A perfect fit means no error at all. All 1 responses are predicted to be exactly 1, and all 0 responses exactly 0. When it occurs, there will be no unique set of parameter values that will produce the perfect fit.

David Nichols Principal Support Statistician and Manager of Statistical Support SPSS Inc.

> -----Original Message----- > From: KB [SMTP:cg815@FREENET.TORONTO.ON.CA] > Sent: Friday, March 12, 1999 3:40 PM > To: SPSSX-L@UGA.CC.UGA.EDU > Subject: Re: ordinal logistic regression > > Hi John, > > I read your explanations in the news group about the logistic question and > I > have a question for you : > > I am having some troubles to understand what is happening with a > particular > dataset. > Let me explain it to you and if you can give me the answer it would help > me > in my researches. > > I am working on a project involving different classification and > optimization techniques. > In order to demonstrate a theory we are working on, my supervisor and I, > we > are using some dataset published in the literature. > Therefore, we agreed to use a simple dataset in the banking industry to > work > on the logistic regression feature and extract some results. > This dataset has 5 independent variables (x1, ..., x5) and a dependent > variable Y which can take the values of 0 and 1 (0 for bankrupt banks and > 1 > for solvent ones). This dataset is based on a sample of 66 observations. > > Here is for your info the meaning of each variable: > x1 = (working capital) / (total assets) > x2 = (retained earnings) / (total assets) > x3 = (earnings before interest and taxes) / (total assets) > x4 = (market-value equity) / (book value of total liabilities) > x5 = sales / (total assets) > > y = 0 if bankrupt after 2 years (33 firms) > y = 1 if solvent after 2 years (33 firms) > > The paper says: > "A multiple logistic regression model was fitted using variables x2 and > x3. > The other thre variables did not substantially add to the explanatory > power > of the model. The fitte logistic model is: > > g(X) = ln( P/(1-P)) = -0.550 + 0.157x2 + 0.194x3 > > the predicted probabilities Phat(X) for the remaining solvent is given by > : > > Phat(x) = exp(g(x)) / [1 + exp(g(x))] > > .... if we classify any firm with Phat(X) greater than 0.5 to be a solvent > firm, the model given above misclassifies only two firms, one from the > bankrupt and one from the solvent category...." > > Well, I used the same hypothesis and tried the logistic regression with > SPSS. Here is the options I chose : > > Input file : bkfirm.sav (the attached file) > Statistics ---> Regression ---> Logistic > Dependent variable : Y > Covariates : X1, ...., X5 > > Options : Classification plots, Correlation of estimates, Iteration > history, > Hsmer-Lemeshow goodness-of-fit, 99 iterations. > > First, I tried it with the two variables X2 and X3 and I found exactly the > same results. That was great ! > > Then I tried to do it with the five variables and SPSS prompted me the > following message (cf. > bkfirm2.spo) : > "Estimation terminated at iteration number 25 because > a perfect fit is detected. This solution is not unique." > > at the end of the output it couldn't compute the logistic parameters and > prompted the following message instead: > ">Warning # 18582 > >Covariance matrix cannot be computed. Remaining statistics will be > >omitted." > > > May be you have an idea about what's happening and how I can handle that > kind of problem. > > My ultimate goal is to run many logistic regressions on the same dataset > by > moving firms from 0 to 1 and vice versa. > About 100 simulations would be a first step and it will help to validate > some predicting theory based on other techniques such as goal programming > and DEA. > > Thanks for your help. > > J.Hendrickx a icrit dans le message ... > >In article <7c7i7t$dli@bobs.unbc.ca>, zumbob@unbc.edu says... > >> Can I run an ordinal logistic regression in SPSS? If so, how? I don't > need > >> too many options; a logit link function will do. I am running the most > >> recent version of SPSS. > >> > >SPSS version 9 can run multinomial logistic regression but apparently not > >ordered logistic regression (I don't have version 9 yet, it always takes > >half a year longer here before a new release is available, so I can't say > >for certain). I have a macro that uses MATRIX to estimate an ordered > >logistic model; it will run on any version of SPSS. Get OLOGIT2.INC at > >http://baserv.uci.kun.nl/~johnh/mlogist/. > > > >SPSS