1
$\begingroup$

I have a set of data including information of age groups, education level, and meteorological data from 2000 to 2006.

The dataset was spilt by age groups(65-74,75-84,85+),education level(university or above, other), weather (summer, winter), and run the following regression:

Death ~ Temp_lag3 + NO3 + rainfall + relative_humidity, family = quasipoisson()

However, the coefficient of Temp_lag3 is insignificant.

If I run full model regression:

Death ~ Temp_lag3 + NO3 + rainfall + relative_humidity + age_group + education + weather, family = quasipoisson()

The coefficient of Temp_lag3 becomes significant. Also, the coefficients of age_group are significant.

How do we interpret such contradictory results?

Improve this question
$\endgroup$
4
  • 1
    $\begingroup$ If you use the "dredge" function of the MuMIn package, is Temp-lag3 supported in the models within a delta AIC of 2 ? $\endgroup$
    – CaroZ
    Commented Oct 24, 2023 at 14:35
  • $\begingroup$ Would you mind to explain why "dredge" function will relate to my question? The dataset used in subgroup analysis is different from full model analysis. Can this function deal with sub group analysis too...? Thank you. $\endgroup$
    – doraemon
    Commented Oct 25, 2023 at 1:39
  • $\begingroup$ Sorry I did not understand you had subsetted the data. It means that for the first model you provide, you ran one separate one on each age class ? $\endgroup$
    – CaroZ
    Commented Oct 25, 2023 at 8:44
  • $\begingroup$ Yes, I mean Death ~ Temp_lag3 + NO3 + rainfall + relative_humidity, family = quasipoisson(), subset = (age_group='65-74') etc. $\endgroup$
    – doraemon
    Commented Oct 25, 2023 at 8:53

1 Answer 1

Reset to default
3
$\begingroup$

By subsetting the data, you are decreasing the sample size available for the detection of your simple effects. When you consider your whole dataset together on the other hand, you increase the detection power of these effects. Note that it is incorrect to subset your dataset by age class if you want for example to know whether Temp_lag3 has a different effect in one age class compared to the other. For this you would need to test for the interaction between Temp_lag3 and age class in your whole dataset.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Would you mind to explain why it is incorrect to subset your dataset by age class if you want for example to know whether Temp_lag3 has a different effect in one age class compared to the other? The reason why I did sub-group analysis was the assumption of heterogeneity of heat/ cold effects in sub-groups. $\endgroup$
    – doraemon
    Commented Oct 26, 2023 at 1:44
  • $\begingroup$ I am thinking about whether it is releted to Simpson's paradox or not. I need more time to update my question and provide some sample data $\endgroup$
    – doraemon
    Commented Jun 19 at 16:16
  • 2
    $\begingroup$ If you want to know whether age Temp-lag3 has a different effect between age classes, it means you want to compare the effect of Temp.lag3 between the age classes. You are not doing this if you are analysing the age classes separately. If you want to compare the effect of peanut butter and chocolate on palatability of a cake, palatability of the sample cakes containing peanut butter and palatability of the sample cakes containing chocolate need to be in the same dataset. Otherwise you are not comparing them, you are describing them separately. $\endgroup$
    – CaroZ
    Commented Jul 2 at 14:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.