Using ggplot2 to plot predicted probabilities
Manuel Spínola
I want to plot the the results (predicted probabilities) of a logistic regression model with 5 categorical predictors (factors). Using grid:expand I create the following object (see below) with the combinations (32) of each level of the factors. Each factor has 2 levels.
Is there a way to represent this graphically in ggplot2?
Thank you very much in advance.
Manuel
> grid
iluminacion animales cielo.raso piso paredes condicion ucl lcl
1 0 0 0 0 0 0.3955409 0.5282769 0.26280477
2 1 0 0 0 0 0.2881366 0.4219901 0.15428300
3 0 1 0 0 0 0.5093029 0.6329608 0.38564512
4 1 1 0 0 0 0.3909893 0.5308140 0.25116467
5 0 0 1 0 0 0.3629202 0.5072199 0.21862040
6 1 0 1 0 0 0.2605554 0.3967200 0.12439083
7 0 1 1 0 0 0.4746677 0.6159702 0.33336522
8 1 1 1 0 0 0.3585214 0.5063956 0.21064711
9 0 0 0 1 0 0.4289568 0.5914158 0.26649775
10 1 0 0 1 0 0.3172410 0.4807612 0.15372074
11 0 1 0 1 0 0.5436852 0.6945761 0.39279424
12 1 1 0 1 0 0.4242906 0.5939351 0.25464619
13 0 0 1 1 0 0.3953829 0.5740095 0.21675624
14 1 0 1 1 0 0.2880010 0.4571859 0.11881614
15 0 1 1 1 0 0.5091378 0.6828918 0.33538380
16 1 1 1 1 0 0.3908320 0.5734134 0.20825054
17 0 0 0 0 1 0.3651755 0.5230834 0.20726754
18 1 0 0 0 1 0.2624367 0.3946978 0.13017550
19 0 1 0 0 1 0.4770974 0.6421186 0.31207620
20 1 1 0 0 1 0.3607648 0.5107178 0.21081188
21 0 0 1 0 1 0.3336766 0.5141625 0.15319072
22 1 0 1 0 1 0.2364983 0.3825463 0.09045025
23 0 1 1 0 1 0.4426765 0.6381145 0.24723848
24 1 1 1 0 1 0.3294490 0.5012936 0.15760432
25 0 0 0 1 1 0.3977140 0.5106134 0.28481456
26 1 0 0 1 1 0.2900027 0.3903442 0.18966116
27 0 1 0 1 1 0.5115721 0.6202424 0.40290176
28 1 1 0 1 1 0.3931537 0.4962814 0.29002601
29 0 0 1 1 1 0.3650223 0.5212488 0.20879575
30 1 0 1 1 1 0.2623088 0.3938309 0.13078659
31 0 1 1 1 1 0.4769325 0.6378933 0.31597173
32 1 1 1 1 1 0.3606124 0.5075854 0.21363949
Manuel Spínola, Ph.D.
Instituto Internacional en Conservación y Manejo de Vida Silvestre
Universidad Nacional
Apartado 1350-3000
Heredia
COSTA RICA
mspi...@una.ac.cr
mspin...@gmail.com
Teléfono: (506) 2277-3598
Fax: (506) 2237-7036
Personal website: Lobito de río
Institutional website: ICOMVIS
Andrie de Vries
There is always a way, but first you have to describe in words what
you want the plot to look like.
Here is a starting point. First, I put your data into a data.frame,
and then plotted some of the variables using a combination of colour
and facets.
