[Anshul Anand]

But do I have to look at margins, dydx? I understand that interpretation of sign can now be tricky. But can't I just look for my categorical independent variables at

 margins X, atmeans predict (outcome(#))

and for continuous variable wage_of_hour I could use

 margins, dydx ...

.

Because this looks fine for me, I don't need more analysis. Just a simple comparison of the workers. Because here I can say female occupation has the highest average probability for level of employment:

2. according to margins:
The average probability of being in a female occupation with a level of employment with 50-89% is 52.28%

The average probability of being in a male occupation with a level of employment with 50-89% is 11.16%

The average probability of being in a mix occupation with a level of employment with 50-89& is 36.56%

Sorry again for disturbing,

[Joshua D Merfeld]

It depends what you want. Using dydx will give you the marginal effect (that is, the effect on probability by increasing the independent variable by one). Omitting dydx gives you the predicted probability of being in a certain outcome, not the change in probability. The way you are talking about it, it sounds like you want to compare the two approaches, but you can't do that, because they have different meanings and interpretations.

Your interpretation:

2. according to margins:
The average probability of being in a female occupation with a level of employment with 50-89% is 52.28%
The average probability of being in a male occupation with a level of employment with 50-89% is 11.16%
The average probability of being in a mix occupation with a level of employment with 50-89& is 36.56%

is more or less correct, although I would change "average" to "predicted" and add "at the means of the other variables."

[Anshul Anand]

Thank you Joshua D Merfeld!!! Yes, I am only interested in of being in a certain outcome. Thank you very much to you all!!!

[Anshul Anand]

Ok, I need some help for the interaction term:

 mlogit occupation i.occupational_status i.flexible_working i.level_of_employment c.wage_per_hour#i.agegroup, base(mix_occupation)

the wage per hour is in range from 10 to 7682 (3 values are above 1000). I categorized age in four groups because these are the groups where wages changes due to age.

the output I get for the interaction is:

for female occupation:
agegroup#c.wage_per_hour  
  
20_29    -.0140473    .0017506    -8.02    0.000    -.0174784    -.0106163
30_39    -.0214578    .0011523    -18.62    0.000    -.0237163    -.0191993
40_49    -.0165352    .0010012    -16.52    0.000    -.0184975    -.0145729
50_64    -.0144538    .0009742    -14.84    0.000    -.0163632    -.0125445

for male occupation:
agegroup#c.wage_per_hour  

    20_29    -.0052381    .0014492    -3.61    0.000    -.0080785    -.0023977
    30_39    -.0041117    .0007272    -5.65    0.000    -.005537    -.0026865
    40_49    -.0001267    .0002574    -0.49    0.623    -.0006312    .0003778
    50_64    -.0015584    .0006172    -2.52    0.012    -.0027681    -.0003487

I am not sure how in this I should use margins atmeans predict (outcome(#)) because of this table with wage per hour and agegroup:

