Logistic regression problem: propensity score matching

Dear Paul,
Thanks for your reply.

I am using logistic regression because my response variable is categorical 
- and this seems to be recommended in the literature (by Heckman, Smith 
and others).
I think that Prof. Ripley's point here is that although one can use multnom 
in the nnet package to fit a binary (or binomial) logistic regression, it 
is more common to do so using the glm (generlized linear model) function. 
One normally would use multinomial logistic regression only for a 
polytomous (several-category) response variable. Applied to a dichotomous 
response, it will give the same results as a binary logistic regression.
. . .

I have MASS but was unable to locate logistic regression, which I was 
advised was the standard method for my problem.
In MASS (4th edition), logit models are discussed in chapter 7 on 
generalized linear models (see, in particular, section 7.2). In my R and 
S-PLUS Companion, to which you referred in your original message, these 
models are discussed in chapter 5 on generalized linear models (see, in 
particular, section 5.2.1).

I hope that this helps,
  John
Thanks again.

Prof Brian Ripley wrote:

1) Why are you using multinom when this is not a multinomial logistic 
regression?  You could just use a binomial glm.

2) The second argument to predict() is `newdata'.  `sample' is an R 
function, so what did you mean to have there?  I think the predictions 
should be a named vector if `sample' is a data frame.

3) There are many more examples of such things (and more explanation) in 
Venables & Ripley's MASS (the book).

On Wed, 4 Jun 2003, Paul Bivand wrote:

I am doing one part of an evaluation of a mandatory welfare-to-work 
programme in the UK.
As with all evaluations, the problem is to determine what would have 
happened if the initiative had not taken place.
In our case, we have a number of pilot areas and no possibility of 
random assignment.
Therefore we have been given control areas.
My problem is to select for survey individuals in the control areas who 
match as closely as possible the randomly selected sample of action area 
participants.
As I understand the methodology, the procedure is to run a logistic 
regression to determine the odds of a case being in the sample, across 
both action and control areas, and then choose for control sample the 
control area individual whose odds of being in the sample are closest to 
an actual sample member.

So far, I have following the multinomial logistic regression example in 
Fox's Companion to Applied Regression.
Firstly, I would like to know if the predict() is producing odds ratios 
(or probabilities) for being in the sample, which is what I am aiming for.
You asked for `probs', so you got probabilities.

Secondly, how do I get rownames (my unique identifier) into the output 
from predict() - my input may be faulty somehow and the wrong rownames 
being picked up - as I need to export back to database to sort and match 
in names, addresses and phone numbers for my selected samples.

My code is as follows:
londonpsm <- sqlFetch(channel, "London_NW_london_pilots_elig", 
rownames=ORCID)
attach(londonpsm)
mod.multinom <- multinom(sample ~ AGE + DISABLED + GENDER + ETHCODE + 
NDYPTOT + NDLTUTOT + LOPTYPE)
lonoutput <- predict(mod.multinom, sample, type='probs')
london2 <- data.frame(lonoutput)

The Logistic regression seems to work, although summary() says the it is 
not a matrix.

what is `it'?

The output looks like odds ratios, but I would like to know whether this 
is so.

No.

______________________________________________
R-help at stat.math.ethz.ch mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help
-----------------------------------------------------
John Fox
Department of Sociology
McMaster University
Hamilton, Ontario, Canada L8S 4M4
email: jfox at mcmaster.ca
phone: 905-525-9140x23604
web: www.socsci.mcmaster.ca/jfox

Logistic regression problem: propensity score matching

Thread (5 messages)