You can build on this by labelling your variables using descriptive
names (1 and 0 are good for modelling, but not descriptive for
graphing). You can build on this even more by combining some of your
predictors into super-categories. For example, you can combine 'piso'
and 'paredes' into a single variable 'piso.paredes' with four
different levels, and use this combined variable as a faceting
variable.
library(ggplot2)
library(reshape)
d <- structure(list(n = 1:32, iluminacion = c(0L, 1L, 0L, 1L, 0L,
1L, 0L, 1L, 0L, 1L, 0L, 1L, 0L, 1L, 0L, 1L, 0L, 1L, 0L, 1L, 0L,
1L, 0L, 1L, 0L, 1L, 0L, 1L, 0L, 1L, 0L, 1L), animales = c(0L,
0L, 1L, 1L, 0L, 0L, 1L, 1L, 0L, 0L, 1L, 1L, 0L, 0L, 1L, 1L, 0L,
0L, 1L, 1L, 0L, 0L, 1L, 1L, 0L, 0L, 1L, 1L, 0L, 0L, 1L, 1L),
cielo.raso = c(0L, 0L, 0L, 0L, 1L, 1L, 1L, 1L, 0L, 0L, 0L,
0L, 1L, 1L, 1L, 1L, 0L, 0L, 0L, 0L, 1L, 1L, 1L, 1L, 0L, 0L,
0L, 0L, 1L, 1L, 1L, 1L), piso = c(0L, 0L, 0L, 0L, 0L, 0L,
0L, 0L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 0L, 0L, 0L, 0L, 0L,
0L, 0L, 0L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L), paredes = c(0L,
0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L,
1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L,
1L), condicion = c(0.3955409, 0.2881366, 0.5093029, 0.3909893,
0.3629202, 0.2605554, 0.4746677, 0.3585214, 0.4289568, 0.317241,
0.5436852, 0.4242906, 0.3953829, 0.288001, 0.5091378, 0.390832,
0.3651755, 0.2624367, 0.4770974, 0.3607648, 0.3336766, 0.2364983,
0.4426765, 0.329449, 0.397714, 0.2900027, 0.5115721, 0.3931537,
0.3650223, 0.2623088, 0.4769325, 0.3606124), ucl = c(0.5282769,
0.4219901, 0.6329608, 0.530814, 0.5072199, 0.39672, 0.6159702,
0.5063956, 0.5914158, 0.4807612, 0.6945761, 0.5939351, 0.5740095,
0.4571859, 0.6828918, 0.5734134, 0.5230834, 0.3946978, 0.6421186,
0.5107178, 0.5141625, 0.3825463, 0.6381145, 0.5012936, 0.5106134,
0.3903442, 0.6202424, 0.4962814, 0.5212488, 0.3938309, 0.6378933,
0.5075854), lcl = c(0.26280477, 0.154283, 0.38564512, 0.25116467,
0.2186204, 0.12439083, 0.33336522, 0.21064711, 0.26649775,
0.15372074, 0.39279424, 0.25464619, 0.21675624, 0.11881614,
0.3353838, 0.20825054, 0.20726754, 0.1301755, 0.3120762,
0.21081188, 0.15319072, 0.09045025, 0.24723848, 0.15760432,
0.28481456, 0.18966116, 0.40290176, 0.29002601, 0.20879575,
0.13078659, 0.31597173, 0.21363949)), .Names = c("n",
"iluminacion",
"animales", "cielo.raso", "piso", "paredes", "condicion", "ucl",
"lcl"), class = "data.frame", row.names = c(NA, -32L)) #
d$iluminacion <- as.factor(d$iluminacion)
d$animales <- as.factor(d$animales)
d$cielo.raso <- as.factor(d$cielo.raso)
d$piso <- as.factor(d$piso)
d$paredes <- as.factor(d$paredes)
meltd <- melt(d, id.vars=1:6)
str(meltd)
p <- ggplot(meltd, aes(x=iluminacion, y=value, colour=variable)) +
facet_grid(animales~cielo.raso) +
geom_point()
print(p)
Andrie
> Instituto Internacional en Conservaci�n y Manejo de Vida Silvestre
> Apartado 1350-3000
> Heredia
> COSTA RICA
> mspinol...@gmail.com
> Tel�fono: (506) 2277-3598
> Fax: (506) 2237-7036
> Personal website: Lobito de r�o
> <https://sites.google.com/site/lobitoderio/>
> Institutional website: ICOMVIS <http://www.icomvis.una.ac.cr/>
Manuel Spínola
Something like you have done with facet will be very good. I was thinking also on a plot like the result from a cluster analysis or a classification tree where the end will be a point range display showing the point estimate and the 95% CI (the predicted probabilities).