        agegroup
wage    29    39    49    64    Total
                    
10    5    8    7    2    22
11    5    9    12    10    36
12    7    9    8    9    33
13    5    5    9    13    32
14    4    8    13    26    51
15    6    11    12    12    41
16    13    13    13    13    52
17    11    21    27    18    77
18    27    18    33    39    117
19    33    37    38    30    138
20    37    60    44    60    201
21    63    61    57    56    237
22    75    91    69    58    293
23    97    89    86    89    361
24    96    98    111    105    410
25    116    101    119    130    466
26    127    130    113    111    481
27    147    122    123    133    525
28    153    153    143    142    591
29    140    150    156    160    606
30    143    165    157    144    609
31    154    149    167    166    636
32    166    171    148    166    651
33    150    199    187    179    715
34    121    180    164    164    629
35    112    160    171    185    628
36    100    190    203    172    665
37    72    195    206    158    631
38    89    174    182    181    626
39    86    194    213    161    654
40    69    171    191    188    619
41    65    196    183    189    633
42    73    184    191    176    624
43    45    180    163    153    541
44    49    161    186    160    556
45    32    156    155    159    502
46    42    147    164    145    498
47    28    122    150    143    443
48    25    133    141    138    437
49    28    143    138    126    435
50    14    117    140    121    392
51    25    117    116    103    361
52    15    103    140    125    383
53    14    110    119    118    361
54    15    99    121    116    351
55    8    90    82    96    276
56    7    82    98    96    283
57    7    81    92    70    250
58    10    79    81    87    257
59    5    59    81    65    210
60    7    63    94    82    246
61    7    74    80    80    241
62    4    68    79    67    218
63    2    58    94    84    238
64    4    55    71    59    189
65    3    53    58    61    175
66    3    45    64    72    184
67    3    35    71    60    169
68    2    37    64    65    168
69    0    29    51    44    124
70    2    36    62    52    152
71    2    36    55    34    127
72    2    28    51    59    140
73    2    26    38    37    103
74    1    28    53    48    130
75    0    16    38    43    97
76    1    27    37    35    100
77    0    22    47    43    112
78    0    18    48    33    99
79    0    17    30    36    83
80    0    11    25    31    67
81    1    21    24    27    73
82    2    22    31    22    77
83    0    18    25    24    67
84    2    13    30    21    66
85    0    13    18    31    62
86    1    11    28    14    54
87    0    16    24    20    60
88    1    11    19    25    56
89    0    8    24    11    43
90    0    11    19    30    60
91    2    10    20    13    45
92    1    7    15    17    40
93    0    14    18    15    47
94    0    6    12    11    29
95    0    6    19    19    44
96    0    8    20    12    40
97    0    8    18    14    40
98    1    5    15    15    36
99    0    4    7    10    21
100    0    5    11    19    35
101    0    3    13    7    23
102    0    5    12    12    29
103    0    11    13    13    37
104    1    5    8    8    22
105    0    6    10    4    20
106    0    5    9    10    24
107    0    3    5    6    14
108    0    3    10    11    24
109    0    2    4    5    11
110    0    6    5    6    17
111    0    5    7    10    22
112    0    2    6    6    14
113    0    4    4    5    13
114    0    4    2    5    11
115    1    0    1    4    6
116    0    1    6    7    14
117    0    4    7    4    15
118    0    1    4    5    10
119    1    4    7    11    23
120    0    0    5    1    6
121    0    0    5    3    8
122    1    1    5    6    13
123    0    1    2    5    8
124    1    3    3    3    10
125    0    0    3    2    5
126    0    2    5    1    8
127    0    1    6    5    12
128    0    1    0    1    2
129    0    2    5    4    11
130    0    2    6    5    13
131    0    1    1    0    2
132    0    1    1    3    5
133    0    0    0    5    5
134    0    1    2    1    4
135    0    2    3    2    7
136    0    1    3    2    6
137    0    0    1    2    3
138    0    0    5    3    8
139    0    0    2    1    3
140    0    1    1    1    3
141    0    2    2    2    6
142    0    0    2    0    2
143    0    0    0    1    1
144    0    0    3    2    5
145    0    0    1    1    2
146    0    0    2    4    6
147    0    0    3    0    3
148    0    0    3    1    4
149    0    3    3    1    7
150    0    0    2    0    2
151    0    0    0    1    1
152    0    0    0    1    1
153    0    0    0    2    2
154    0    1    2    0    3
155    0    1    4    2    7
156    0    1    1    3    5
157    0    0    1    1    2
158    0    0    3    2    5
159    0    1    2    2    5
160    0    1    0    1    2
161    0    0    1    1    2
162    1    2    2    3    8
163    0    1    1    0    2
164    0    1    1    4    6
165    0    0    0    1    1
166    0    0    3    2    5
167    0    0    2    0    2
168    0    0    2    0    2
169    0    0    2    1    3
170    0    0    0    2    2
171    0    1    1    3    5
172    0    0    2    2    4
173    0    0    3    3    6
174    0    1    1    0    2
175    0    1    2    0    3
176    0    0    1    1    2
178    0    2    1    1    4
179    1    0    0    1    2
180    0    1    2    0    3
181    0    0    1    0    1
182    0    1    0    1    2
183    0    0    0    2    2
184    0    1    1    0    2
186    0    1    1    0    2
187    0    0    1    0    1
188    0    0    1    0    1
189    0    0    1    1    2
190    0    1    2    1    4
191    0    0    0    1    1
192    0    0    0    2    2
193    0    0    1    0    1
194    0    0    0    2    2
195    0    0    0    2    2
196    0    2    0    1    3
197    0    0    1    0    1
198    0    0    1    0    1
199    0    0    0    1    1
203    0    0    3    0    3
204    0    1    0    1    2
207    0    2    0    1    3
209    0    1    0    1    2
210    0    0    0    1    1
212    0    0    1    0    1
213    0    0    1    0    1
215    0    0    1    0    1
216    0    0    1    1    2
219    0    2    1    1    4
224    1    0    1    0    2
225    1    0    0    0    1
227    0    0    1    0    1
232    0    0    0    4    4
233    0    0    1    0    1
237    0    0    0    1    1
240    0    0    0    1    1
241    0    0    0    1    1
249    0    0    0    1    1
254    0    1    0    0    1
258    0    1    0    0    1
260    0    0    0    2    2
263    0    1    0    0    1
268    0    0    1    0    1
271    0    0    1    0    1
276    0    0    0    1    1
295    0    0    0    1    1
302    0    0    0    1    1
305    0    0    0    1    1
308    0    0    0    1    1
309    0    0    1    1    2
314    0    0    0    1    1
319    0    0    0    1    1
320    0    0    1    0    1
331    0    1    0    0    1
343    0    1    0    0    1
350    0    0    0    1    1
354    0    1    0    0    1
362    0    0    1    0    1
363    0    0    0    1    1
364    0    0    0    1    1
371    0    1    0    0    1
377    0    0    1    0    1
400    0    0    0    1    1
407    0    1    0    0    1
410    0    0    0    1    1
411    0    1    0    0    1
413    0    1    0    0    1
417    0    0    0    1    1
433    0    1    0    0    1
460    1    0    0    0    1
508    0    1    0    0    1
657    0    0    1    0    1
676    0    0    0    1    1
703    0    0    0    1    1
885    0    0    0    1    1
889    0    0    1    0    1
903    0    0    0    1    1
1091    0    0    1    0    1
3144    0    0    0    1    1
7682    0    0    1    0    1
                    
