I am studying the determinants of the location choice of a number of foreign investments in a given country. I am looking specifically at what drives the decision of companies to choose a particular industrial district for their investment among all districts available (354 of them). The determinants I am exploring are both variables describing the different 'alternatives' (that is, the different districts), such as labor cost, quality of local infrastructures, human capital...etc, and variables describing the characteristics of 'cases' (that is, the investments), such as the size of the investor, its country of origin, its degree of technological prowess...etc. To do so, I used an alternative-specific mixed logit model. The regression call was similar to this one:
After obtaining logit coefficients, I wanted to derive average marginal effects but encountered a few problems...
The first problem is both conceptual and practical. What I would like to obtain are average marginal (dydx) effects showing how much the probability of choosing a given (but random) district increases or decreases when a particular district characteristic (e.g. distr_char1) changes by one unit. I am assuming that the factors driving such decision are district characteristics - that is, a firm would be indifferent if it wasn't for those characteristics. When looking at the Stata help guide, however, I couldn't find a way to get this. The help file instead focuses on how to derive the marginal effect of choosing one specific district when one of its characteristics changes by one unit. Yet, I am not looking at the marginal effect related to one specific district, but at that of choosing any district in which that variable happened to change by one unit. Is there a way to obtain this with margins after cmmixlogit?
The second problem is somewhat linked to the first one because I would like to get the average response probability of choosing a certain (random) district at different levels of some alternative-specific (district) and case-specific (investor) variables. Again the problem is that it seems possible to obtain this response probability only when focusing on a certain, specific district and not when imagining a random district.
The third problem is technical. Since estimating the model took many hours, I saved the estimates to use them at a later stage. When I then use them and try to compute the marginal effects, however, I get a message saying that "no cases remain after removing invalid observations". I wonder how this is possible since the variable I am trying to get the marginal effect of was in the equation. Here is the code I have been using
Would you know how to solve (any of) these problems?
Code:
use dataset cmset investment_id district_id cmmixlogit chosen_district distr_char1 distr_char2 distr_char3....c.distr_char2#c.distr_char3 c.distr_char1#c.distr_char1), nocons casevars(invest_char1 invest_char2 invest_char3) scalemetric(unconstrained) difficult est save reg_results, replace
The first problem is both conceptual and practical. What I would like to obtain are average marginal (dydx) effects showing how much the probability of choosing a given (but random) district increases or decreases when a particular district characteristic (e.g. distr_char1) changes by one unit. I am assuming that the factors driving such decision are district characteristics - that is, a firm would be indifferent if it wasn't for those characteristics. When looking at the Stata help guide, however, I couldn't find a way to get this. The help file instead focuses on how to derive the marginal effect of choosing one specific district when one of its characteristics changes by one unit. Yet, I am not looking at the marginal effect related to one specific district, but at that of choosing any district in which that variable happened to change by one unit. Is there a way to obtain this with margins after cmmixlogit?
The second problem is somewhat linked to the first one because I would like to get the average response probability of choosing a certain (random) district at different levels of some alternative-specific (district) and case-specific (investor) variables. Again the problem is that it seems possible to obtain this response probability only when focusing on a certain, specific district and not when imagining a random district.
The third problem is technical. Since estimating the model took many hours, I saved the estimates to use them at a later stage. When I then use them and try to compute the marginal effects, however, I get a message saying that "no cases remain after removing invalid observations". I wonder how this is possible since the variable I am trying to get the marginal effect of was in the equation. Here is the code I have been using
Code:
use dataset, clear
Code:
cmset investment_id district_id
estimates use "reg_results"
estimates esample:
margins, dydx(distr_char2)
Would you know how to solve (any of) these problems?
