| |
 |
|
|
Science Forum Index » Space - Consult Forum » Logistic Regression Q
Page 1 of 1
|
| Author |
Message |
| Patsfan |
 Posted: Mon Jan 15, 2007 12:22 pm |
|
|
|
Guest
|
When interpreting the results of a logistic regression, is it valid to
discuss/interpret statistically significant relationships between
individual predictor variables and the dependent variable if the
overall model is not significant? Many thanks in advance! |
|
|
| Back to top |
|
| Adam |
| Posted: Mon Jan 15, 2007 3:23 pm |
|
|
|
Guest
|
"Patsfan" <rmirabile@maguireassoc.com> wrote in message
news:1168878154.189539.304990@l53g2000cwa.googlegroups.com...
Quote: When interpreting the results of a logistic regression, is it valid to
discuss/interpret statistically significant relationships between
individual predictor variables and the dependent variable if the
overall model is not significant? Many thanks in advance!
There's no reason why you shouldn't discuss them, but I would be very wary
of attaching any importance to individual significant relationships under
those circumstances.
Do you think you might have too many predictor variables in your model? It
might be that by using a more appropriate model with fewer irrelevant
predictors the overall model would be significant and your inferences would
be more robust. However, you need to be careful that you are not just basing
your choice of predictor variables purely on which ones of the many you have
used give you significant results, or you run a serious risk of a type I
error.
HTH
Adam |
|
|
| Back to top |
|
| Patsfan |
| Posted: Mon Jan 15, 2007 6:12 pm |
|
|
|
Guest
|
Thanks for your helpful response! The interpretation is for a
dissertation write-up and the independent variables in the model were
included for theory testing purposes, and thus entered using the
"Enter" method (as opposed to stepwise, etc.). So it sounds like from
your response, I could conclude that the non-significant (i.e.,
irrelevant) predictors are decreasing the significance of the overall
model?
Adam wrote:
Quote: "Patsfan" <rmirabile@maguireassoc.com> wrote in message
news:1168878154.189539.304990@l53g2000cwa.googlegroups.com...
When interpreting the results of a logistic regression, is it valid to
discuss/interpret statistically significant relationships between
individual predictor variables and the dependent variable if the
overall model is not significant? Many thanks in advance!
There's no reason why you shouldn't discuss them, but I would be very wary
of attaching any importance to individual significant relationships under
those circumstances.
Do you think you might have too many predictor variables in your model? It
might be that by using a more appropriate model with fewer irrelevant
predictors the overall model would be significant and your inferences would
be more robust. However, you need to be careful that you are not just basing
your choice of predictor variables purely on which ones of the many you have
used give you significant results, or you run a serious risk of a type I
error.
HTH
Adam |
|
|
| Back to top |
|
| David Winsemius |
| Posted: Mon Jan 15, 2007 11:47 pm |
|
|
|
Guest
|
"Patsfan" <rmirabile@maguireassoc.com> wrote in
news:1168899140.872695.301590@a75g2000cwd.googlegroups.com:
Quote: Thanks for your helpful response! The interpretation is for a
dissertation write-up and the independent variables in the model were
included for theory testing purposes, and thus entered using the
"Enter" method (as opposed to stepwise, etc.). So it sounds like from
your response, I could conclude that the non-significant (i.e.,
irrelevant) predictors are decreasing the significance of the overall
model?
Adam wrote:
"Patsfan" <rmirabile@maguireassoc.com> wrote in message
news:1168878154.189539.304990@l53g2000cwa.googlegroups.com...
When interpreting the results of a logistic regression, is it valid
to discuss/interpret statistically significant relationships
between individual predictor variables and the dependent variable
if the overall model is not significant? Many thanks in advance!
There's no reason why you shouldn't discuss them, but I would be very
wary of attaching any importance to individual significant
relationships under those circumstances.
