meglm with family(binomial) and link(logit)

[Robert Wagner]

Hi All,

I am using -meglm- with binomial family and logit link to estimate the probability of job vacancies being filled at company level. Several explanatory variables are included at firm and regional level.

The simplified form of the estimation equation is as follows:

meglm share_positions_filled independent_variables|| region: || company_id:, family(binomial) link(logit)

I am concerned that the chosen model does not fit the structure of the dependent variable, which represents the number of filled positions divided by the number of offered positions. It therefore takes on values in the interval [0;1], with the values 0 and 1 accounting for almost half of all observations.

My question: Does this model make sense in the context of the structure of the dependent variable, or would another model specification be more suitable?


I also modeled the dependent variable as binomial by specifying it as the number of successes (positions_filled) from a series of binomial trials (positions_offered).
For these cases meqrlogit and melogit were used:

me(qr)logit positions_filled_total independent_variables  || region: || company_id:, binomial(positions_offered)

Unfortunately for both melogit and meqrlogit, the error ‘initial values not feasible’ occurred.


Thank you in advance for your helpful answers!

[Clyde Schechter]

Unfortunately for both melogit and meqrlogit, the error ‘initial values not feasible’ occurred.

Before you give up on this:

1. You might have better luck with:

gen `c(obs_t)' obs_no = _n
expand 2
by obs_no, sort: gen byte filled = _n-1
gen fwt = cond(filled, positions_filled_total, positions_offered - positions_filled_total)
me(qr)logit filled independent_variables [fweight = fwt] || region: || company_id:

2. The problem may be attributable to one or more of your independent variables. Try a simpler model with only the independent variable that is the focus of your research question. If that model converges, try adding back other independent variables one at a time. Probably you can get most of your independent variables back, and that model might be satisfactory.

I suspect others will have suggestions for alternatives to the logistic model as well.

[Robert Wagner]

Thank you very much for your valuable input Clyde, I will try out the suggested approach.