Total    3,002    6,638    7,620    7,257    24,517

I think this is bad... however the dataset contains also a variable, where they categorize the wage per hour:

 
                                     20_29 30_39 40_49 50_64 Total

2.Bis 15 Fr./Std.            38       55        68      77      238
3.15 - 20 Fr./Std.          210    228       202    227      867
4.20 - 25 Fr./Std.          691    657       683    670   2,701
5.25 - 30 Fr./Std.         823     902       871    933   3,529
6.30 - 35 Fr./Std.         598  1,045    1,063  1,021   3,727
7.35 - 40 Fr./Std.         339  1,037    1,066    986    3,428
8.40 - 45 Fr./Std.        152     824       847    782    2,605
9.45 - 50 Fr./Std.          69    569       702     642    1,982
10.50 Fr./Std=<           82  1,321    2,118   1,919   5,440

Total                      3,002  6,638    7,620    7,257  24,517

Would be an interaction term with that categorized wage per hour and agegroup better?

[Joshua D Merfeld]

Given that you have used -fvvarlist-, the interaction term just represents the effect of wave within each age group. That is, the first coefficient represents the effect of wage for those aged 20 to 29 and so on.

Could you explain a little more what problem you think exists in the data? What do you think is "bad"?

[Clyde Schechter]

I think this is bad...

Why? It isn't at all surprising to find that a wage per hour variable has a skew distribution with a long tail and, in a finite data set, sparsity at individual values. And in no way does that impair your analysis. If you were treating wage_per_hour as a discrete variable, you would have a problem with lots of singleton dummies and collinearity with age group, but as a continuous variable, I just don't see a problem. Is there something I'm missing here?

