Accounting for gender difference in adoption study using multinomial logistic regression

[Kwaku Agyei Adu]

Hi,
I am Kwaku from Kwame Nkrumah University of Science and Technology, Ghana. I am working on paper with an objective: the impact of adoption of improved rice technology on income and also time allocation. However, we would like to estimate the gender effect on the various analysis.

I have four levels of adoption:

1. No adoption
2. Adoption of improve rice variety
3. Adoption of fertilizer application
4. Adoption of both improved rice variety and fertilizer application.

Thus, I am using the multinomial logistic regression approach to estimate the determinants of adoption of the various technologies due to the four options in the first objective. The second level of analysis is to estimate the extent of adoption of the various technologies with the extent of utilization of rice variety and fertilizer being the dependent variables using the multivariate tobit regression approach.
The third level of analysis is where I want to assess the impact of adoption of the various technologies on two outcome variables: first income and second, time allocations using the BFG approach. I want to account for the gender difference in adoption. How do I do about this using the multinomial logistic approach. I know you include an interaction term between gender and a predictor variable. I read you interact gender with all the variables and use a joint test to test the significance of all the gender variables. One Doctor also mentioned I should rather interact gender with the adoption status (i.e gender*adoption of technologies).

My question is the approach right for the MNL and if so which of the two approaches will be right (gender*all the explanatory variables or gender*adoption status) and at which stage of the analysis must I include the interaction term; the first stage which estimates the determinants of adoption using the multinomial logit or the impact assessment level on the outcome variables? Again, I would like to know how to deal with interaction term (gender and predictor variables)

I’d much appreciate your generous suggestions.

[Carlo Lazzaro]

Kwaku:
welcome to the list.
Please read and act on the FAQ on how to post more effectively (in my opinion your post poses too many issues).
That said, the substantive matter is how to include gender in whatever regression model you're interested in:

i.gender

If you want to interact gender with different levels of technology adoption (let's name the latter -tech_adop-):

i.gender ##i.tech_adop

For more on categorical variables and interactions, see -help fvvarlist-.

As an aside, for the future, please post what tou typed and what Stata gave you back via CODE delimiters (as per FAQ again). Thanks.

[Kwaku Agyei Adu]

Dear Carlo,

Thank you for the prompt. I am sorry for not coming out clear. However, if I want to account for gender difference (1 for male and 0 otherwise) in adoption (four levels) using multinomial logit, how do I go about it?

[Carlo Lazzaro]

Kwaku:
simply include among your predictors:

i.gender##i.tech_adop

[Kwaku Agyei Adu]

Carlo,
I used this command
mlogit Adoption_Both_Tech gender log_Age Marital_Status log_Years_Sch log_Eco_Active_HH Land_Ownership log_Farm_Size Awarenes Improved_Rice_Tech Attend_Training_Programme Access_Credit Labour_Intensive_Improved_Rice_V Labour_Intensive_Fert_Use Capital_Intensive_Fert_Use log_seed_cost log_fert_cost log_TLU log_Domestic_Hours i.gender##i.Adoption_Both_Tech

for the estimation but unfortunately I am getting this feed from stata

note: 1.gender omitted because of collinearity

and then it keeps iterating.

[Carlo Lazzaro]

Kwaku:
I overlooked that -Adoption_Both_Tech- was your depvar (and not a predictor): hence you cannot interact idt with -gender-.
Try the following code:

mlogit Adoption_Both_Tech i.gender log_Age Marital_Status log_Years_Sch log_Eco_Active_HH Land_Ownership log_Farm_Size Awarenes Improved_Rice_Tech Attend_Training_Programme Access_Credit Labour_Intensive_Improved_Rice_V Labour_Intensive_Fert_Use Capital_Intensive_Fert_Use log_seed_cost log_fert_cost log_TLU log_Domestic_Hours

[Kwaku Agyei Adu]

Thank you Carlo,

I tried it and work. However, I get the same results for using same command except for using gender instead of i.gender. What is the implication?

Regards,
Kwaku

[Carlo Lazzaro]

Kwanku:
did you get a coefficient for -gender- this time?

[Kwaku Agyei Adu]

Carlo,
these are the two commands used and both produced the same results. I did have same coefficients for gender and i.gender in both estimations though

mlogit Adoption_Both_Tech gender log_Age Marital_Status log_Years_Sch log_Eco_Active_HH Land_Ownership log_Farm_Size Awarenes_Improved_Rice_Tech Attend_Training_Programme Access_Credit Labour_Intensive_Improved_Rice_V Labour_Intensive_Fert_Use Capital_Intensive_Fert_Use log_seed_cost log_fert_cost log_TLU log_Domestic_Hours


mlogit Adoption_Both_Tech i.gender log_Age Marital_Status log_Years_Sch log_Eco_Active_HH Land_Ownership log_Farm_Size Awarenes_Improved_Rice_Tech Attend_Training_Programme Access_Credit Labour_Intensive_Improved_Rice_V Labour_Intensive_Fert_Use Capital_Intensive_Fert_Use log_seed_cost log_fert_cost log_TLU log_Domestic_Hours

I tried this as well but i.gender was omitted because of collinearity.

mlogit Adoption_Both_Tech i.gender gender log_Age Marital_Status log_Years_Sch log_Eco_Active_HH Land_Ownership log_Farm_Size Awarenes_Improved_Rice_Tech Attend_Training_Programme Access_Credit Labour_Intensive_Improved_Rice_V Labour_Intensive_Fert_Use Capital_Intensive_Fert_Use log_seed_cost log_fert_cost log_TLU log_Domestic_Hours


Regards,
Kwaku

[Carlo Lazzaro]

Kwaku:
you results make sense.
in the first two codes yiou have the same results because gended has two categories only (hence -gender- =-i.gender-, however, I do prefer the latter way of writing categorical variables, that makes me prepared to deal with indicator variables with >2 categories);
in you third code, -gender- is omitted due to collinearity with -i.gender-.

[Kwaku Agyei Adu]

Carlo,
So can it be said that the differences in gender has been accounted for based on the second code and that comparisons can be made between males and females (using females as base category). If so, is the interpretation the same as that of the normal coefficient?

Regards,

Kwaku

[Carlo Lazzaro]

Kwaku:
if the reference category is 0 for female, yes.
Other things being equal, you should have obtained the same coefficient for -gender- and -i.gender-.
As an aside, for the future, please post what tou typed and what Stata gave you back via CODE delimiters (as per FAQ). Thanks.

[Kwaku Agyei Adu]

Carlo,
I did obtained same coefficient for -gender- and -i.gender- for the two codes. I'd much appreciate any reference on this discussion.

Regards,
Kwaku

[Carlo Lazzaro]

Kwaku:
the best reference would be -fvvarlist- entry in Stata .pdf manual (that you can quote in your research report).

[Kwaku Agyei Adu]

Thanks Carlo. I appreciate the time and education