A question about OR

#1
Hello everybody
I have a question about interpretation of OR in logistic regression.

My result is like this:
Dependent variable is satisfaction with family life, explanatory variable is MMSE (cognitive functions, where 0 score means great difficulties)
OR (95% CI ) - 1.088 (1.088 - 1.523), P value is .003, ROC .896.

Can I say that: satisfaction with family life can be explained by good cognitive functions in 88% of the cases? or it is 8% of cases?

Thank u very much for the answer.
 

maartenbuis

TS Contributor
#2
Are you sure that the confidence interval is 1.088 - 1.523, while the point estimate is 1.088 (i.e. the lower bound of the confidence interval)?

Anyhow, lets assume the odds ratio is 1.088. An odds ratio is a ratio of odds (for once a name in statistics makes actual sense!). So a unit change in MMSE is associated with a 8.8% increase in the odds of being satisfied with your family live.
 

hlsmith

Omega Contributor
#3
Can you refresh us on how the mini mental status exam is scored and what its range is, to make sure you are treating it correctly in the model and model interpretation.
 
#4
maartenbuis - ush, seems i have written it wrong, there should be "2". It is OD (95% CI ) - 1.288 (1.088 - 1.523).

hlsmith - MMSE is scored between 0 -30, cut-off score > 24p.
 
#6
maartenbuis - I did not dichotomize. It is just norm value for MMSE as it is described in the test description. Do u think i should have dichotomized?
 

maartenbuis

TS Contributor
#7
No, it usually a bad idea to dichotomize. I am still confused about what you mean with "cut-off score". An odds ratio of 1.288 is extremely high (as in unrealistic and probably wrong) for a variable that ranges from 0 to 30. Can you give us the output when you type

Code:
desc MMSE
sum MMSE, detail
 
#8
maartenbuis - with cut off score they mean that a person with > 24p has good cognitive abilities. And person under - 24 p has not so good connive abilities.
 

maartenbuis

TS Contributor
#9
I am still not sure what MMSEsum is. Can you give us descriptive statistics (mean, standard deviation, range, the 5th, 10th 25th 50th 75th 90th 95th percentiles) of that variable.

To be concrete, I am afraid that MMSEsum is dichotomized without you realizing it.

Technically your output could also be consistent with a MMSEsum that runs from 0 to 30. No-one is predicted to be satisfied with family live (classification table), i.e. the predicted probability is always below 50%, which means that the log(odds) is always negative. If MMSEsum runs from 0-30 than the maximum log(odds) is -7.598 + 30*.253 = -.008.

However, I am still very very very worried about the effect size of 1.288 for a variable that runs from 0-30. That sounds unrealistically large: it would mean that the odds would increase by (exp(30*.253)-1)*100%=197,731% if you compare the lowest MMSEsum with the highest MMSEsum.
 

maartenbuis

TS Contributor
#10
People don't seem to be sattisfied with their family life. That sounds untypical to me. Is there something special about your sample (e.g. these are couples in therapy)?
 
#11
maartenbuis - MMSE has different parts like language, orientation in time, orientation in space, spatial abilities etc. Each part has its own score. We summarize the scores from different parts and it is total, sum score what matters. I have just coded it MMSEsum, so i know that it is a total score and not a part of it.

It is not dochotomized :)
for my study group the mean is 26,5, median 28, 95 % CI - 26-27, SD - 4.1, range 0-30, variance 16.8

p.s. about people not satisfied with family life - it is a special group of sick people.
 
Last edited:

maartenbuis

TS Contributor
#12
Is the range you report the theoretical range or the empricial range? The latter would be very surprising given the mean and standard deviation you report.

The effect is still very large but a bit more realistic: A standard deviation increase in MMSE is associated with a (exp(.253*4.1)-1)*100 = 182% increase in the odds of being sattisfied with family life.
 
#13
I am afraid, i do not know what is the difference between the theoretical range or the empricial range I will read it after writing this post.

What i have written, was based on my study group.

U are so good as such calculations. I wish i could learn such things too. I can only use SPSS, nothing more.
 

maartenbuis

TS Contributor
#14
A theoretical range is what is theoretically possible. An empirical range is what is the largest and smalles value for the variable MMSEsum in the data.
 

hlsmith

Omega Contributor
#15
You had someone with a mmse score of 0, is what maartrenbuis is asking. I am also curious about this. Was the mmse conducted as part of normal care or for the study. Typically patient participation in a study is based on a mmse, patients with a low score are deemed unable to consent to participate in a study - lack the capacity to make a reasonable judgment for themselves.

maartenbuis, the mmse is a very very short quiz to access mental capacity, I was just trying to get at whether it was appropriate to treat as a continuous variable in the model. And as badibad mentioned - your familiarity with odds ratio is very impressive!

Edit: appropriateness of MMSE when interpreting an OR may be up for consideration, e.g., a one unit increase in score related to a ??% increase in odds of satisfaction..., etc.
 

hlsmith

Omega Contributor
#16
A standard deviation increase in MMSE is associated with a (exp(.253*4.1)-1)*100 = 182% increase in the odds of being sattisfied with family life.
maartenbuis, I get your calculation here, but how and when would such a calculation be useful? Can you give an example of how the calculation may be used to examine one's data or sample?
 

hlsmith

Omega Contributor
#18
Seems as though you are using Logistic regression, will any other variables be in the model with mmse and satisfaction?

Side question, I guess this opens the door to how would you quantified satisfaction with family life if these individuals have a null mental status. How is your dependent variable collected?
 

maartenbuis

TS Contributor
#19
maartenbuis, I get your calculation here, but how and when would such a calculation be useful? Can you give an example of how the calculation may be used to examine one's data or sample?
What I did with my computation was to get the odds ratio when MMSE was standardize to have a standard deviation of 1. If I had access to the data I would probably just standardize the variables rather than doing it like I did in my posts. So the answer to your question is: I would perform such calculations if I did not have (quick) access to the data.

Sometimes, I do stuff like that when I first see the results of a model and try out different units of the explanatory variable to get a quick feel for the size of the effect. If I find the results of these preliminary computations useful, I would probably repeat it by creating new explanatory variables with the desired unit, as I am less likely to make an error that way...