Chow test in logistics regression

#1
I carried out a logistic model on malnutrition status a whole underfive children and I found that my variable poverty is significant. But, when I rerun models for each sex of children, I noticed that the poverty variable is not significant in any of the two models. How can I explain this ?
 

obh

Active Member
#2
Hi Traore,

What is the dependent variable (DV)? and what are the independent variables (IV)?

If for example, you have 30 boys and 30 girls
Do you mean that when you run logistic regression over 30 boys or over 30 girls the IV poverty is not significant?
And when you run it over 60 children than IV poverty is significant?
 

obh

Active Member
#4
Hi,

You didn't answer all my questions :), so I will try to have a general answer.

1. Statistical tests with a greater sample size (60) will have greater power to identify smaller effect sizes than tests with a smaller sample size (30).
If you want to get better intuition you may do a simple exercise. Run a two-sample t-test on two sets of numbers with some difference in the average.
(for example group1: 2,3,2,3,2,3,2,3,2,3,2,3 and group2: 3,4,3,4,3,4,3,4,3,4,3,4 see what p-value you got.
Now run the same test on half of the groups (group1: 2,3,2,3,2,3 group2: 3,4,3,4,3,4)

2. Maybe If you have only boys or only girls you actually fix one IV, if this IV is relevant to the model the fixed model may not be as good as the wider model.

So the main question is why do you expect the poverty variable to be significant in any of the two models?
 
#5
The dependent variable is child stunting that is a reflect of living standard. This relationship is well established in the litterature. It would be logical to have this IV significant in at least in one model, if not the two models, Since it is not significant in any model, I am somehow obliged to formulate a recommendation to increase the sample size for the next round of the survey. To this end, it is important I have robust arguments to support this suggestion, with for example the calculation of effect size in logistic regression.
 

obh

Active Member
#6
Hi Traore,

If the theoretical model says it should be significant, and the full model with boys and girls is significant, so poverty is significant.
I expect that with a larger sample size the partial models will also be significant...

PS what are the p-values of poverty in each of the 3 models?
 
#7
Here is the information asked. Hope that it will do.

Both sexes:
Number of obs = 2569
-----------------Odds Ratio----Std. Err. -------T----------P>t
Poverty_1-------1.585762------0.3234038-----2.26------0.024
Poverty_2-------1.516249------0.298936------2.11-------0.035

Male:
Number of obs = 1317
----------------Odds Ratio----Std. Err. ------T------P>t
Poverty_1------1.430976-------0.3880431---1.32---0.187
Poverty_2------1.35826--------0.3575763----1.16---0.245

Female:
Number of obs = 1252
-----------------Odds Ratio-------Std. Err. -------t---------P>t
Poverty_1----1.967428--------0.7056009-------1.89-------0.060
Poverty_2----1.899081--------0.6526607-------1.87-------0.063

I have put spaces between items with hyphens to make it more readable.
 

obh

Active Member
#8
Hi Traore,

What is poverty_1, poverty_2? dummy variable of categorical variable with 3 values?
What are the possible values of poverty?

The only male regression and the only female regression have less power than both sexes regression.
Generally, a sample size of more than 1000 should be big enough to identify large effects size.

It looks like poverty is potentially significant for females, but may not be significant for males even with larger sample size.
So with larger sample size, at least for females poverty should be significant.
For me, there is no difference between the p-value of 0.05 or 0.06 (although you need to put the border)

The significance level is not important by itself, you should also look at the effect size (better a standardized one)