Do you think you might have too many predictor variables in your
model? It might be that by using a more appropriate model with fewer
irrelevant predictors the overall model would be significant and your
inferences would be more robust. However, you need to be careful that
you are not just basing your choice of predictor variables purely on
which ones of the many you have used give you significant results, or
you run a serious risk of a type I error.
I am unclear what you mean by the "model not being significant". If there
are control variables which theory or prior knowledge requires being in
the model, then they should form a base model. I would imagine that if
prio knowledge demanded their inclusion that you would have an
interesting time explaining why they did not improve the fit of the model
but <stuff> happens. Your main interest would then lay in comparing
the incremental impact of putting in variables of special interest on the
expanded model's fit. The decision about the contribution of those
variables will be driven by comparing nested model fits. The change in
deviance is distributed chi-square with the number of degrees of freedom
in the added variables.
Have you done any regression diagnostics at all? Any checking for
interactions? Leverage analysis? Any identification of influential data
points? Checked the fit of the model in deciles of predicted risk?
There are many important components of a proper data analysis that seem
to be missing. It is entirely too easy (and entirely too limited) to
force one set of variables into a regression model and then start
speculating about what a single model might mean. I have a hard time
imagining that only one model would be tested after all the effort of
gathering data for a thesis paper.
--
David winsemius |
|
|
| Back to top |
|
| Adam |
| Posted: Tue Jan 16, 2007 5:24 am |
|
|
|
Guest
|
"Patsfan" <rmirabile@maguireassoc.com> wrote in message
news:1168899140.872695.301590@a75g2000cwd.googlegroups.com...
Quote: Thanks for your helpful response! The interpretation is for a
dissertation write-up and the independent variables in the model were
included for theory testing purposes, and thus entered using the
"Enter" method (as opposed to stepwise, etc.). So it sounds like from
your response, I could conclude that the non-significant (i.e.,
irrelevant) predictors are decreasing the significance of the overall
model?
I think you need to be careful about concluding predictors are irrelevant
just because they are non-significant. As David says in his helpful
response, you need to consider many things when choosing your variables for
the model. Statistical significance is only one factor, and unlikely to be
the most important.
Adam |
|
|
| Back to top |
|
| Richard Ulrich |
| Posted: Wed Jan 17, 2007 12:25 am |
|
|
|
Guest
|
On 16 Jan 2007 09:22:55 -0800, "hume" <Winzar@gmail.com> wrote:
Quote: Significant parameter estimates in the presence of a non-significant
overall model is STRONG evidence of multicollinearity.
I would say that more generally, you have that backwards.
If the overall model is 'significant', multicollinearity can
reduce the p-value of the 'partial regression coefficients'.
The overall model can be *very* highly significant, without
any of the correlated predictors contributing much that is
unique.
If there are a lot of variables, especially if they are independent,
then you have a good chance of capitalizing on chance, so
that one or two coefficients will be nominally significant.
Quote: That is, your
significant predictor is a strong linear function of one or more of the
other predictors. Try removing some of your non-significant predictors
and I'll bet a dollar that your "significant" predictor is no longer so
significant.
If you look at variations of models, p-values will vary.
If they vary by a lot, you might have some multi-collinearity,
but a really strong single variable will produce an overall
significant model.
Quote:
Patsfan wrote:
When interpreting the results of a logistic regression, is it valid to
discuss/interpret statistically significant relationships between
individual predictor variables and the dependent variable if the
overall model is not significant? Many thanks in advance!
There were good comments posted about the role of
covariates and the attention that should be paid to setting
up one's hypotheses in the first place. Something that
remains intriguing is an "exploratory finding", to be regarded
lightly, if at all.
--
Rich Ulrich, wpilib@pitt.edu
http://www.pitt.edu/~wpilib/index.html |
|
|
| Back to top |
|
| |
|
Page 1 of 1
All times are GMT - 5 Hours
The time now is Mon Sep 08, 2008 11:55 am
|
|