is offering a separate program GOLDMINER for ordered logistic > >regression. See http://www.spss.com/software/goldminer/ for more > >information. There's a demo available and a paper with a description of > >the model under the "White Papers" tab at this site. > > > >However, goldminer is not a proportional odds cumulative logistic > >regression model. It's referred to in the white paper mentioned above as > >a "parallel log-odds" model. What it does is impose a linear restriction > >on a multinomial logistic model. The model has 3 intercept parameters for > >a dependent variable with 4 categories, but each independent variable has > >only one parameter. This indicates the effect of a unit's increase of the > >independent on the logit for a category of the dependent versus the > >previous category, which is equal for all adjacent categories. > > > >I was disappointed by this model, since it can already be estimated using > >COXREG.COXREG can be used to estimate a multinomial logistic or > >conditional logit model, where the latter is actually a more flexible > >form of the former. See > >http://baserv.uci.kun.nl/~johnh/mlogist/mlogist.html#The person/choice > >for details. In a conditional logit specification of the model, it is > >fairly simple to impose different response functions, i.e. the type of > >restriction imposed on the dependent variable. > > > >The following program will reproduce the results in > >http://www.spss.com/software/goldminer/biomedical.htm, except that the > >effects have opposite signs due to a different order of the dependent > >variable. Not super user-friendly, I admit, but very flexible. The point > >is that this type of model should have been an option in the new > >multinomial logistic program. > > > >Goldminer would be more worthwhile if it could estimate a proportional > >odds model as well as this parallel odds model. An even more worthwhile > >improvement would be to implement a Stereotyped Ordered Regression (SOR) > >model. A SOR model estimates a scale for the dependent variable based on > >the effects of the independent variables. So it's basically a parametric > >version of CATREG in SPSS categories. See Hendrickx & Ganzeboom, > >"Occupational status attainment in the Netherlands", European > >Sociological Review 14: 387-403 for an application of this model. > > > >Good luck, > >John Hendrickx > > > > > >SPSS syntax for estimating a parallel odds model. > >------------------------------------------------------------------------ > >data list free /mental ses events. > >var labels mental 'Mental Impairment' > > /ses 'socioeconomic status' > > /events 'Life Events'. > > > >value labels mental 1 'Well' 2 'Mild' 3 'Moderate' 4 'Impaired' > > /ses 0 'low' 1 'high'. > > > >* Agresti 1990: 325. > >begin data. > >1 1 1 > >1 1 9 > >1 1 4 > >1 1 3 > >1 0 2 > >1 1 0 > >1 0 1 > >1 1 3 > >1 1 3 > >1 1 7 > >1 0 1 > >1 0 2 > >2 1 5 > >2 0 6 > >2 1 3 > >2 0 1 > >2 1 8 > >2 1 2 > >2 0 5 > >2 1 5 > >2 1 9 > >2 0 3 > >2 1 3 > >2 1 1 > >3 0 0 > >3 1 4 > >3 0 3 > >3 0 9 > >3 1 6 > >3 0 4 > >3 0 3 > >4 1 8 > >4 1 2 > >4 1 7 > >4 0 5 > >4 0 4 > >4 0 4 > >4 1 8 > >4 0 8 > >4 0 9 > >end data. > > > >crosstabs events ses by mental /statistics=chisq. > > > >* Transform the data to a person-choice file for estimation as a > >conditional > >* logit model. > >* Each respondent has 4 records, one for each category of mental > >impairment. > >* The dependent variable K has the value 1 for the record corresponding > >with > >* the respondent's score on mental and 2 for the other records. > >* A stratifying variable RESP indicates respondents, a copy of K is used > >to > >* indicate the censored/not censored status of records in COXREG. > > > >compute resp=$casenum. > >loop Y=1 to 4. > >xsave outfile='%temp%\temp.sav'. > >end loop. > >execute. > > > >get file='%temp%\temp.sav'. > > > >compute K=2. > >if (mental=Y) K=1. > >compute cens=K. > >compute mental=Y. > >compute mentlin=mental. > > > >* Estimate the conditional logit model using coxreg. > >* The first model is an intercept model, the second uses linear logits > >* as the response function for events and ses. > >* The L^2 and df are the same as Goldminer, the effects of events and ses > >have > >* the same values but opposite signs. > > > >coxreg K > > /status=cens(1) > > /strata=resp > > /contrast(mental)=indicator(1) > > /method=enter mental > > /method=enter mental mentlin*events mentlin*ses. > > << File: bkrpfirm.sav >> << File: bkfrim2.spo >>