Hi everyone,
I’m analyzing the association between cholesterol and echocardiograph measures using imputed data, and looking at BMI as a potential effect modifier. Given the multiple readings of cholesterol over a 2hr period I am using the area under the curve for this particular analysis. My dataset has 38 observations.
I ran regress including only the interaction term (c.chol_total#i.bmi_grp) for each outcome and separately ran regress using the interaction term between chol_total and BMI while also including the main effects (c.chol_total##i.bmi_grp).
My results differ greatly between the 2 models and some of the outcomes reported opposite directionality (BMI group 3). I then ran regress stratified by BMI and the results seem most consistent with the interaction + main effects model, however, the stratified analysis is heavily limited by sample size.
My questions are; what is driving this difference between the simple interaction and interaction + main effects model to the point of completely changing the directionality of the coefficient? I know the stratified model is the least "robust", but between the two interaction models I've shown, which would be preferable to report?
The reasoning behind my question is that the stratified model and the interaction + main effects model report negative coefficients for BMI group 3, which is rather unexpected and not observed in the simple interaction model. However, if the interaction + main effects model is suitable, this finding could be important - I just want to know if it is in fact a valuable finding worth expanding upon or just a result of poor sample size or analytical error on my part. Would appreciate some insight
Here is a snippet of my data;
I’m analyzing the association between cholesterol and echocardiograph measures using imputed data, and looking at BMI as a potential effect modifier. Given the multiple readings of cholesterol over a 2hr period I am using the area under the curve for this particular analysis. My dataset has 38 observations.
I ran regress including only the interaction term (c.chol_total#i.bmi_grp) for each outcome and separately ran regress using the interaction term between chol_total and BMI while also including the main effects (c.chol_total##i.bmi_grp).
Code:
mi estimate: regress outcome3 c.chol_total##i.bmi_grp
Multiple-imputation estimates Imputations = 20
Linear regression Number of obs = 38
Average RVI = 0.0000
Largest FMI = 0.0000
Complete DF = 32
DF adjustment: Small sample DF: min = 30.17
avg = 30.17
max = 30.17
Model F test: Equal FMI F( 5, 30.2) = 2.94
Within VCE type: OLS Prob > F = 0.0280
--------------------------------------------------------------------------------------
outcome3 | Coefficient Std. err. t P>|t| [95% conf. interval]
---------------------+----------------------------------------------------------------
chol_total | -.0007062 .0010682 -0.66 0.514 -.0028871 .0014748
|
bmi_grp |
2 | -.3883076 .2920682 -1.33 0.194 -.9846484 .2080333
3 | 1.254069 .7681965 1.63 0.113 -.314424 2.822562
|
bmi_grp#c.chol_total |
2 | .0038234 .0021201 1.80 0.081 -.0005054 .0081521
3 | -.0062009 .0059295 -1.05 0.304 -.0183076 .0059058
|
_cons | 2.736967 .1502497 18.22 0.000 2.43019 3.043745
--------------------------------------------------------------------------------------
Code:
mi estimate: regress outcome3 c.chol_total#i.bmi_grp
Multiple-imputation estimates Imputations = 20
Linear regression Number of obs = 38
Average RVI = 0.0000
Largest FMI = 0.0000
Complete DF = 34
DF adjustment: Small sample DF: min = 32.16
avg = 32.16
max = 32.16
Model F test: Equal FMI F( 3, 32.2) = 3.01
Within VCE type: OLS Prob > F = 0.0446
--------------------------------------------------------------------------------------
outcome3 | Coefficient Std. err. t P>|t| [95% conf. interval]
---------------------+----------------------------------------------------------------
bmi_grp#c.chol_total |
1 | -.0003244 .0009957 -0.33 0.747 -.0023521 .0017034
2 | .0008706 .0011002 0.79 0.435 -.00137 .0031112
3 | .0031674 .0014029 2.26 0.031 .0003103 .0060245
|
_cons | 2.672764 .1323394 20.20 0.000 2.403251 2.942277
--------------------------------------------------------------------------------------
My questions are; what is driving this difference between the simple interaction and interaction + main effects model to the point of completely changing the directionality of the coefficient? I know the stratified model is the least "robust", but between the two interaction models I've shown, which would be preferable to report?
The reasoning behind my question is that the stratified model and the interaction + main effects model report negative coefficients for BMI group 3, which is rather unexpected and not observed in the simple interaction model. However, if the interaction + main effects model is suitable, this finding could be important - I just want to know if it is in fact a valuable finding worth expanding upon or just a result of poor sample size or analytical error on my part. Would appreciate some insight
Here is a snippet of my data;
Code:
clear
input double(outcome1 outcome2 outcome3) float(chol_total bmi_grp)
8.5 .2602 2.7842 64.21331 1
12.2 .41125 2.95675 75.131195 2
6.25 .1944 2.8997 59.8591 1
10.4 .3802 3.2707 124.33892 2
. . . 72.491135 2
13.2 .31085 2.04485 77.10976 2
9.4 .27925 2.69575 30.807266 1
16 .5775 3.03285 121.6245 3
9.55 .27085 2.5666 49.1961 1
6.55 .2011 2.8676 89.26665 2
8.066666667 .1903 2.165266667 26.76571 1
14.55 .42735 2.5077 122.41145 2
10.9 .3829 3.13 151.64449 2
10.13333333 .322033333 2.865766667 153.1595 3
15.9 .56775 3.0026 152.31947 3
12.45 .48135 3.3811 121.64539 1
12.25 .42085 3.00315 171.43484 2
12.45 .36635 2.5658 60.3415 1
7.15 .20685 2.696 108.66035 2
11 .3171 2.56075 164.97496 1
1.1 .00253 2.276 110.54124 1
6.3 .1944 2.8713 180.49765 2
7.85 .1954 2.2981 122.66583 2
9.85 .32135 2.93915 57.20695 1
6.566666667 .193266667 2.766766667 110.4878 1
14.8 .57565 3.3089 92.86025 3
14.5 .5129 3.01655 141.99864 2
9.25 .3116 3.0463 175.73875 1
7.55 .2381 2.9118 188.0964 2
13.65 .33575 2.1284 228.8729 1
11.75 .3458 2.5976 227.68706 1
12.35 .4547 3.22585 115.00507 3
11.65 .40255 3.05625 218.1458 2
12.5 .34765 2.43475 276.64615 1
4.7 .1385 2.812633333 89.67432 2
4.8 .12395 2.4547 64.75948 2
6.95 .1602 2.1468 148.5786 2
10.46666667 .352233333 3.011233333 141.43736 3
13.25 .51 3.3253 116.11664 3
10.9 .3829 3.13 151.64449 2
1.1 .00253 2.276 110.54124 1
end
Comment