Yes, I agree with you, I used number to facilitate the process but the 1 and 0 are basically presence - absence for each categorical predictor.
I am trying to plot the result of a logistic model when your predictors are categorical. I have plotting without any problem the result of a model when variables are continuous using stat_smooth, but I have trouble to express graphically the model for categorical predictors.
Best,
Manuel
Manuel Spínola, Ph.D.
Instituto Internacional en Conservación y Manejo de Vida Silvestre
Apartado 1350-3000
Heredia
COSTA RICA
mspin...@gmail.com
Teléfono: (506) 2277-3598
Fax: (506) 2237-7036
Personal website: Lobito de río
Institutional website: ICOMVIS
Rob James
You might want to look at Prof Harrell's rms package which is a remarkable collection of regression tools reflecting his philosophy, including many graphical displays.
It now (as of V3.0?) produces data frames from the Predict command and these df slide nicely into ggplot2.
Harrell and Hadley are clearly friends and while Harrell's current rms package does not yet make (extensive use) of ggplot, I suspect that plans are afoot.
I will say that have had less success (probably reflecting lack of effort) in converting Harrell's elegant bplot (heat-map) outputs into something equally elegant in ggplot2. But the fautl lies not in the software, but in the operator. For displaying categorical predictors, etc, you can start with the rms standard outputs and then use the dataframes which get generated by Predict to feed ggplot. That said, rms has a very particular and specific take on things like bootstrapping, calibration, and to use it well really requires engaging with Harrell's larger methodological and statistical perspective.
---------------------------
Robert C James, PhD
On Mon 11/10/10 05:33 , Manuel Spínola mspin...@gmail.com sent:
Institutional website: ICOMVIS
--
Manuel Spínola, Ph.D.
Instituto Internacional en Conservación y Manejo de Vida Silvestre
Universidad Nacional
Apartado 1350-3000
Heredia
COSTA RICA
mspi...@una.ac.cr
mspin...@gmail.com
Teléfono: (506) 2277-3598
Fax: (506) 2237-7036
Personal website: Lobito de río
Institutional website: ICOMVIS
--
You received this message because you are subscribed to the ggplot2 mailing list.
Please provide a reproducible example: http://gist.github.com/270442
To post: email ggp...@googlegroups.com
To unsubscribe: email ggplot2+u...@googlegroups.com
More options: http://groups.google.com/group/ggplot2
Manuel Spínola
Dennis Murphy
The problem is that with five factors each at two levels, there are 32 factor combinations. I suppose you could combine the levels of two factors twice to produce a 4 x 4 plot, and then make one plot with all 16 combinations at level 0 of the fifth factor and another at level 1 of the fifth factor. You could then use the 'paste two graphs on one page' trick using the ideas from section 8.4 of Hadley's book. I suspect that plot will be a bit cramped, but it's one approach to the problem...
HTH,
Dennis
Manuel Spínola
Yes, that is the problem, the 32 combinations. This was my best model according to AIC and I am trying to plot the results.
I will try your suggestion.
Best,
Manuel
James Howison
Fiss PC (2007) A Set-theoretic Approach to Organizational Configurations. Academy of Management Review. 32:1180-1198.
--J
On Oct 12, 2010, at 06:48, Manuel Spínola wrote:
> Thank you very much Dennis.
> Yes, that is the problem, the 32 combinations. This was my best model according to AIC and I am trying to plot the results.
> I will try your suggestion.
