps
delta: 1 unit
I have each womans id, their county id (geographic area) that these women are living in, they are also nested in family groups for which I have a family group id, however, as I drop anyone else from the family group who isn’t a mother from the sample, each family group now only contains the mother.
In my analysis I tested for attrition by creating a variable equal to one if mothers had left the sample, based on having filled a questionnaire in wave 1 but not in wave 2 and wave 3:
Code:
. drop if gender==1
(1,980 observations deleted)
* Total attrition left sample:
. generate leftsamp=.
(3,381 missing values generated)
. replace leftsamp = 1 if has_y5_questionnaire == 0 & has_y10_questionnaire == 0
(1,530 real changes made)
. replace leftsamp = 0 if has_y5_questionnaire == 1 | has_y10_questionnaire == 1
(1,851 real changes made)
.
.
.
. tab leftsamp
leftsamp | Freq. Percent Cum.
------------+-----------------------------------
0 | 1,851 54.75 54.75
1 | 1,530 45.25 100.00
------------+-----------------------------------
Total | 3,381 100.00
. tab has_y0_questionnaire
has_y0_ques |
tionnaire | Freq. Percent Cum.
------------+-----------------------------------
0 | 9 0.27 0.27
1 | 3,372 99.73 100.00
------------+-----------------------------------
Total | 3,381 100.00
. tab binary_health_y if leftsamp == 1
.
Following this I look at the differences between the sample stayers and the sample leavers, and whether this difference is significant:
Code:
. . tab binary_health_y
binary_heal |
th_y | Freq. Percent Cum.
------------+-----------------------------------
Bad | 595 28.19 28.19
Good | 1,516 71.81 100.00
------------+-----------------------------------
Total | 2,111 100.00
.
.
.
. tab binary_health_y if leftsamp == 1
binary_heal |
th_y | Freq. Percent Cum.
------------+-----------------------------------
Bad | 185 37.22 37.22
Good | 312 62.78 100.00
------------+-----------------------------------
Total | 497 100.00
.
.
.
tab binary_health_y leftsamp, column row nokey chi2 lrchi2 V exact gamma taub
binary_hea | leftsamp
lth_y | 0 1 | Total
-----------+----------------------+----------
Bad | 410 185 | 595
| 68.91 31.09 | 100.00
| 25.40 37.22 | 28.19
-----------+----------------------+----------
Good | 1,204 312 | 1,516
| 79.42 20.58 | 100.00
| 74.60 62.78 | 71.81
-----------+----------------------+----------
Total | 1,614 497 | 2,111
| 76.46 23.54 | 100.00
| 100.00 100.00 | 100.00
Pearson chi2(1) = 26.2308 Pr = 0.000
likelihood-ratio chi2(1) = 25.2839 Pr = 0.000
Cramér's V = -0.1115
gamma = -0.2704 ASE = 0.051
Kendall's tau-b = -0.1115 ASE = 0.023
Fisher's exact = 0.000
1-sided Fisher's exact = 0.000
.
.
I obviously wanted to do something to deal with this attrition bias.
Searching the forums I followed the advice from this post to consider inverse probability of attrition weighting: https://www.statalist.org/forums/for...istrative-data
And followed the steps linked to here:
http://www.chronicpoverty.org/upload...N-revfinal.pdf
I cloned the health variable from earlier as cbinary_health and created a variable A (for attrition) that was equal to 1 if binary health in waves 2 and 3 was missing and 0 otherwise. I also generated a lagged health value, although I don’t know if I did this right as this is a study measured at years 0, 5 and 10, so maybe it needs to be lagged differently.
Code:
gen lcbinary_health_y0 = (cbinary_health_y0 +1)
. gen A=1 if cbinary_health_y5==.& cbinary_health_y10==.
(3,510 missing values generated)
.
. replace A=0 if A!=1
(3,510 real changes made)
.
. tab A
A | Freq. Percent Cum.
------------+-----------------------------------
0 | 1,848 54.66 54.66
1 | 1,533 45.34 100.00
------------+-----------------------------------
Total | 3,381 100.00
. tab binary_health_y
binary_heal |
th_y | Freq. Percent Cum.
