# Lec2LogisticRegression2 [modalità compatibilità] ```18/06/2014
LOGISTIC REGRESSION 2
dr. Ferruccio Biolcati-Rinaldi
Applied Multivariate Analysis - Module 2
Graduate School in Social and Political Sciences
University of Milan - Academic Year 2013-2014
How to fit, interpret and evaluate
a logistic regression model?
1
18/06/2014
1. Fit and interpretation
Logit coefficients
• b coefficients
• Hard to understand
2
18/06/2014
Odds ratios: exp(b)
(Field 2005, 225-226)
•
•
•
Crucial to the interpretation of logistic regression is the value of exp(b),
which is an indicator of the change in odds resulting from a unit change in
the predictor. As such, it is similar to the b-coefficient in logistic regression
but easier to understand (because it doesn’t require a logarithmic
transformation).
Ratio of odds = odds after a unit change in the predictor / original odds
This proportionate change in odds is exp(b), so we can interpret exp(b) in
terms of the change in odds:
– If the value is greater than 1 then it indicates that as the predictor increases,
the odds of the outcome occurring increase;
– conversely, a value less than 1 indicates that as the predictor increases, the
odds of the outcome occurring decrease.
•
Transforming for a better interpretation: Δ% = [Exp(β)-1]*100
– E.g.: exp(β)=1.2 Δ%=20%; exp(β)=0.8 Δ%=-20%;
exp(β)=2.4 Δ%=140%; exp(β)=0.2 Δ%=-80%;
Odds ra@o ≠ Risk ra@o
(Menard 2002, 57)
• A risk ratio is a ratio of two probabilities
• The two are &laquo;approximately&raquo; equal under
certain fairly restrictive conditions(a base rate
less than.10)
• In general, the use of an odds ratio to
“represent” a risk ratio will overstate the
strength of the relationship. E.g.:
– odds ratio of 4.5 for females
– risk ratio of 2.8 for females (.487/.173)
3
18/06/2014
Odds ratio and probability:
do not over-emphasise odds ratio
• Probability
– Change in church attendance from 0.39 to 0.38
– Δabsolute = -0.01 Δrelative = -2.5%
• Odds ratio
– From 0.639 (=0.39/0.61) to 0.613 (=0.38/0.62)
– Odds ratio = 0.959 (=0.613/0.639)
– Δrelative = -4,1%
Probability interpretation
Sorry for some slides in Italian…
4
18/06/2014
“Pensi al suo rapporto con la religione. Quale, fra le
seguenti definizioni, descrive meglio le Sue credenze
religiose attuali?” (%) (CGP 2001, 231-248)
Credo in Ges&ugrave; Cristo e negli insegnamenti della Chiesa cattolica
54,3
Credo in Ges&ugrave; Cristo ma solo in parte negli insegnamenti della Chiesa
cattolica
28,6
Non mi sento membro di nessuna religione in particolare, ma credo
nell’esistenza di una Divinit&agrave; o Essere superiore o Energia spirituale
o Dimensione soprannaturale
9,6
Sono membro di una religione o culto o setta o nuovo movimento
religioso o confessione diversa da quella cattolica
1,3
Sono assolutamente ateo/a, cio&egrave; credo che non esista nessuna
Divinit&agrave; o Essere superiore o Energia spirituale o Dimensione
soprannaturale
5,1
Non sa/non risponde
1,1
Totale
100,0
(N)
(2.171)
Regressione logistica binomiale
Un esempio
(CGP 2001, 231-248)
• Credere senza appartenere* = α
+ β1 et&agrave; (deviazioni dalla media, 45,6)
+ β2 femmina
+ β3 Italia Centrale
+ β4 Italia Meridionale