> Best,
>
> Manuel
>
>
> On 11/10/2010 02:17 p.m., Dennis Murphy wrote:
>> Hi:
>>
>> The problem is that with five factors each at two levels, there are 32 factor combinations. I suppose you could combine the levels of two factors twice to produce a 4 x 4 plot, and then make one plot with all 16 combinations at level 0 of the fifth factor and another at level 1 of the fifth factor. You could then use the 'paste two graphs on one page' trick using the ideas from section 8.4 of Hadley's book. I suspect that plot will be a bit cramped, but it's one approach to the problem...
>>
>> HTH,
>> Dennis
>>
>> On Mon, Oct 11, 2010 at 10:10 AM, Manuel Spínola <mspin...@gmail.com <mailto:mspin...@gmail.com>> wrote:
>>
>> Thank you very much Rob.
>>
>> I will take a look at the rms.
>> I am following the idea of Andrew Gelman of plotting the model.
>> instead of presenting a table with the coefficients, something
>> that I do not see very often, at least in ecology. Now, I found
>> that having several categorical predictors, and even some of them
>> with several levels, is challenging for showing the results, and
>> even more challenging to make the plots. I tried the effects
>> package from John Fox, but the plot for categorical predictor is
>> not what I would like for this type of variables.
>> Best,
>>
>> Manuel
>>
>> 2010/10/11 Rob James <r...@aetiologic.ca <mailto:r...@aetiologic.ca>>
>>
>> You might want to look at Prof Harrell's rms package which is
>> a remarkable collection of regression tools reflecting his
>> philosophy, including many graphical displays.
>>
>> It now (as of V3.0?) produces data frames from the Predict
>> command and these df slide nicely into ggplot2.
>>
>> Harrell and Hadley are clearly friends and while Harrell's
>> current rms package does not yet make (extensive use) of
>> ggplot, I suspect that plans are afoot.
>>
>> I will say that have had less success (probably reflecting
>> lack of effort) in converting Harrell's elegant bplot
>> (heat-map) outputs into something equally elegant in ggplot2.
>> But the fautl lies not in the software, but in the operator. For displaying categorical predictors, etc, you can start with
>> the rms standard outputs and then use the dataframes which get
>> generated by Predict to feed ggplot. That said, rms has a very
>> particular and specific take on things like bootstrapping,
>> calibration, and to use it well really requires engaging with
>> Harrell's larger methodological and statistical perspective.
>>
>>
>>
>> ---------------------------
>> Robert C James, PhD
>>
>> *On Mon 11/10/10 05:33 , Manuel Spínola mspin...@gmail.com
>> <mailto:mspin...@gmail.com> sent:
>> *
>> -- *Manuel Spínola, Ph.D.*
>> Instituto Internacional en Conservación y Manejo de Vida
>> Silvestre
>> Universidad Nacional
>> Apartado 1350-3000
>> Heredia
>> COSTA RICA
>> mspi...@una.ac.cr
>> mspin...@gmail.com
>> Teléfono: (506) 2277-3598
>> Fax: (506) 2237-7036
>> Personal website: Lobito de río
>> <https://sites.google.com/site/lobitoderio/>
>> Institutional website: ICOMVIS
>> <http://www.icomvis.una.ac.cr/>
>> -- You received this message because you are subscribed to
>> the ggplot2 mailing list.
>> Please provide a reproducible example:
>> http://gist.github.com/270442
>>
>> To post: email ggp...@googlegroups.com
>> <mailto:ggp...@googlegroups.com>
>> To unsubscribe: email ggplot2+u...@googlegroups.com
>> <mailto:ggplot2%2Bunsu...@googlegroups.com>
>> More options: http://groups.google.com/group/ggplot2
>>
>> -- You received this message because you are subscribed to the
>> ggplot2 mailing list.