------------+-----------------------------------
Bad | 595 28.19 28.19
Good | 1,516 71.81 100.00
------------+-----------------------------------
Total | 2,111 100.00
Code:
**** BINARY HEALTH
* Calculate unrestricted attrition probit
* Binary health Attrition:
* Vars that might effect health
. xi: probit A cbmi_y0 i.cown_education_y0 i.cmedical_card_y0 i.cemployment_y0 i.cmaritalstatus_y0 cage_y0 cpsum_unemployed_total_cont_y0 lcbinary_health_y0, robust clus
> ter(current_county_y)
i.cown_ed~on_y0 _Icown_educ_1-6 (naturally coded; _Icown_educ_1 omitted)
i.cmedical_c~y0 _Icmedical__0-1 (naturally coded; _Icmedical__0 omitted)
i.cemploymen~y0 _Icemployme_1-8 (naturally coded; _Icemployme_1 omitted)
i.cmaritalst~y0 _Icmaritals_1-6 (naturally coded; _Icmaritals_1 omitted)
note: _Icemployme_5 != 0 predicts success perfectly
_Icemployme_5 dropped and 6 obs not used
Iteration 0: log pseudolikelihood = -1600.524
Iteration 1: log pseudolikelihood = -1495.6619
Iteration 2: log pseudolikelihood = -1495.0598
Iteration 3: log pseudolikelihood = -1495.0051
Iteration 4: log pseudolikelihood = -1494.9972
Iteration 5: log pseudolikelihood = -1494.9961
Iteration 6: log pseudolikelihood = -1494.9959
Iteration 7: log pseudolikelihood = -1494.9959
Probit regression Number of obs = 2,376
Wald chi2(19) = 4469.68
Prob > chi2 = 0.0000
Log pseudolikelihood = -1494.9959 Pseudo R2 = 0.0659
(Std. Err. adjusted for 30 clusters in current_county_y)
------------------------------------------------------------------------------------------------
| Robust
A | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------------------------+----------------------------------------------------------------
cbmi_y0 | -.0009359 .0110034 -0.09 0.932 -.0225023 .0206304
_Icown_educ_2 | -3.405426 .638231 -5.34 0.000 -4.656336 -2.154517
_Icown_educ_3 | -4.045781 .37698 -10.73 0.000 -4.784649 -3.306914
_Icown_educ_4 | -4.026321 .3782838 -10.64 0.000 -4.767743 -3.284898
_Icown_educ_5 | -3.972752 .4498496 -8.83 0.000 -4.854441 -3.091063
_Icown_educ_6 | -4.165562 .3707544 -11.24 0.000 -4.892227 -3.438897
_Icmedical__1 | .0902858 .1105417 0.82 0.414 -.126372 .3069436
_Icemployme_2 | .0486977 .3572737 0.14 0.892 -.6515459 .7489412
_Icemployme_3 | -.1855261 .4400484 -0.42 0.673 -1.048005 .6769528
_Icemployme_4 | -.4687752 .2495238 -1.88 0.060 -.9578329 .0202825
_Icemployme_5 | 0 (omitted)
_Icemployme_7 | -.2044312 .0910754 -2.24 0.025 -.3829358 -.0259267
_Icemployme_8 | -.2936189 .2797146 -1.05 0.294 -.8418495 .2546116
_Icmaritals_2 | .0404128 .214149 0.19 0.850 -.3793115 .4601371
_Icmaritals_4 | .2333091 .6966846 0.33 0.738 -1.132168 1.598786
_Icmaritals_5 | 1.065994 .8066381 1.32 0.186 -.5149877 2.646976
_Icmaritals_6 | .0826954 .1384134 0.60 0.550 -.1885898 .3539807
cage_y0 | -.0436179 .0093682 -4.66 0.000 -.0619793 -.0252565
cpsum_unemployed_total_cont_y0 | .0430288 .0257745 1.67 0.095 -.0074883 .0935459
lcbinary_health_y0 | -.2567768 .0923911 -2.78 0.005 -.4378599 -.0756936
_cons | 5.376694 .6043276 8.90 0.000 4.192233 6.561154
------------------------------------------------------------------------------------------------
Then I employ a Wald test for whether attrition is random on those variables that were significant in this probit
Code:
. test _Icown_educ_2 _Icown_educ_3 _Icown_educ_4 _Icown_educ_5 _Icown_educ_6 _Icemployme_2 _Icemployme_3 _Icemployme_4 _Icemployme_5 _Icemployme_7 _Icemployme_8 cage_y0
> lcbinary_health_y0
( 1) [A]_Icown_educ_2 = 0
( 2) [A]_Icown_educ_3 = 0
( 3) [A]_Icown_educ_4 = 0
( 4) [A]_Icown_educ_5 = 0
( 5) [A]_Icown_educ_6 = 0
( 6) [A]_Icemployme_2 = 0
( 7) [A]_Icemployme_3 = 0
( 8) [A]_Icemployme_4 = 0
( 9) [A]o._Icemployme_5 = 0
(10) [A]_Icemployme_7 = 0
(11) [A]_Icemployme_8 = 0
(12) [A]cage_y0 = 0
(13) [A]lcbinary_health_y0 = 0
Constraint 9 dropped
chi2( 12) = 2513.77
Prob > chi2 = 0.0000
. * Below we test if any of the above groups of variables are individually different from zero:
.
