ApolloChoiceModelling forum

Hi

I am running a MNL-model in wtp-space. From earlier posts here in the forum I am aware that the priors used to estimate wtp parameters are crucial for the model output. I have run the apollo_searchStart code, and plugged in the result from this code as priors in the model. This only returned wtp-estimates identical to the priors, and thus to the output from the search-code. I have tried to apply model estimation results using only the pilot data as priors. And I have set all priors equal to zero (except for cost parameter). Every set of priors give widely different final wtp-estimates.

Is there a procedure to follow in order to get reliable wtp-estimates in a MNL-model?

Here is the model I try to estimate:
### List of utilities: these must use the same names as in mnl_settings, order is irrelevant
V = list()
V[['Alt1']] = cost*(seabed*SB1 + salmon*SAL1 + jobs*JOB1 + (COST1/1000))+ asc_SQ
V[['Alt2']] = cost*(seabed*SB2 + salmon*SAL2 + jobs*JOB2 + (COST2/1000))
V[['Alt3']] = cost*(seabed*SB3 + salmon*SAL3 + jobs*JOB3 + (COST3/1000))

Using the following priors:
asc_SQ = 0.1178
seabed = 0.1880
salmon = 0.1388
jobs = -0.0305
cost = -0.2580,

the model output is as follows;

Estimation method : bfgs
Model diagnosis : successful convergence
Number of individuals : 1123
Number of observations : 8546

Number of cores used : 1
Model without mixing

LL(start) : -9173.329
LL(0) : -9388.741
LL(final) : -9173.329
Rho-square (0) : 0.0229
Adj.Rho-square (0) : 0.0224
AIC : 18356.66
BIC : 18391.92


Estimated parameters : 5
Time taken (hh:mm:ss) : 00:00:1.72
pre-estimation : 00:00:0.8
estimation : 00:00:0.21
post-estimation : 00:00:0.71
Iterations : 2
Min abs eigenvalue of Hessian : 126.6346

Estimates:
Estimate s.e. t.rat.(0) p(2-sided) t.rat(1) p(2-sided)
asc_SQ 0.11780 0.085775 1.373 0.169648 -10.29 0.000
seabed 0.18801 0.018727 10.039 0.000000 -43.36 0.000
salmon 0.13880 0.022163 6.263 3.783e-10 -38.86 0.000
jobs -0.03049 0.009743 -3.129 0.001754 -105.77 0.000
cost -0.25800 0.025900 -9.961 0.000000 -48.57 0.000
Rob.s.e. Rob.t.rat.(0) p(2-sided) Rob.t.rat.(1) p(2-sided)
asc_SQ 0.090477 1.302 0.192928 -9.751 0.000
seabed 0.021584 8.710 0.000000 -37.620 0.000
salmon 0.018852 7.363 1.803e-13 -45.681 0.000
jobs 0.009529 -3.199 0.001378 -108.142 0.000
cost 0.026816 -9.621 0.000000 -46.913 0.000


Best regards,
Margrethe

Margrethe

so first, a point on notation/language. What you are talking about here are not priors, but starting values. You are using classical estimation, not Bayesian estimation.

Next, the reason your model below converges immediately to the starting values is that the search start algorithm itself conducts some estimation steps, and seems to have ended with a set of values that are a maximum likelihood estimate.

In relation to your statement that you get different results with different starting values, can you first confirm what version of Apollo you are using

Thanks

Stephane