• Credere senza appartenere * = - 1,818
- 0,023 et&agrave; (deviazioni dalla media, 45,6)
- 0,527 femmina
+ 0,147 Italia Centrale
– 0,467 Italia Meridionale
5
18/06/2014
Et&agrave;
X1
X2
X3
X4
Et&agrave;
Femmina
Centro
Sud
logit
P(Y= 1)
P(1 - P)
30
-15,6
0
0
0
-1,46
0,19
60
14,4
0
0
0
-2,15
0,10
0,15
0,09
30
-15,6
0
1
0
-1,31
0,21
0,17
60
14,4
0
1
0
-2,00
0,12
0,10
30
-15,6
0
0
1
-1,93
0,13
0,11
60
14,4
0
0
1
-2,62
0,07
0,06
30
-15,6
1
0
0
-1,99
0,12
0,11
60
14,4
1
0
0
-2,68
0,06
0,06
30
-15,6
1
1
0
-1,84
0,14
0,12
60
14,4
1
1
0
-2,53
0,07
0,07
30
-15,6
1
0
1
-2,45
0,08
0,07
60
14,4
1
0
1
-3,14
0,04
0,04
20
-25,6
0
0
0
-1,23
0,23
0,18
30
-15,6
0
0
0
-1,46
0,19
0,15
40
-5,6
0
0
0
-1,69
0,16
0,13
50
4,4
0
0
0
-1,92
0,13
0,11
60
14,4
0
0
0
-2,15
0,10
0,09
70
24,4
0
0
0
-2,38
0,08
0,08
Logit: Z = -1,818 -0,023*X1 -0,527*X2 +0,147*X3 -0,467*X4
Nota: l'et&agrave; &egrave; espressa in scarti dalla media (45,6)
Fonte: adattamento da Corbetta, Gasperoni e Pisati (2001, 231-248)
P(Y=1)=exp(logit)/(1+exp(logit))
Conditional-effects plot
(CGP 2001, 231-248)
P(Y=1) Nord
P(Y=1) Centro
P(Y=1) Sud
0,25
0,20
0,15
0,10
0,05
0,00
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70
Et&agrave;
6
18/06/2014
Binomial logistic regression models
– Logit(p^i) = α + β1Xi1 + β2Xi2
• Multiplicative form (non linear effect):
– p^i = Pr(Yi = 1|Xi) =
exp(α + β1Xi1 + β2Xi2) / 1 + exp(α + β1Xi1 + β2Xi2)
Average marginal effects
(Kohler and Kreuter 2012, 361-362)
• There is not an obvious way to express the
influence of an independent variable on the
dependent variable with one single number
• A marginal effect is the slope of a regression
line at a specific point
• Because the marginal effects differ with the
levels of the covariates, it makes sense to
calculate the average of all marginal effects of
the covariate patterns observed in the
dataset. This is the average marginal effect
7
18/06/2014
2. Model fit evaluation
R2, F, and sum of squared errors
(Menard 2002, 17-24)
•
•
•
•
•
•
•
•
•
Total Sum of Squares (SST) (Mean of Y)
Error Sum of Squares (SSE) (Predicted value of Y)
Regression Sum of Squares (SSR) = SST – SSE
The multivariate F ratio is used to test whether the
improvement in the prediction using Y Regression instead of Y
Mean could be attributed to random sampling variation.
Two equivalent hypothesis
F=
R2, coefficient of determination, explained variance
Proportional reduction in error statistic, form 0 to 1
R2 = SSR/SST
8
18/06/2014
Likelihood
(Menard 2002, 17-24)
• log LL as SS
• log LL, negative values, to maximize
• - 2 log LL, posi@ve values from 0 to ∞, to
minimize
– D0 as SST (Mean of Y)
– DM as SSE (Predicted value of Y)
– GM as SSR
• R2McF = 1 – (ln L(MFull) / ln L(MIntercept))
• If model MIntercept = MFull, R2McF equals 0, but
R2McF can never exactly equal 1.
• Pseudo R2 in Stata
9
18/06/2014
Likelihood ratio statistic
(Aldrich and Nelson 1984, 54-61)
• c = -2log(L0/L1) = (-2logL0) – (-2logL1) = -2(logL0 – logL1) [F]
– where L1 is the value of the likelihood function for the full model as
fitted [SSE]
– and L0 is the maximum value for the likelihood function if all
coefficients except the intercept are 0 [SST].