. test _Icemployme_2 _Icemployme_3 _Icemployme_4 _Icemployme_5 _Icemployme_7 _Icemployme_8
( 1) [A]_Icemployme_2 = 0
( 2) [A]_Icemployme_3 = 0
( 3) [A]_Icemployme_4 = 0
( 4) [A]o._Icemployme_5 = 0
( 5) [A]_Icemployme_7 = 0
( 6) [A]_Icemployme_8 = 0
Constraint 4 dropped
chi2( 5) = 8.88
Prob > chi2 = 0.1139
. test _Icown_educ_2 _Icown_educ_3 _Icown_educ_4 _Icown_educ_5 _Icown_educ_6
( 1) [A]_Icown_educ_2 = 0
( 2) [A]_Icown_educ_3 = 0
( 3) [A]_Icown_educ_4 = 0
( 4) [A]_Icown_educ_5 = 0
( 5) [A]_Icown_educ_6 = 0
chi2( 5) = 176.33
Prob > chi2 = 0.0000
. test cage_y0
( 1) [A]cage_y0 = 0
chi2( 1) = 21.68
Prob > chi2 = 0.0000
. test lcbinary_health_y0
( 1) [A]lcbinary_health_y0 = 0
chi2( 1) = 7.72
Prob > chi2 = 0.0054
So when I calculate inverse probability weights below, I exclude the above as causing attrition.
Code:
. * Calculate inverse probability weights
* First do the regression with everything in from before
.
.
. xi: probit A cbmi_y0 i.cown_education_y0 i.cmedical_card_y0 i.cemployment_y0 i.cmaritalstatus_y0 cage_y0 cpsum_unemployed_total_cont_y0 lcbinary_health_y0, robust clus
> ter(current_county_y)
i.cown_ed~on_y0 _Icown_educ_1-6 (naturally coded; _Icown_educ_1 omitted)
i.cmedical_c~y0 _Icmedical__0-1 (naturally coded; _Icmedical__0 omitted)
i.cemploymen~y0 _Icemployme_1-8 (naturally coded; _Icemployme_1 omitted)
i.cmaritalst~y0 _Icmaritals_1-6 (naturally coded; _Icmaritals_1 omitted)
note: _Icemployme_5 != 0 predicts success perfectly
_Icemployme_5 dropped and 6 obs not used
Iteration 0: log pseudolikelihood = -1600.524
Iteration 1: log pseudolikelihood = -1495.6619
Iteration 2: log pseudolikelihood = -1495.0598
Iteration 3: log pseudolikelihood = -1495.0051
Iteration 4: log pseudolikelihood = -1494.9972
Iteration 5: log pseudolikelihood = -1494.9961
Iteration 6: log pseudolikelihood = -1494.9959
Iteration 7: log pseudolikelihood = -1494.9959
Probit regression Number of obs = 2,376
Wald chi2(19) = 4435.22
Prob > chi2 = 0.0000
Log pseudolikelihood = -1494.9959 Pseudo R2 = 0.0659
(Std. Err. adjusted for 30 clusters in current_county_y)
------------------------------------------------------------------------------------------------
| Robust
A | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------------------------+----------------------------------------------------------------
cbmi_y0 | -.0009359 .0110034 -0.09 0.932 -.0225023 .0206304
_Icown_educ_2 | -3.405426 .6382683 -5.34 0.000 -4.656409 -2.154444
_Icown_educ_3 | -4.045781 .3770116 -10.73 0.000 -4.784711 -3.306852
_Icown_educ_4 | -4.026321 .3783153 -10.64 0.000 -4.767805 -3.284836
_Icown_educ_5 | -3.972752 .4498364 -8.83 0.000 -4.854415 -3.091089
_Icown_educ_6 | -4.165562 .3708327 -11.23 0.000 -4.892381 -3.438743
_Icmedical__1 | .0902858 .1105417 0.82 0.414 -.126372 .3069436
_Icemployme_2 | .0486977 .3572737 0.14 0.892 -.6515459 .7489412
_Icemployme_3 | -.1855261 .4400484 -0.42 0.673 -1.048005 .6769528