As a general principle, converting a continuous variable into a discrete variable by applying cutpoints is a bad idea, often a really terrible idea. Unless there is reason to believe that something truly discontinuous happens to the outcome you are studying when the predictor variable crosses a cutpoint, you are just throwing away information. For example if you have two categories of 15-20 and 20-25, you are treating 20.01 and 19.99 as radically different wages, but you are also saying that 20.01 and 24.99 are essentially the same. There are only a few situations in which that isn't just nonsense.

Discrete variables should be used for things that are inherently discrete. Continuous variables should be used as continuous (with non-linear specifications through the use of transformations, splines, polynomial terms, or fractional polynomials if need be). Grouped values of continuous variables should be avoided unless that is the only form in which the data is available or when legal/privacy considerations require it.

See Royston P, Altman DG, Sauerbrei W. Dichotomizing continuous predictors in multiple regression: a bad idea. Statistics in Medicine 2006; 25:127-141 for a full explanation.

[Anshul Anand]

thanks for the answers. My intuition was that it is bad... So it would make more sense to do "c.wage_per_hour#c.age" where age ranges from 20 to 64. If I do that, I got:

for female occupation:
c.wage_per_hour#c.age  -.0002318   .0000151   -15.37   0.000    -.0002613   -.0002022

for male occupation:
c.wage_per_hour#c.age   -3.77e-06   5.56e-06    -0.68   0.498    -.0000147    7.13e-06

But I don't know how to interpret this with "margins, atmeans predict (outcome(#))" because its continuous. And also I would only be able to make an statement about female occupation and mix occupation (because the effect on male occupation is insignificant).

[Clyde Schechter]

To use -margins- with a continuous by continuous interaction you just need to select some interesting values of wage per hour and age and put them into an -at()- option. So, say some interesting values of wage_per_hour are 10, 30, 100, 300, 1000, 3000, and some interesting values of age are 20, 30, 40, 50, and 60. Then you can do this:

margins, at(wage_per_hour = (10 30 100 300 1000 3000) age = (20(10)60)) ///
    predict(outcome(1)) predict(outcome(2)) predict(outcome(3))

to get the predicted probabilities of all three outcomes at all of those combinations of wage_per_hour and age. Note that both wage_per_hour and age appear in the same -at()- option. If you put them in separate -at()-s, you will not get combinations of the two, but each listed separately--which, in an interaction model, is not very useful. Note also that if you are using Stata version 14, you do not have to specify those three -predict()- options because giving you all three outcome probabilities is now the default (thank you Richard Williams for pointing that out earlier in this thread).

And you may also want to look at -margins, dydx()- with similar at() specifications.

[Clyde Schechter]

And also I would only be able to make an statement about female occupation and mix occupation (because the effect on male occupation is insignificant).

No. Given the complex way in which probabilities of outcomes go up and down in multinomial models, the non-significance of the interaction coefficient in one outcome level does not mean that the probability of that outcome is unaffected by the interaction. It may well show considerable variation as a result of the effects of the interaction on the probabilities of the other outcome levels and the constraint that the sum of the three level probabilities must always be 1. So you still need to examine the actual predicted probabilities and marginal effects at each level of the outcome.

Apart from that very strong consideration in the case of -mlogit-, even in models that don't have these complications, if there is a strong theoretical reason to include an interaction in a model, it should stay there even if in your particular data you don't get a significant coefficient for it. If you included the interaction term just "on a hunch" and found it not significant then, by all means, you can remove it. (Caveat: don't make a fetish out of p < 0.05. p = 0.052 is not meaningfully different from p = 0.05 or p = 0.048.)

[Anshul Anand]

sorry for not giving any response. I was working on my thesis... Joshua D Merfeld, you said:

Given that you have used -fvvarlist-, the interaction term just represents the effect of wave within each age group. That is, the first coefficient represents the effect of wage for those aged 20 to 29 and so on.

. Does that mean, higher wage for the agegroup 20_29 makes it less liekly being in a female occupation than the reference category? Maybe this helps (where lohn=wage_per_hour and alter=agrgroup):

 margins, at(lohn=(10(5)100) alter=29) predict(outcome(1))
marginsplot

The graph of marginsplot:

And Clyde Schechter:

(Caveat: don't make a fetish out of p < 0.05. p = 0.052 is not meaningfully different from p = 0.05 or p = 0.048.)