>> Please provide a reproducible example:
>> http://gist.github.com/270442
>>
>> To post: email ggp...@googlegroups.com
>> <mailto:ggp...@googlegroups.com>
>> To unsubscribe: email ggplot2+u...@googlegroups.com
>> <mailto:ggplot2%2Bunsu...@googlegroups.com>
>> More options: http://groups.google.com/group/ggplot2
>>
>>
>> -- You received this message because you are subscribed to the
>> ggplot2 mailing list.
>> Please provide a reproducible example: http://gist.github.com/270442
>>
>> To post: email ggp...@googlegroups.com
>> <mailto:ggp...@googlegroups.com>
>> To unsubscribe: email ggplot2+u...@googlegroups.com
>> <mailto:ggplot2%2Bunsu...@googlegroups.com>
>> More options: http://groups.google.com/group/ggplot2
>>
>>
>
>
> --
> *Manuel Spínola, Ph.D.*
> Instituto Internacional en Conservación y Manejo de Vida Silvestre
> Universidad Nacional
> Apartado 1350-3000
> Heredia
> COSTA RICA
> mspi...@una.ac.cr
> mspin...@gmail.com
> Teléfono: (506) 2277-3598
> Fax: (506) 2237-7036
> Personal website: Lobito de río <https://sites.google.com/site/lobitoderio/>
> Institutional website: ICOMVIS <http://www.icomvis.una.ac.cr/>
>
Andrie de Vries
fact be simplified a little bit. The following code combines two sets
of factors to display four variables in facets, and uses the fifth as
a categorical variable in the plot itself.
It helps, of course, to understand that the data really is one
variable, with upper and lower confidence limits.
library(ggplot2)
lani <- paste("animales", 0:1)
lcie <- paste("cielo.raso", 0:1)
lpis <- paste("piso", 0:1)
lpar <- paste("paredes", 0:1)
d$iluminacion <- as.factor(lilu[d$iluminacion + 1])
d$animales <- as.factor(lani[d$animales + 1])
d$cielo.raso <- as.factor(lcie[d$cielo.raso +1])
d$piso <- as.factor(lpis[d$piso + 1])
d$paredes <- as.factor(lpar[d$paredes + 1])
d$factor1 <- paste(d$animales, d$cielo.raso)
d$factor2 <- paste(d$piso, d$paredes)
p <- ggplot(d, aes(x=iluminacion)) +
facet_grid(factor1~factor2) +
geom_point(aes(y=condicion)) +
geom_linerange(aes(ymin=lcl, ymax=ucl))
print(p)
---
Andrie
On Oct 12, 4:32 pm, James Howison <ja...@freelancepropaganda.com>
wrote:
>
> Fiss PC (2007) A Set-theoretic Approach to Organizational Configurations. Academy of Management Review. 32:1180-1198.
>
> --J
>
> On Oct 12, 2010, at 06:48, Manuel Spínola wrote:
>
> > Thank you very much Dennis.
> > Yes, that is the problem, the 32 combinations. This was my best model according to AIC and I am trying to plot the results.
> > I will try your suggestion.
> > Best,
>
> > Manuel
>
> > On 11/10/2010 02:17 p.m., Dennis Murphy wrote:
> >> Hi:
>
> >> The problem is that with five factors each at two levels, there are 32 factor combinations. I suppose you could combine the levels of two factors twice to produce a 4 x 4 plot, and then make one plot with all 16 combinations at level 0 of the fifth factor and another at level 1 of the fifth factor. You could then use the 'paste two graphs on one page' trick using the ideas from section 8.4 of Hadley's book. I suspect that plot will be a bit cramped, but it's one approach to the problem...
>
> >> HTH,
> >> Dennis
>
> >> <mailto:mspinol...@gmail.com> sent:
>
> read more »
Manuel Spínola
Best,
Manuel
Manuel Spínola, Ph.D.
Universidad Nacional
Apartado 1350-3000
Heredia
COSTA RICA
mspi...@una.ac.cr
mspin...@gmail.com
Teléfono: (506) 2277-3598
Fax: (506) 2237-7036
Personal website: Lobito de río