_Icemployme_4 | -.4687752 .2495238 -1.88 0.060 -.9578329 .0202825
_Icemployme_5 | 0 (omitted)
_Icemployme_7 | -.2044312 .0910754 -2.24 0.025 -.3829358 -.0259267
_Icemployme_8 | -.2936189 .2797146 -1.05 0.294 -.8418495 .2546116
_Icmaritals_2 | .0404128 .214149 0.19 0.850 -.3793115 .4601371
_Icmaritals_4 | .2333091 .6966846 0.33 0.738 -1.132168 1.598786
_Icmaritals_5 | 1.065994 .8066381 1.32 0.186 -.5149877 2.646976
_Icmaritals_6 | .0826954 .1384134 0.60 0.550 -.1885898 .3539807
cage_y0 | -.0436179 .0093682 -4.66 0.000 -.0619793 -.0252565
cpsum_unemployed_total_cont_y0 | .0430288 .0257745 1.67 0.095 -.0074883 .0935459
lcbinary_health_y0 | -.2567768 .0923911 -2.78 0.005 -.4378599 -.0756936
_cons | 5.376694 .6043547 8.90 0.000 4.19218 6.561207
------------------------------------------------------------------------------------------------
.
.
. gen sample=e(sample)
. predict pxav
(option pr assumed; Pr(A))
(1005 missing values generated)
.
* Repeat this regression excluding those things that cause attrition:
.
.
. xi: probit A cbmi_y0 i.cmedical_card_y0 i.cmaritalstatus_y0 cpsum_unemployed_total_cont_y0, robust cluster(current_county_y)
i.cmedical_c~y0 _Icmedical__0-1 (naturally coded; _Icmedical__0 omitted)
i.cmaritalst~y0 _Icmaritals_1-6 (naturally coded; _Icmaritals_1 omitted)
Iteration 0: log pseudolikelihood = -1796.0272
Iteration 1: log pseudolikelihood = -1730.1673
Iteration 2: log pseudolikelihood = -1729.9723
Iteration 3: log pseudolikelihood = -1729.972
Iteration 4: log pseudolikelihood = -1729.972
Probit regression Number of obs = 2,643
Wald chi2(7) = 301.88
Prob > chi2 = 0.0000
Log pseudolikelihood = -1729.972 Pseudo R2 = 0.0368
(Std. Err. adjusted for 30 clusters in current_county_y)
------------------------------------------------------------------------------------------------
| Robust
A | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------------------------+----------------------------------------------------------------
cbmi_y0 | .0013239 .0107957 0.12 0.902 -.0198353 .0224831
_Icmedical__1 | .2700192 .1163751 2.32 0.020 .0419282 .4981101
_Icmaritals_2 | .2895128 .1267742 2.28 0.022 .0410399 .5379856
_Icmaritals_4 | .5160599 .4218635 1.22 0.221 -.3107773 1.342897
_Icmaritals_5 | 1.258782 .6736031 1.87 0.062 -.0614555 2.57902
_Icmaritals_6 | .509274 .1101539 4.62 0.000 .2933764 .7251717
cpsum_unemployed_total_cont_y0 | .0336633 .0306106 1.10 0.271 -.0263324 .093659
_cons | -.6886433 .351835 -1.96 0.050 -1.378227 .0009407
------------------------------------------------------------------------------------------------
.
. predict pxres
(option pr assumed; Pr(A))
(738 missing values generated)
* After calculating the predicted probabilities from the restricted attrition probit, the inverse probability weights are calculated straightforwardly by taking the ratio of the restricted to unrestricted probabilities.
