Dear all, I have a question regarding an observational study I analyzed a few years ago and reanalyzed recently.
I thank you in advance for your time.
I want to investigate whether there are differences in efficacy between men and women in a cohort of patients. There is no specific treatment; patients simply attend my outpatient clinic. The outcome I'm examining is the post-pre variation in Neuropathic pain, measured as a continuous outcome (lower values are better).
The first analysis I performed was a multivariable model in which the effect of the "sex" variable was adjusted through manual backward selection, and the model was selected based on AUC and Akaike Criteria.
In my final model (see below), I observed that men had a coefficient of +0.37, indicating they went worse than women (as a more negative delta is considered better).
The margins at means are provided below.
For the second analysis, I constructed the best model to predict "sex" and used it to calculate a Propensity score for "being male." The propensity score-adjusted sex effect was 0.385, which was nearly the same as the multivariable regression model.
Yesterday, I applied the same propensity score model using the command "teffects psmatch," and I obtained an ATE of -0.389.
So, while the magnitude is the same, the sign is inverse. I'm now wondering if this analysis is yielding an opposite conclusion compared to the two previous analyses, or if the result is essentially the same, and the negative sign is appropriate.
Where am I getting confused?
Gianfranco
I thank you in advance for your time.
I want to investigate whether there are differences in efficacy between men and women in a cohort of patients. There is no specific treatment; patients simply attend my outpatient clinic. The outcome I'm examining is the post-pre variation in Neuropathic pain, measured as a continuous outcome (lower values are better).
The first analysis I performed was a multivariable model in which the effect of the "sex" variable was adjusted through manual backward selection, and the model was selected based on AUC and Akaike Criteria.
In my final model (see below), I observed that men had a coefficient of +0.37, indicating they went worse than women (as a more negative delta is considered better).
Code:
---------------------------------------------------------------------------------
| Robust
deltaNRS | Coefficient std. err. t P>|t| [95% conf. interval]
----------------+----------------------------------------------------------------
1.sex | .374365 .3611812 1.04 0.301 -.3387997 1.08753
DN4 | .281132 .1225688 2.29 0.023 .0391157 .5231484
CARDIOVASCOLARI | .4990604 .3565306 1.40 0.163 -.2049216 1.203042
PSICHIATRICHE | .8165899 .4646979 1.76 0.081 -.1009722 1.734152
BUPRENORFINA | 1.531804 .6156641 2.49 0.014 .3161538 2.747454
CODEINA | .6894478 .3850612 1.79 0.075 -.0708688 1.449764
TRAMADOLO | .9433767 .4287724 2.20 0.029 .0967508 1.790003
Ozono | -1.141873 .3753492 -3.04 0.003 -1.883012 -.4007327
NRSentrata | -.3709376 .1254566 -2.96 0.004 -.618656 -.1232193
_cons | -2.015001 1.160065 -1.74 0.084 -4.305588 .2755868
---------------------------------------------------------------------------------
Code:
. margins sex, vce(unconditional) atmeans
Adjusted predictions Number of obs = 174
------------------------------------------------------------------------------
| Unconditional
| Margin std. err. t P>|t| [95% conf. interval]
-------------+----------------------------------------------------------------
sex |
0 | -3.425054 .2091705 -16.37 0.000 -3.838068 -3.012039
1 | -3.050689 .287412 -10.61 0.000 -3.618194 -2.483184
------------------------------------------------------------------------------
Yesterday, I applied the same propensity score model using the command "teffects psmatch," and I obtained an ATE of -0.389.
Code:
. teffects psmatch ( deltaNRS ) (sex Fumo i.Diagnosi NEUROLOGICHE REUMATOLOGICHE ORTOPEDICHE UROLOGICHE BUPRENORFINA OSSICODONE Altrotrattamento, lo
> git) if ID!=154 & ID!=176
Treatment-effects estimation Number of obs = 174
Estimator : propensity-score matching Matches: requested = 1
Outcome model : matching min = 1
Treatment model: logit max = 8
------------------------------------------------------------------------------
| AI robust
deltaNRS | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
ATE |
sex |
(1 vs 0) | -.3886426 .9014637 -0.43 0.666 -2.155479 1.378194
------------------------------------------------------------------------------
Where am I getting confused?
Gianfranco
