Hi I have a question whether we can assess the validity of a model for a data set using prediction.

<Situation>
1. I have a data set consists of 21 countries.
2. I ran the multinomial logit regression using only the first 20 countries.

<What I want to check>
1. my model specification is valid for the 21th country (that seems to be heterogenous)

<What I have done>
Let y is the "nominal" dependent variable having values 1, 2, and 3
Let dummies be dy1 = 1 if y = 1, 0 if y = 2 or 3 /// dy2 = 1 if y =2, 0 if y = 1 or 3 /// dy3 = 1 if y = 3, 0 if y = 1 or 2
country_id indicates the identifier of countreis
Let x1 x2 x3 are the regressors

Code:
mlogit y x1 x2 x3 if 1<=country_id & country_id <=20
keep if country_id == 21
predict pr* , pr
summarize pr* d1 d2 d3, sep(3)
Then I got this result
Code:
  Variable   |        Obs        Mean      Std. Dev.      Min       Max
-------------+--------------------------------------------------------------------------
         pr1 |      5,808    .1188443    .0433502    .020736    .2662867
         pr2 |      5,808    .5227578    .0685922   .3551487    .7098638
         pr3 |      5,808    .3583978    .0774935   .2025161    .5655785
-------------+--------------------------------------------------------------------------
        dy1  |      5,808    .2458678    .4306376         0            1
        dy2  |      5,808    .6194904     .485554         0            1
        dy3  |      5,808    .1346419      .34137         0            1
As we can see, when I predict the choice probability using the 21th country data, the prediction is very poor.

In this case, can I say that the model specification is not appropriate fort the 21th country?

Thank you for your time spending to read this question.