1/89
Comprehensive practice flashcards covering Discrete Variable Models: Ordered Categorical Models (Ordered Logit and Probit), Unordered Categorical Models (Multinomial Logit), Count Data Models (Poisson, Poisson PML, and Negative Binomial), and Response Heterogeneity (HOPIT with Anchoring Vignettes).
What is an ordered categorical variable?
A discrete variable whose response categories have a clear, natural ordering, but whose values are not continuous (generally coded as y=1,2,…,J).
What are three examples of ordered categorical variables mentioned in the lecture?
Self-assessed health, risk preferences/attitudes, firm sales growth (low to high), and student course evaluations.
In the working mothers practical example, what dataset and sample size were used?
The 1977 and 1989 US General Social Survey (GSS), with a sample size of N=2,293.
How is the dependent variable warm defined and coded in the working mothers example?
Based on agreement with the statement that a working mother can establish just as warm and secure a relationship with her child: 1=Strongly Disagree (SD), 2=Disagree (D), 3=Agree (A), and 4=Strongly Agree (SA).
What are the two key explanatory variables in the baseline ordered model for warm working mothers?
x1=yr89 (a dummy variable with 1977 as the reference category) and x2=ed (education in years).
Why can standard linear regression models not be used for ordered categorical outcomes?
Because the outcome variable is discrete and ordinal; numerical category codes reflect ranking rather than continuous quantitative differences or equal intervals.
How is the underlying continuous process modeled in ordered regression models?
Through a continuous latent variable y∗=β1x1+β2x2+u, which determines the observed discrete response y.
What is the mathematical relationship mapping the latent variable y∗ to the observed category y=j?
y=j if τj−1≤y∗<τj for j=1,…,J, where τ0=−∞, τJ=+∞, and τ1≤τ2≤⋯≤τJ−1 are unknown thresholds to be estimated.
What distributional assumption on the error term u distinguishes ordered probit from ordered logit?
In ordered probit, u follows a standard normal distribution (u∼N(0,1)), whereas in ordered logit, u follows a standard logistic distribution.
What is the formula for P(y=j∣x1,x2) in the ordered probit model?
P(y=j∣x1,x2)=Φ(τj−β1x1−β2x2)−Φ(τj−1−β1x1−β2x2), where Φ(⋅) is the standard normal cumulative distribution function (CDF).
In an ordered logit model, why does P(y=1∣x1,x2) simplify to a single fraction?
Because τ0=−∞, causing exp(τ0−β1x1−β2x2)=0, so the subtracted term equals zero: P(y=1∣x1,x2)=1+exp(τ1−β1x1−β2x2)exp(τ1−β1x1−β2x2).
In an ordered logit model with J=4, why does P(y=4∣x1,x2) simplify to 1−1+exp(τ3−β1x1−β2x2)exp(τ3−β1x1−β2x2)?
Because τ4=+∞, which makes the leading term 1+exp(τ4−β1x1−β2x2)exp(τ4−β1x1−β2x2)=1.
In the working mothers ordered logit regression, what were the estimated coefficients for yr89 and ed?
β^yr89=0.5146 (z=6.50) and β^ed=0.1024 (z=8.22).
What were the three estimated cut-points (/cut1, /cut2, /cut3) in the baseline ordered logit model for warm?
τ1=−0.5120, τ2=1.2401, and τ3=3.0286.
What is the test statistic for testing a single hypothesis H0:βk=β∗ in ordered logit or probit models?
z=σ^β^kβ^k−β∗∼N(0,1) asymptotically.

What econometric estimation concept is depicted in this figure?
Maximum Likelihood estimation, illustrating that the parameter estimate θ^ML maximizes the log-likelihood function logL.
How is the total log-likelihood function logL defined across n observations in an ordered model?
logL=∑i=1nlogf(yi∣x1i,x2i), where f(yi∣⋅)=P(y=j∣⋅) for the observed response category j.
What is the formula and asymptotic distribution of the Likelihood Ratio Test (LRT) statistic?
LRT=2(logLUR−logLR)∼χ2(q), where q is the number of parameter restrictions.
In Stata, what command performs a Wald test of multiple coefficient restrictions after ologit?
The test command (e.g., test male white age prst), which evaluates restrictions using an asymptotic χ2(q) distribution.
How is McFadden's Pseudo-R2 defined in ordered logit and probit models?
Pseudo-R2=1−logL0logL, where logL is the log-likelihood of the estimated model and logL0 is the log-likelihood of an intercept-only model.
If β2>0 in an ordered model, what are the unambiguous signs of the marginal effects on P(y=1) and P(y=J)?
The marginal effect on P(y=1) is strictly negative, and the marginal effect on the highest category P(y=J) is strictly positive.
Why can the sign of the marginal effect on intermediate categories (j=2,…,J−1) not be determined by the sign of βk alone?
Because intermediate probabilities depend on shifts between two adjacent cut-points, meaning their marginal effects can be positive or negative depending on the specific values of the covariates.