• That is, the computed chi-square value tests the hypothesis that all
coefficients except the intercept are 0, which is exactly the
hypothesis that is tested in regression using the “overall” F statistic.
• The degrees of freedom for this chi-square statistic is K-1 (i.e., the
number of coefficients constrained to be zero in the null
hypothesis).
• The formal test, then, is performed by comparing the computed
statistic c to a critical value (χ2(K-1, α)) taken from a table of chisquare distribution with K-1 degrees of freedom and significance
level α.
Joint hypothesis tests for subsets
of coefficients (Aldrich and Nelson 1984, 54-61)
•
•
•
•
•
•
•
•
We have defined procedures for testing hypotheses about individual coefficients
(t-test and confidence interval) and about set of all coefficients except the
intercept. In many circumstances, hypotheses of interest center on the importance
or significance of more than one coefficient, but not all of them.
Religion example…
A procedure that can be used to test joint hypotheses is a variant on the likelihood
ratio test reported above for the overall fit of the model.
c = -2log(L2/L1)
where L1 is – as above – the value of the likelihood function for the full model as
fitted
and L2 is the likelihood value that results from constraining some coefficients to be
zero, leaving the other coefficients unconstrained.
c will follow a chi-square distribution with degrees of freedom equal to the
number of constraints imposed.
L2 can be no larger than L1, and will be equal to L1 only if the coefficients of
interest have coefficients of zero in the unconstrained model. The smaller L2 (and
hence the larger c statistic) the less likely it is that the constrained coefficients are
in fact zero.
10
18/06/2014
Credente senza appartenenza piuttosto che cattolico
(soglia = 0,11; pp = 62,3%; λ = -257,2%)
Valori osservati di Y
Totale
Valori previsti di Y
1 (s&igrave;)
2 (no)
1 (s&igrave;)
121
656
777
2 (no)
87
1.144
1.231
Totale
208
1.800
2.008
Pass by score (soglia = 0,50; pp = 60,0%; λ = 0,0%)
Valori osservati di Y
Totale
Valori previsti di Y
1 (s&igrave;)
2 (no)
1 (s&igrave;)
12
12
2 (no)
8
18
24
26
Totale
20
30
50
Pass by score and exp (soglia = 0,50; pp = 72,0%; λ = 36,4%)
Valori osservati di Y
Totale
Valori previsti di Y
1 (s&igrave;)
2 (no)
1 (s&igrave;)
16
8
24
2 (no)
6
20
26
Totale
22
28
50
11
18/06/2014
actual VS. predicted values
(CGP 2001, 244-248)
Valori previsti di Y
Valori osservati di Y
Totale
1 (s&igrave;)
0 (no)
1 (s&igrave;)
Veri positivi (VP)
Falsi positivi (FP)
R1
0 (no)
Falsi negativi (FN) Veri negativi (VN)
R0
Totale
Poterepredittivo =
C1
VP + VN
*100
N
C0
λ=
N
VP + VN − max(C1, C 0)
*100
N − max(C1, C 0)
3. Other topics
12
18/06/2014
Zero cell
• Zero cell count occurs when the dependent variable is invariant for
one or more values of a categorical independent variable
• E.g.: other ethnicity and marijuana
• odds = 1/(1-1) = 1/0 = + infinity, ln(odds) = + infinity
• odds = 0/(1-0) = 0/1 = 0, ln(odds) = - infinity
• Very high estimated standard error
• Particularly nominal values
• 3 options
– Accepting …
– Recoding the categorical independent variable in a meaningful way
– Adding a constant to each cell of the contingency table
Complete separation
• If you are too successful in predicting the
dependent variable with a set of predictors
• Coefficients and standard errors extremely
large, pseudo R2=1
• Quasi-complete separation, bivariate
relationship
• Nothing intrinsically wrong, but be suspicious:
problems in the data or in the analysis
13
18/06/2014
Rare events
• http://www.statisticalhorizons.com/logisticregression-for-rare-events