Cannot estimate Margins following -npregress-

[Tim Murray]

I am trying to estimate a nonparametic regression using -npregress-. I first tried to estimate the following:

npregress kernel sin_totHomeExp age_female if P == 1 & age_female >= 50, vce(bootstrap, seed(1234))

I get an error saying that there are insignificant observations to compute the bootstrapped standard errors. So I tried adding the -nodrop- option and estimated the following:

npregress kernel sin_totHomeExp age_female if P == 1 & age_female >= 50, vce(bootstrap, seed(1234) nodrop)

I get results for this:

                              Observed Estimate        Bootstrapped Std. Err.         z                 P>|z|
Mean
sin_totHomeExp               5.837566                         0.411449                 141.88            0.000

Effect
age_female                  -0.0603533                       0.0048706                -12.39             0.000

After this, I want to get the marginal effects, so I run this line of code:

margins, at(age_female=(50(2)85))

and I get the following error:

data have changed since estimation.  Margins and predictafter npregress will not work if your covariates, your dependent variable, or the predictions from npregress have changed since estimation. r(198)

unfortunately, I cannot share the data because I am working with a restricted data set, but I am hoping someone has some insight as to the problem. I am running the regression and margins command at the same time, so I am not changing the data or anything at all inbetween npregress and margins.

Thanks for any help!
Tim

[FernandoRios]

Hi Tim
I think this problem will be very hard to diagnose without additional information or a replicable example.
Perhaps the most daunting problem i see is that bootstrap standard errors cannot be directly applied. My guess is that age does not have enough support (around the tails), which may be causing this.
Can you provide a tabulation of age_female, and perhaps a tabstat (mean and standard errors) of sin_totHomeExp by age_female? In both cases using the sample defined by "if P == 1 & age_female >= 50"
That may help have a better idea of your data, and why npregress is not working.

Fernando

[Tim Murray]

Hey Fernando, age_female is just the age, ranges from 50-102. I included the -tabstat- output with mean and standard deviation

Age_Female	Mean	SD
50	6.358338	2.996968
51	6.245804	3.573873
52	6.927383	2.74296
53	7.213501	2.468586
54	6.486439	3.284938
55	6.459813	3.494563
56	6.718764	3.008857
57	6.151103	3.326089
58	6.500867	3.326089
59	6.405158	3.308358
60	6.830825	2.853765
61	6.233559	3.452721
62	6.477195	3.25382
63	6.537533	3.374326
64	6.374115	3.309824
65	6.390091	3.231187
66	6.135569	3.414742
67	6.160622	3.362239
68	6.075977	3.233201
69	6.080395	3.335637
70	5.66325	3.588735
71	5.597188	3.50087
72	5.769733	3.40418
73	5.75599	3.529218
74	5.548663	3.432273
75	5.363811	3.494965
76	5.850755	3.459182
77	5.562073	3.396962
78	5.150945	3.65316
79	4.785413	3.703598
80	5.390926	3.567981
81	5.761652	3.362377
82	5.091838	3.624754
83	4.655346	3.605155
84	5.026394	3.680572
85	5.260563	3.559989
86	4.996206	3.433092
87	4.901424	4.169815
88	4.922944	3.838504
89	4.4093194	3.957232
90	3.321275	3.821344
91	4.769178	3.955306
92	5.144357	4.116713
93	5.358893	3.568693
94	.	.
95	1.569964	3.510547
96	.	.
97	2.273401	3.937646
98	.	.
99	6.848347	.
100	.	.
101	.	.
102	.	.

[FernandoRios]

I see. So, my guess is that npregress is having problems estimating the model when age>90. Probably you do not have enough observations above that threshold to run bootstrap standard errors, or estimate the npregress kernel.
would you try running your command, say, for age<90 ? or perhaps 80?
Since npregress is based on local regressions, i think the estimations would not be affected much by excluding some outlier observations.

[Tim Murray]

Thanks Fernando, I have tried that, I guess the error is with calculating the standard errors with bootstrapping, when I run the regression without bootstrapping the standard errors it works just fine. There are insufficient observations to bootstrap the standard errors without the -nodrop- option coded.

I saw this in this post https://www.statalist.org/forums/for...with-summarize