[Carlo Lazzaro]

Sophie:
your results do not show evidence of a squared relationship for -age-.
That said, only the time dimension of your panel seems to play a role in explaining variations in the regressand.
That said, what strikes me is that you have -age- only as other predictors.
Are you sure that you gave a fair and true view of the data generating process?

[sophie maene]

Hi, yes sorry here it is!

 xtoprobit shealth c.age##c.age i.wave, vce(cluster mergeid_n)

Fitting comparison model:

Iteration 0:   log likelihood = -220787.09  
Iteration 1:   log likelihood = -216727.71  
Iteration 2:   log likelihood = -216726.89  
Iteration 3:   log likelihood = -216726.89  

Refining starting values:

Grid node 0:   log likelihood = -209010.45

Fitting full model:

Iteration 0:   log pseudolikelihood = -209010.45  
Iteration 1:   log pseudolikelihood = -200663.47  
Iteration 2:   log pseudolikelihood =  -196345.3  
Iteration 3:   log pseudolikelihood = -195566.04  
Iteration 4:   log pseudolikelihood = -195532.62  
Iteration 5:   log pseudolikelihood = -195532.47  
Iteration 6:   log pseudolikelihood = -195532.47  

Random-effects ordered probit regression        Number of obs     =    154,711
Group variable: mergeid_n                       Number of groups  =     62,399

Random effects u_i ~ Gaussian                   Obs per group:
                                                              min =          1
                                                              avg =        2.5
                                                              max =          7

Integration method: mvaghermite                 Integration pts.  =         12

                                                Wald chi2(8)      =    6777.25
Log pseudolikelihood  = -195532.47              Prob > chi2       =     0.0000

                         (Std. Err. adjusted for 62,399 clusters in mergeid_n)
------------------------------------------------------------------------------
             |               Robust
     shealth |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         age |  -.0634849   .0137425    -4.62   0.000    -.0904197   -.0365502
             |
 c.age#c.age |   .0000222   .0000901     0.25   0.805    -.0001543    .0001987
             |
        wave |
          2  |  -.1946302    .014936   -13.03   0.000    -.2239043   -.1653562
          3  |  -.3675597   .0161868   -22.71   0.000    -.3992852   -.3358342
          4  |   -.243044   .0154814   -15.70   0.000     -.273387    -.212701
          5  |  -.1574564   .0153565   -10.25   0.000    -.1875545   -.1273583
          6  |  -.1856671   .0154187   -12.04   0.000    -.2158873    -.155447
          7  |  -.2027917   .0157322   -12.89   0.000    -.2336263   -.1719571
-------------+----------------------------------------------------------------
       /cut1 |  -6.501648   .5206987                     -7.522199   -5.481098
       /cut2 |  -4.913978   .5207371                     -5.934604   -3.893352
       /cut3 |  -3.246265   .5207259                     -4.266869   -2.225661
       /cut4 |  -2.076403   .5207411                     -3.097037    -1.05577
-------------+----------------------------------------------------------------
   /sigma2_u |   1.575051   .0200396                       1.53626    1.614822
------------------------------------------------------------------------------

[Carlo Lazzaro]

Sophie:
why not sharing -xtoprobit- outcome table, too? Thanks.

[sophie maene]

Dear,

I estimated a random effects ordered probit model:

 xtoprobit shealth c.age##c.age i.wave, vce(cluster mergeid_n)

To see whether the turning point was included in my data i ran as was suggested in this thread.

nlcom -_b[age]/(2*_b[c.age#c.age])

The data in my sample runs from 65-95 but I receive

_nl_1:  -_b[age]/(2*_b[c.age#c.age])

------------------------------------------------------------------------------
     shealth |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       _nl_1 |   1427.445   5472.045     0.26   0.794    -9297.566    12152.46
------------------------------------------------------------------------------

The age 1427.445 simply does not exists. Any idea what is going on?

When I run:

 margins, at (age=(65(5)95)) nose
marginsplot

I get this margins plot:



Not sure of this margins plot proves that I should use age squared or not.

Thank you in advance!

[Aye Aye Khaine]

Thank you so much, Professor Clyde!
Much appreciated!

[Clyde Schechter]

You are reading the graph accurately, but you are misinterpreting its meaning. You cannot, from this graph, say anything about the height for age z-score of any one child, nor can you make statements such as "all children are above -2 z score." The predicted values that are plotted in that graph are the mean height-for-age z-scores at that age, adjusted for other variables in your model. So the average height for age is always above -2, but there may be many children whose height for age dips below that value. And you can't say anything about any single child from this graph. It is a graph of the average values (adjusted) only.

[Aye Aye Khaine]

Hi Clyde,
Please let me try the following in the graph and please correct me if I am not making sense of the graph. the dependent variable was height for age z score (standard deviation)unit. Children <-2 z score are considered stunted.

The graph means that

at about 2 months of age, his or her height for age z score (or their height growth according to that age) is growing between .-5 (fish) and .-6 height for age z score. as a child reaches to 6 months old, their height for age z score between -8 and -9. Older children in my sample average height for age z score is low in general. However, all children are above -2 z score.


Is that how I should be making sense?

Thanks so much for your previous reply.

Hope to hear again soon. Please help and I got to get over this.

[Clyde Schechter]

In #16 you show results of the -utest- command, and the pvalue is 6.40e-16. In ordinary scientific notation that's 6.4x10^-16. So why do you say p > .10? You know, I'm one of those people who doesn't like the concept of statistical significance at all, and am often dismissive of p-values, but even I would be loathe to go against a pvalue of 6.4x10^-16. Even though it is absurd to pretend that p-values have anything approaching that level of accuracy, you don't get a pvalue that small without something going on that's worth talking about.

Even ignoring p-values, the U-shapedness of the graph is unmistakeable. Even if the p-value really were > .10, I would call this a definitely U-shaped relationship.

[Aye Aye Khaine]

Hello, here is the graph on the regression I posted with age and age squared. U shape test's p value was way above p>.10. so, not sure I need to do age squared??? please advise and hope to hear from you soon.
Best regards,
Aye Aye

[Aye Aye Khaine]

Hi,
How do I get the graph to show? I just did the simple cut and paste and it is not showing.

[Clyde Schechter]

No graph is showing in #18.

[Aye Aye Khaine]

Hello,

I did the margins plot on the regression model I presented earlier with the squared of age. My age ranges are from zero to sixty months. I got the following graph. I don't think it is actually U shaped. It is just a liner declining function as children ages and their height are not growing. Am I right? Is the graph below showing?

margins, at (agemo= (2 (4) 65))
marginsplot

[Aye Aye Khaine]

Hi Clyde,
I am sorry. Please disregard my #12.
Warm regards,
Aye Aye Khaine

[Aye Aye Khaine]

Hello,
I did the U test. How shall I interpret? P value is insignificant. Does that mean that there is no shape? my model is linear without the squared term?



. utest agemo sq_age

Specification: f(x)=x^2
Extreme point: 42.57625

Test:
H1: U shape
vs. H0: Monotone or Inverse U shape

-------------------------------------------------
| Lower bound Upper bound
-----------------+-------------------------------
Interval | .00274 64.56667
Slope | -.0673668 .0347968
t-value | -15.02906 8.016549
P>|t| | 1.65e-50 6.40e-16
-------------------------------------------------

Overall test of presence of a U shape:
t-value = 8.02
P>|t| = 6.40e-16




Thanks so much

[Aye Aye Khaine]

Hi Clyde,
Thank you. I keep making such basic mistake.
Thanks for catching that!