ORDINAL
RESPONSE
MODELS
Nguyen Quang
[email protected]
ORDINAL DISCRETE
VARIABLE
• Many discrete outcomes have natural ordering
• self-reported health status (Likert scale)
• the degree to which patients adhere to physician’s
prescription (Likert scale)
• the degree to which someone agree with a statement
[totally disagree/disagree /neutral /agree/totally agree]
• What if these are our dependent variable?
• OLS: the variable has no quantitative meaning
• MNL: appropriate for non-ordinal discrete variable
• ORDERED PROBIT/LOGISTIC MODEL
AN EXAMPLE
Dependent variable: blood pressure
• y = 0 if normal or elevated blood pressure
(systolic < 140 and diastolic < 90 mm Hg)
• y = 1 if hypertension stage 1 (systolic 140-
159 OR diastolic 90-99 mm Hg)
• y = 2 if hypertension stage 2 (systolic 160 or
higher OR diastolic 100 mm Hg or higher)
We want to analyze how regressors affect blood
pressure.
• Let 𝒚∗ = an indicator of blood pressure and model it as
𝑦 ∗ = 𝑋𝛽 + 𝜀
• Higher 𝒚∗ indicates higher blood pressure
• Let 𝒚 = 0, 1, 2 be the categories of blood pressure, where 0 = Normal, 1 =
Hypertension stage 1, 2 = Hypertension stage 2.
• Decision rule: 𝑦 = 0 if 𝑦 ∗ < 𝑢"
𝑦 = 1 if 𝑢" < 𝑦 ∗ < 𝑢#
𝑦 = 2 if 𝑦 ∗ > 𝑢#
THE ORDERED PROBIT MODEL
THE ORDERED PROBIT MODEL
• We assume 𝒚∗ is a function of independent variables X and error terms
𝑦 ∗ = 𝛽"𝑥"+. . . +𝛽$ 𝑥$ + 𝜀 = 𝑋𝛽 + 𝜀
• Coefficients 𝜷 and critical values 𝒖𝟏 , 𝒖𝟐 will be estimated by the model
THE PROBABILITY
• Consider the case of Pr 𝑦' = 0 = Pr 𝑦'∗ ≤ 𝑢"
= Pr 𝑥' 𝛽 + 𝜀' ≤ 𝑢"
= Pr 𝜀' ≤ 𝑢" − 𝑥' 𝛽
= 1 − Φ 𝑥' 𝛽 − 𝑢"
• Note the error term is normally distributed.
THE PROBABILITY
• The case of Pr 𝑦' = 2 = Pr 𝑦'∗ > 𝑢#
= Pr 𝑥' 𝛽 + 𝜀' > 𝑢#
= Pr 𝜀' > 𝑢# − 𝑥' 𝛽
= 1 − Φ 𝑢# − 𝑥' 𝛽
= Φ 𝑥' 𝛽 − 𝑢#
PROBABILITY OF Y=1
Pr 𝑦' = 1 = Pr 𝑢" ≤ 𝑦'∗ ≤ 𝑢#
= Pr 𝑦'∗ ≤ 𝑢# − Pr 𝑦'∗ ≤ 𝑢"
= Pr 𝑥' 𝛽 + 𝜀' ≤ 𝑢# − Pr 𝑥' 𝛽 + 𝜀' ≤ 𝑢"
= Pr 𝜀' ≤ 𝑢# − 𝑥' 𝛽 − Pr 𝜀' ≤ 𝑢" − 𝑥' 𝛽
= Φ 𝑢# − 𝑥' 𝛽 − Φ 𝑢" − 𝑥' 𝛽
= 1 − Φ 𝑥' 𝛽 − 𝑢# − 1 + Φ 𝑥' 𝛽 − 𝑢"
= Φ 𝑥' 𝛽 − 𝑢" − Φ 𝑥' 𝛽 − 𝑢#
THE PROBABILITIES
• Pr 𝑦' = 0 = 1 − Φ 𝑥' 𝛽 − 𝑢"
• Pr 𝑦' = 1 = Φ 𝑥' 𝛽 − 𝑢" − Φ 𝑥' 𝛽 − 𝑢#
• Pr 𝑦' = 2 = Φ 𝑥' 𝛽 − 𝑢#
• Note that the sum of all probabilities is always 1.
LIKELIHOOD FUNCTION
• The log-likelihood function
𝐿𝐿 = 9 9 𝑌$ ln Pr 𝑦' = 𝑘
' $
1 𝑖𝑓 𝑦' = 𝑘
𝑌$ =
0 𝑖𝑓 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
A CASE STUDY: BLOOD PRESSURE
• Data collected from 300 outpatients visited a hospital in Kien Giang in 2017.
• Data file: ordinal.xlsx
• Dep. Var: blood pressure bp
• bp = 0 if normal or elevated blood pressure (systolic < 140 ; diastolic < 90 mm Hg)
• bp = 1 if hypertension stage 1 (systolic 140-159 OR diastolic 90-99 mm Hg)
• bp = 2 if hypertension stage 2 (systolic ≥ 160 OR diastolic ≥ 100 mm Hg)
• Independent variables:
• age: in years
• female: 1 = female, 0 otherwise
• married: 1 = married and living with, 0 otherwise.
• bmi: body mass index
• workout: 1 = physical exercises at least 15m a day, 0 otherwise
• fam: 1 = having family history of blood pressure, 0 otherwise
• eat: 1 = consuming too much salt, 0 otherwise (self-reported)
• rel: Likert scale 1 = very good family relationship, 5 = very bad family relationship
• cigar: number of cigarettes per day
A CASE STUDY: BLOOD PRESSURE
IMPORTING DATA
DEPENDENT VARIABLE:
BLOOD PRESSURE
CATEGORICAL
REGRESSORS
BIVARIATE ANALYSIS:
CONTINUOUS REGRESSORS
BIVARIATE ANALYSIS:
CATEGORICAL REGRESSORS
BIVARIATE ANALYSIS: CATEGORICAL
REGRESSORS (CONT)
ORDERED
PROBIT
MODEL
ADDING THE P-VALUES
LR TEST FOR OVERALL SIGNIFICANCE
THE FITTED VALUES
(PROBABILITY)
PREDICTION AFTER
ORDERED PROBIT MODEL
PREDICT FOR A SPECIFIC VALUE
GOODNESS of fit: r-squares
HYPOTHESIS
TESTING
MARGINAL EFFECTS
THE ORDERED LOGIT MODEL
• Similar to logit and probit, whether the model is ordered logit or probit
depends on the assumption on the distribution of the error terms 𝜺
• logistic: ordered logit model
• normal: ordered probit model
• Probit is more popular
ORDERED LOGIT
MODEL IN RSTUDIO