The problem is, that we students have to follow the rules of the university ):

And Joshua D Merfeld, your tipp with changing the baseoutcome was really helpful. Because I did just now understand that: If the baseoutcome is mix occupation I interpret the coefficient relativ to the baseoutcome. changing the baseoutcome to men occupation gives me the opposite sign of mix occupation, meaining the interpretation remains as a whole the same. Now I can compare men occupation with women occupation. That is what I wanted . The only problem is to understand the interaction effect as above mentioned.

[Anshul Anand]

I mean the interaction effect as mentioned in #19.

Given that you have used -fvvarlist-, the interaction term just represents the effect of wave within each age group. That is, the first coefficient represents the effect of wage for those aged 20 to 29 and so on.

. Does that mean, that an increase in wage makes it less likely to be in a female occupation relativ to the mix occupation?

[Joshua D Merfeld]

Hi Anshul,

Yes, your interpretation that higher wages make one less likely to be in a female job for those that are 20-29 is correct. The -margins, dydx- command gives you the change in probability of being in the outcome of interest, though, not as a comparison to the base category, and the -margins- command without dydx specified gives you the (expected conditional) probability of being in the category of interest, again not as a comparison to the base category. The only coefficient that is with respect to the base category is the coefficient from the mlogit command itself, prior to doing any margins commands.

Also, I don't think any universities have rules regarding p-values (at least not that I've heard of). Clyde is 100% correct that you shouldn't make a big deal out of 0.05.

Josh

[Anshul Anand]

Hi Joshua D Merfeld! Many Thanks!!! Last questions. If I have as baseoutcome the mix occupation and the result is being in a female occupation is more likely relativ to the mix occupation if working part-time and now I change the baseoutcome to male occupation and the result is being in a female occupation is more likely relativ to the male occupation if working part-time. Can I say after that: Summing together, being in a female occupation is more likely?

And I am not sure whether I should do three regression : one for female, one for male and one with all. My hypotheses are: Being in a female occupation is more likely if working part-time ; Being in a female occupaation is more likely if having flexible working time; Being in a male occupation is more likely if having high wages; Being in a male occupation is more likely if having high position (with comparison interpretation).

[Benedikt Schrotberger]

I did not want to open a new topic because i´m basicly having the same problem. I´m not sure how to interpret my results of the margins after a logit regression.

I used the code:

mlogit wechsel $var2
margins, dydx(*) post noestimcheck

I used post noestimcheck because of the size of the dataset it takes hours to calculate the margins and I had to do alot modells so I wanted to have results.

This is the output I recieve with outreg2:

 


(1)
(2)
(3)
(4)
(5)

VARIABLES
LABELS
0b.gender
1.gender
age
0b.staat
1.staat









1bno._predict

0.000


0.000




(0.000)


(0.000)


2o._predict

0.000


0.000




(0.000)


(0.000)


3o._predict

0.000


0.000




(0.000)


(0.000)


4o._predict

0.000


0.000




(0.000)


(0.000)


5o._predict

0.000


0.000




(0.000)


(0.000)


1bn._predict


0.001
0.002***

0.034***




(0.002)
(0.000)

(0.003)

2._predict


0.010***
-0.001***

-0.016***




(0.001)
(0.000)

(0.002)

3._predict


-0.007***
-0.004***

-0.016***




(0.002)
(0.000)

(0.002)

4._predict


-0.036***
0.001***

-0.005***




(0.001)
(0.000)

(0.002)

5._predict


0.033***
0.001***

0.003




(0.001)
(0.000)

(0.002)

The Variable wechsel only has 5 posible outcomes so I have marginal effects for each of the outcomes. But how do I interpret them now ?
For example gender:
Is the probabiility of being in second class of wechsel 1 % higher for woman than for man ?

If this is the interpretation then what is the advantage of multinominal logit compared to a logit regression where i recode wechsel ?

[Benedikt Schrotberger]

Even a small hint where to look to find the correct interpretation method would be very nice. Or would it be more use full to open a new topic ?