What is the difference between an average marginal effect and a conditional marginal effect?
An average marginal effect calculates the marginal effect for every individual observation and averages them across the sample (margins, dydx(*)), while a conditional marginal effect evaluates the effect at specified values of the covariates (margins, dydx(*) at(...)).
What characterizes an unordered categorical variable?
A discrete outcome whose categories have no natural ordering or hierarchy (e.g., transportation mode, choice of healthcare provider, or firm innovation strategy).
How many sets of slope coefficients are estimated in a Multinomial Logit (MNL) model with M categories?
M−1 sets of coefficients, because one category must be set as the base/reference category whose parameters are normalized to zero.
In the Portuguese healthcare provider example, what were the three categories of the variable choice?
1=no specialist consultations in the past 3 months (base outcome), 2=consultation only with public specialist(s), and 3=consultation with private specialist(s).
What were the sample size and explanatory variables in the healthcare choice MNL example?
The sample size was N=22,772, with explanatory variables female (dummy variable) and educ (education in years).
What is the formula for the probability of the base category (y=1) in a 3-category MNL model?
P(y=1∣x1,x2)=1+exp(β0,2+β1,2x1+β2,2x2)+exp(β0,3+β1,3x1+β2,3x2)1.
What is the formula for the probability of an alternative category m (m>1) in an MNL model?
P(y=m∣x1,x2)=1+∑k=2Mexp(β0,k+β1,kx1+β2,kx2)exp(β0,m+β1,mx1+β2,mx2).
What does the sign of an MNL coefficient β1,m indicate?
It indicates the sign of the effect of x1 on the probability of choosing category m relative to the base category (P(y=1)P(y=m)), not on the absolute probability P(y=m).
Does a positive coefficient β1,m>0 guarantee that an increase in x1 increases the absolute probability P(y=m)?
No; P(y=m) could decrease if the probabilities of other alternative categories increase by an even greater amount.
In the healthcare choice MNL regression, what were the estimated coefficients of female for public and private care?
For public care (m=2), β^female,2=0.1274 (p=0.001); for private care (m=3), β^female,3=0.1416 (p=0.026), both relative to no care.
In the healthcare choice MNL regression, what were the estimated coefficients of educ for public and private care?
For public care (m=2), β^educ,2=−0.0426 (p=0.000); for private care (m=3), β^educ,3=0.0608 (p=0.000), both relative to no care.
Why is it incorrect to state that β^female,2=0.1274 means being female increases the probability of public care by 12.7 percentage points?
Because MNL regression coefficients reflect log-relative-risk changes relative to the base category, not direct percentage point changes in absolute probabilities.
How is the overall hypothesis that female has no effect on any healthcare choice tested in Stata?
By testing the joint hypothesis H0:βfemale,2=0 and βfemale,3=0 using the command test [2]female [3]female.
What was the result of the joint Wald test for female in the healthcare MNL model?
χ2(2)=14.21 with p=0.0008, rejecting the null hypothesis of no gender effect at the 1% significance level.

In this Stata plot of predicted probabilities from an MNL model, how do the choices change as education increases?
P(choice=1) (no care) stays dominant around 0.8, P(choice=2) (public care) decreases from around 0.19 to 0.07, and P(choice=3) (private care) increases from around 0.03 to 0.15.
In the healthcare MNL model, what was the average marginal effect of female on P(choice=1) (no care)?
−0.0207 (p<0.001), meaning being female decreases the probability of having no specialist consultation by approximately 2.07 percentage points on average.
In the healthcare MNL model, what was the average marginal effect of female on P(choice=2) (public care)?
+0.0149 (p=0.002), meaning being female increases the probability of public specialist consultation by approximately 1.49 percentage points on average.
In the healthcare MNL model, what was the average marginal effect of female on P(choice=3) (private care)?
+0.0058 (p=0.050), meaning being female increases the probability of private specialist consultation by approximately 0.58 percentage points on average.
What is the mathematical sum of the marginal effects of a covariate across all categories in an MNL model, and why?
The sum equals zero (∑m=1M∂xk∂P(y=m)=0), because the probabilities of all mutually exclusive categories must always sum to 1.
What Stata command calculates the average marginal effect of covariates on the third category in an MNL model?
margins, dydx(*) predict(outcome(3)).
In the healthcare MNL model, what were the average marginal effects of education on public vs private specialist visits?
Education had a negative marginal effect on public care (−0.0058, p<0.001) and a positive marginal effect on private care (+0.0032, p<0.001).
By default, how does Stata's mlogit command select the base category?
Stata automatically chooses the category with the highest frequency as the base outcome, unless specified otherwise by the user.
Can the sign of the marginal effect of xk on P(y=m) differ from the sign of the coefficient βk,m in an MNL model?
Yes, because the marginal effect depends on the coefficients and probabilities of all other alternatives, not just category m.
What is a count variable?
