Receiving R(2000) "outcome does not vary" for Heckprob even if DV is coded correctly

[Miguel Rey]

Hi guys!

Currently doing an experiment regarding infidelity, my dependent variable is sexual_affair which when tab-ed gives you a 0 and a 1, but when I run the heckprob command, it still gives me the error. What may be the problem?

my main equation is a bit long: heckprob sexual_affair i.educationlevel_men age_men i.relative_productivity Health_Vulnerability i.educationlevel_spouse i.age_spouse workpast12months_spouse Health_VulnerabilitySpouse sexratio infant_mortality life_expectancy premaritalsex, select(paid_sex refusesex_havesex_withanother sex_with_man women_noright_condom embarassed_tobuy_condom keep_it_asecret)

So I tried to use 3 variables (the DV, one explanatory and one in the select just to see if it would work)

heckprob sexual_affair urban, select( paid_sex)

. tab sexual_affair

Committed |
Sexual |
Infidelity | Freq. Percent Cum.
------------+-----------------------------------
0 | 1,946 81.76 81.76
1 | 434 18.24 100.00
------------+-----------------------------------
Total | 2,380 100.00


. tab urban

urban | Freq. Percent Cum.
------------+-----------------------------------
0 | 1,230 51.68 51.68
1 | 1,150 48.32 100.00
------------+-----------------------------------
Total | 2,380 100.00



. tab paid_sex

paid_sex | Freq. Percent Cum.
------------+-----------------------------------
0 | 2,202 92.52 92.52
1 | 178 7.48 100.00
------------+-----------------------------------
Total | 2,380 100.00

But it still gives me:


selection equation:
outcome does not vary; remember:
0 = negative outcome,
all other nonmissing values = positive outcome
r(2000);

Thank you in advance!!!

[Nick Cox]

It's the joint distribution that's crucial here, not the marginal distributions. Try

 
ssc inst groups 
groups sexual_affair urban paid_sex

[Miguel Rey]

Thank you for the reply!!

It gives me this:

groups sexual_affair urban paid_sex

+-----------------------------------------------+
| sexual~r urban paid_sex Freq. Percent |
|-----------------------------------------------|
| 0 0 0 946 39.75 |
| 0 0 1 52 2.18 |
| 0 1 0 852 35.80 |
| 0 1 1 96 4.03 |
| 1 0 0 220 9.24 |
|-----------------------------------------------|
| 1 0 1 12 0.50 |
| 1 1 0 184 7.73 |
| 1 1 1 18 0.76 |
+-----------------------------------------------+

Does it mean that I can no longer use heckprob due to the small frequencies? Or is there a remedy? Thanks a lot!

[ben earnhart]

+-----------------------------------------------+
| sexual~r urban paid_sex Freq. Percent |
|-----------------------------------------------|
| 0 0 0 946 39.75 |
| 0 0 1 52 2.18 |
| 0 1 0 852 35.80 |
| 0 1 1 96 4.03 |
| 1 0 0 220 9.24 |
|-----------------------------------------------|
| 1 0 1 12 0.50 |
| 1 1 0 184 7.73 |
| 1 1 1 18 0.76 |
+-----------------------------------------------+

[Richard Williams]

Actually there is an additional complication: cases with missing data will get dropped. So that may cause the outcome does not vary message.

[Miguel Rey]

Thank you sir Williams! May I ask if the variables are all coded as 0 and 1, with those as "." replaced with 0. Would that still be the case? Thank you!

[Guest]

I am not sure that I understand what you mean. It helped me a great deal to scrutinize the data in the example (http://www.stata.com/manuals13/rheck...f#rheckoprobit) and compare it to the setup of my own data. In this example, you see that :
1. the selection variable must be [0, 1]
2. the dependent variable is missing for all values selection == 0; recoding missings to zero of your dependent variable will not change the results, since the values of "dependent" will be ignored iff "selection == 0". To see this, add the line "replace satisfaction=0 if missing(satisfaction)" before the heckman probit model in the abovementioned example: your results will be the same.
3. the dependent variable ranges from 1 onward, i.e. [1, 2, ...]
also (and I add this because it took me a while to stumble across this) the selection model has an equality sign between the dependent variable, while the probit itself does not.

[Natalie Kyrg]

Dear Members,
I have a similar problem with my heckprob model.
I am receiving this message after running my specification: no observations r(2000);

Could I please ask for your advice whether I have set up my specification correctly?

Here is my model: heckprob achieved reg_CA reg_EEC reg_RUS reg_SEE reg_SEM reg_TUR sec_fi sec_ica sec_infra ins_D ln_size_agg ln_total ratios ln3_length ln_effect implement fm_SA ln_prep cofin_N synd_N disburs_ratio ln_disburs_cum ln_disburs_no exit_year ln_pl_c noi_c npc_c spc_c ln3_pd_start ln3_pd_exit own_priv sover_N sme_flag ln2_avg_GDP avg_CPI ln_avg_INF fh_2003_F fh_2003_NF fh_2015_F fh_2015_NF trans_yrs avg_POP avg_STAB avg_CORR avg_BURE, select (pselect2 = reg_CA reg_EEC reg_RUS reg_SEE reg_SEM reg_TUR ln_size_agg exp_date status_d life_ba life_cr life_ex life_fr stat_can sec_fi sec_ica sec_infra ins_D sovert_dir sovert_gur sec_nv sec_s ln3_pd_start ln_pl_c noi_c npc_c spc_c exp_y ln2_avg_GDP avg_CPI ln_avg_INF fh_2003_F fh_2003_NF fh_2015_F fh_2015_NF trans_yrs avg_POP avg_STAB avg_CORR avg_BURE )

My DV ‘achieved’ refers to the success of a project performance (1 – achieved, 0 otherwise) and my explanatory variables and controls include a mixed of variables which could determine project performance (e.g. region, sector, size of the project, project length and so on).

In my selection equation I have similar explanatory variables but few new ones like project status dummies, life cycle stage dummies etc.

My selection variable (i.e. pselect2) is set correctly (1 – project selected, 0 – project not selected)

I am not sure what to do with the missing values as I understood that I should not replace them with zeros etc as heckprob takes care of that. Also, I have some ‘real’ missing values inside each of the sub-sample and wondering if they could be stopping my regressions?

I am attaching the summary table with missing values indication as well as high level overview of the variables I am including in my specification

Look forward to hearing from you.
Regards,
Natalie

[Puja Biswas]

probit school i.sex loginc i.motivation

outcome does not vary; remember:
0 = negative outcome,
all other nonmissing values = positive outcome
r(2000);
Please provide a solution to this problem while running a probit model

[Jared Greathouse]

All of your values for school that are in the regression are either 0 or 1.

Welcome to Statalist, Puja. Just so you know, you should've shared an example of your data, but fortunately that's not needed here.