. gen attwght=pxres/pxav
(1,005 missing values generated)
Following creating the I weights I would like to apply my weights to my random effects regressions in this panel data as follows:
I regress percentage unemployed and other variables on binary_health_y, which is the health across all waves of this panel data, i.e. the long health. The other variables included in this model are similarly those which have been changed from age_y0 age_y5 age_y10 to age_y as the data was changed from wide to long.
My analysis without the weights is fine, as you can see.
Code:
** Consumption regressions (without and with attrition weights)
. *without inverse probability weights
. xtreg binary_health_y psum_unemployed_total_cont_y i.own_education_y i.maritalstatus_y i.medical_card_y i.employment_y age_y if gender==0 & sample==1, re robust cluste
> r(current_county_y)
Random-effects GLS regression Number of obs = 1,546
Group variable: id Number of groups = 792
R-sq: Obs per group:
within = 0.0375 min = 1
between = 0.0871 avg = 2.0
overall = 0.0753 max = 3
Wald chi2(19) = .
corr(u_i, X) = 0 (assumed) Prob > chi2 = .
(Std. Err. adjusted for 30 clusters in current_county_y)
-----------------------------------------------------------------------------------------------------------------------------
| Robust
binary_health_y | Coef. Std. Err. z P>|z| [95% Conf. Interval]
------------------------------------------------------------+----------------------------------------------------------------
psum_unemployed_total_cont_y | .0027172 .0025494 1.07 0.287 -.0022797 .007714
|
own_education_y |
Primary school education | .3478043 .1273368 2.73 0.006 .0982288 .5973797
Some secondary school | .6342217 .0702333 9.03 0.000 .4965669 .7718764
Complete secondary education | .6390553 .0491846 12.99 0.000 .5426552 .7354555
Some third level education at college, university, RTC | .6315469 .0668039 9.45 0.000 .5006137 .76248
Complete third level education at college, university, RTC | .7517092 .0704716 10.67 0.000 .6135874 .8898309
|
maritalstatus_y |
Cohabiting | -.0624296 .0361448 -1.73 0.084 -.133272 .0084129
Separated | -.1085929 .1341099 -0.81 0.418 -.3714435 .1542578
Divorced | -.0742946 .1289035 -0.58 0.564 -.3269409 .1783516
Widowed | -.2019116 .1486542 -1.36 0.174 -.4932684 .0894452
Single/Never married | -.0849537 .0381968 -2.22 0.026 -.1598181 -.0100893
|
medical_card_y |
Yes | -.1122467 .0331333 -3.39 0.001 -.1771867 -.0473066
|
employment_y |
Unemployed | -.0217951 .0447618 -0.49 0.626 -.1095266 .0659364
Unable to work owing to permanent sickness or disability | -.613174 .0479992 -12.77 0.000 -.7072507 -.5190973
At school/student | -.1256232 .0587738 -2.14 0.033 -.2408176 -.0104288
Seeking work for the first time | -.1833912 .0404457 -4.53 0.000 -.2626634 -.1041191
Employed | -.016472 .0243922 -0.68 0.499 -.0642799 .0313359
Self Employed | .0020492 .0499899 0.04 0.967 -.0959291 .1000276
Wholly retired from paid work | .0638361 .0266037 2.40 0.016 .0116938 .1159783
|
age_y | -.0023342 .0024538 -0.95 0.341 -.0071435 .0024751
_cons | .165668 .0709221 2.34 0.019 .0266633 .3046728
------------------------------------------------------------+----------------------------------------------------------------
sigma_u | .26997013
sigma_e | .34561966
rho | .37893873 (fraction of variance due to u_i)
-----------------------------------------------------------------------------------------------------------------------------
.
Code:
. *with inverse probability weights . xtreg binary_health_y psum_unemployed_total_cont_y i.own_education_y i.maritalstatus_y i.medical_card_y i.employment_y age_y [pw=attwght] if gender==0 & sample==1, re > robust cluster(current_county_y) pweight not allowed with between-effects and random-effects models
My question is basically, what can I do? I would really like to stick to the random effects model with linear probability models in panel data, as this is what I’ve been working hard on for the past number of months, but is there another approach I should be taking here? Or another way I can make this work? GLLAMM had popped into my head but I don’t really know what implication this might hold here or how to apply it. I could really do with some advice.
Comment