Re: Averaging probabilities?
- From: rocketD <darahx@xxxxxxxxx>
- Date: Thu, 20 Mar 2008 10:11:53 -0700 (PDT)
On Mar 19, 5:26 pm, Bruce Weaver <bwea...@xxxxxxxxxxxx> wrote:
On Mar 19, 2:57 pm, rocketD <dar...@xxxxxxxxx> wrote:
Hi all.
<b>Background:</b> I'm working on a study looking at neonatal
hypoglycemia as a binary outcome, using infants who have neonatal
hypoglycemia (cases) matched for comparison with infants who don't
(controls). I've run a univariate model using conditional logistic
regression in Stata 9, and for each observation, have predicted the
conditional probability of developing neonatal hypoglycemia given
gestational age. The probability value output by Stata is called
P(Neonatal Hypoglycemia | Single outcome w/in group)
and I want to plot this against gestational age to show that there is
a steep negative slope up to a certain gestational age, after which
point it levels out. (Translation: risk of hypoglycemia decreases as
gestational age increases, to a point, then becomes stable.) I'd like
to do this cleanly, with one summary probability, like an average,
[that is, one y-value] for each gestational age (x-value), rather than
the 30 or so I have for each category now.
--- snip ---
Does this mean that you treated age as a continuous variable in your
model, but now you want to use age categories in the plot? If age was
treated as a continuous variable, why not treat it that way in your
plot too. I.e., make a scatter-plot with X = age (selected values),
and Y = the fitted probability (or log-odds) of the outcome? If your
model treated age as continuous (and did not contain age-squared or
any other higher order terms), then such a plot with Y = log-odds of
the outcome should give you a straight line. With Y = the fitted
probability of the outcome, it will be curvilinear, as you said.
--
Bruce Weaver
bwea...@xxxxxxxxxxxxxxxxxxxxxxxxxxxxx/wv/bwhomedir
"When all else fails, RTFM."- Hide quoted text -
- Show quoted text -
Hi Bruce and Paul,
Thanks for your replies. I agree on the boxplots, but wanted a
concise graph for publication to illustrate one of our points
visually. We chose our sample from the entire population of babies
born at a certain hospital that fit very specific criteria, and babies
of this nature are generally considered similar anywhere. Of course
that generalization is an assumption that we'll state in the paper.
Regarding the nature of the variable, I used gestational age both as a
continuous and as indicator. Technically, it is continuous, but it is
measured in weeks and for our population, there are only 6 possible
values (37-42). Indicator variables show similar results in the
predicted conditional probabilities, but there is much less power for
each of the individual levels when treating the variable as an
indicator.
The plot Bruce describes is exactly what I've done, but I can't use
logodds because they are not output from a conditional logistic
regression - instead, the program will only predict the probability of
outcome given one case within a group (that means that for each group
of case & controls, the probabilities on the individuals will sum to
1.0). I *might* be able to get the logodds in SAS but haven't the
time needed to figure it out. Anyway, I calculated the average
probability for all observations falling within each age (37, 38, 39,
etc) and overlaid it on the graph, and it does come out the way I
expected, with the curve evident at least through the points.
Thanks for jogging my memory, though, I had to really think about each
post which helped me understand the data and the mechanics of the
analysis a lot better.
Dara
.
- References:
- Averaging probabilities?
- From: rocketD
- Re: Averaging probabilities?
- From: Bruce Weaver
- Averaging probabilities?
- Prev by Date: Re: Acceptance Rejection Method
- Next by Date: Re: Too much reliance of computers?
- Previous by thread: Re: Averaging probabilities?
- Next by thread: Too much reliance of computers?
- Index(es):
Relevant Pages
|
