Is it okay using apllo_weighting for propensity score weighting?

[Hyewon]

Dear Stephane,

I want to use propensity score weighting to control selection biases in experimental groups.
The datasets are Stated Choice survey, and looks like as follows;

Code: Select all

ID	choice	attirbute1		attribute2	 ... 	ipw
1	   1			1			0		0.3478
1	   3			0			1		0.3478
1	   5			1			1		0.3478
2	   3			0			1		3.7865
2 	   4			1			1		3.7865
2	   3			0			1 		3.7865

ipw row is for propensity scores,
I used ordered logit model since the choice part is in Likert scale, and I have two questions as follows;

1. Can I use apollo_weighting in propensity score weighting?
I tried weights = "ipw" in apollo_control and P = apollo_weighting(P, apollo_inputs, functionality) before prepareprob, which follows the manual.
I just add these two in my original basic code for Ordered logit.

2. If it is okay, however, the result is different from other, MASS package, simple ordered logit also.
I will attach the different results.

(1) Result from apollo

Code: Select all

Estimates:
                  Estimate        s.e.   t.rat.(0)    Rob.s.e. Rob.t.rat.(0)
b_auto_tram       -0.05929     0.03786     -1.5658     0.05149       -1.1515
b_auto_vehi        0.04170     0.04618      0.9030     0.16129        0.2585
b_bus             -0.18977     0.04646     -4.0851     0.14342       -1.3232
b_truck_person     0.24773     0.07118      3.4805     0.28958        0.8555
b_truck_guard      0.25147     0.04102      6.1298     0.06824        3.6850
b_video            0.10208     0.03814      2.6766     0.33802        0.3020
b_experi           0.68932     0.19335      3.5651     0.40888        1.6859
b_none             0.00000          NA          NA          NA            NA
tau_choice_1      -1.67527     0.04992    -33.5567     0.31206       -5.3685
tau_choice_2      -0.62904     0.04634    -13.5750     0.27961       -2.2497
tau_choice_3       0.77439     0.04662     16.6089     0.34088        2.2717
tau_choice_4       1.87547     0.05072     36.9775     0.54187        3.4611

(2) Result for MASS package in r

Code: Select all

Call:
polr(formula = ans ~ auto_tram + auto_vehi + bus + truck_person * 
    s21_1 + truck_guard + video1 + exp_group, data = rdata3, 
    weights = ipw)


Coefficients:
                      Value         Std. Error       t value
auto_tram          -0.06321    0.03787  -1.669
auto_vehi           0.04968    0.04620   1.075
bus                -0.18969    0.04646  -4.083
truck_person       -0.28461    0.05187  -5.487
s21_1               0.29061    0.05363   5.419
truck_guard         0.09386    0.04635   2.025
video              0.07793    0.03833   2.033
exp_group           0.55516    0.19438   2.856
truck_person:s21_1  0.16706    0.09305   1.795

Sorry for the differences in the variable name, As you can see, the coefficients are different especially t-values are really different.
For your understanding, there are 3 groups, video, experiment, and none. Therefore, I put none as a baseline.



Thank you for your kind supports for questions, I really want to use this package in my analysis since it is really straightforward to expand what I want to do.
Thank you so much.

With kindest regards,
Hyewon

[dpalma]

Hi Hyewon,

You seem to be using two additional explanatory variables in your MASS model: s21_1 and the interaction between s21_1 and truck_person (truck_person:s21_1). Those two variables do not seem to be included in the Apollo model.

Cheers
David