A discrete, non-negative integer variable (y∈{0,1,2,…}) representing the number of times an event occurs.
What are four empirical examples of count variables presented in the lecture?
Number of doctor visits/diagnostic tests, alcoholic drinks per week, cigarettes smoked per day, and unpaid parking violations.
What was the empirical question studied by Fisman and Miguel (2005/2007) in their UN parking tickets paper?
Whether cultural norms of corruption affect compliance with laws in the absence of legal enforcement, using unpaid parking violations of UN diplomats in Manhattan who had diplomatic immunity until 2002.
In the UN parking violations sample (N=149), what were the sample mean and variance of violations?
The sample mean was 11.28 and the sample variance was 337.93.
What is the defining property of the Poisson distribution regarding its mean and variance?
Equidispersion: the conditional variance equals the conditional mean (E(y)=Var(y)=μ).
How is the conditional expectation function E(y∣x) specified in Poisson and Negative Binomial models?
E(y∣x1,x2)=μ=exp(β0+β1x1+β2x2), ensuring that expected counts are strictly non-negative.
What is overdispersion in count data?
A condition where the conditional variance exceeds the conditional mean (Var(y)>E(y)).
What are the statistical consequences of overdispersion when estimating a count model using standard Poisson Maximum Likelihood?
The parameter estimates of β and E(y∣x) remain consistent, but standard errors are biased downwards, producing inflated z-statistics and spuriously small p-values.
How does the Negative Binomial (NB) model specify the conditional variance to accommodate overdispersion?
Var(y)=μ(1+αμ)=μ+αμ2, where α>0 is the overdispersion parameter.
Under what condition does the Negative Binomial model simplify to the standard Poisson model?
When the overdispersion parameter α=0, which makes Var(y)=μ(1+0)=μ=E(y).
What is Poisson Pseudo-Maximum Likelihood (Poisson PML)?
Estimating a Poisson regression model using robust (sandwich) standard errors (robust option in Stata) to provide valid standard errors without requiring the equidispersion assumption E(y)=Var(y).
How do Poisson PML and Negative Binomial regression compare in estimating probabilities P(y=j)?
Negative Binomial specifies a full distribution and is good at estimating individual probabilities P(y=j), whereas Poisson PML only correctly models the conditional mean and cannot reliably estimate P(y=j) under overdispersion.
In the Poisson ML estimation of the parking tickets data, what were the coefficients and z-statistics for staff and corruption?
For staff, β^=0.0164 (z=9.57); for corruption, β^=0.4873 (z=10.51).
When re-estimating the parking violations model with Poisson PML (robust standard errors), what happened to the z-statistics?
The coefficient estimates remained identical, but the robust standard errors increased, causing the z-statistic for staff to fall to 2.23 (p=0.026) and for corruption to fall to 2.67 (p=0.008).
In the Negative Binomial estimation of parking tickets, what was the estimated overdispersion parameter alpha and its significance test?
α^=1.5840 (Std. Err.=0.2007); the likelihood ratio test of α=0 gave χˉ2=1772.10 (p<0.001), confirming severe overdispersion.
What is the formula for the marginal effect of a continuous variable x2 on E(y∣x) in count models?
∂x2∂E(y∣x)=β2exp(β0+β1x1+β2x2)=β2E(y∣x).
How is the average marginal effect of a continuous variable x2 approximated in count models?
AME≈β2yˉ, where yˉ is the sample mean of the dependent count variable.
In the Poisson PML parking tickets model, what was the average marginal effect of staff?
AME=0.0164049×11.27517≈0.1858 (p=0.036), meaning one additional diplomat increases expected unpaid violations by approximately 0.19.
How is the coefficient βk in a Poisson or Negative Binomial model interpreted as a relative effect?
As a semi-elasticity: βk=E(y∣x)∂E(y∣x)/∂xk, meaning a 1-unit increase in xk changes the expected count by approximately βk×100%.
In the Poisson parking tickets model, what is the relative effect interpretation of β^staff=0.0164?
One additional diplomatic staff member increases the expected number of parking violations by 1.64%, holding other variables fixed.
When comparing fitted counts to actual parking violations, what were the squared correlations for Poisson PML and Negative Binomial?
For Poisson PML, r2=0.1007; for Negative Binomial, r2=0.0056.
According to the countfit results for parking violations, how did Poisson and NB compare in predicting the proportion of zero counts (actual = 0.189)?
The standard Poisson model severely underpredicted zeros at 0.018, while the Negative Binomial model accurately predicted zeros at 0.194.
What is the recommended workaround in Stata if countfit produces the error 'factor-variable and time-series operators not allowed'?
Manually generate dummy variables for each category level (e.g., tab year, gen(yeardummy)) and include the created dummies in the regression rather than using factor notation (e.g., i.year).
What is reporting heterogeneity (or cut-point shift) in self-reported survey data?
A phenomenon where different individuals or groups have different internal thresholds/standards (τj) when mapping the same underlying latent state (y∗) into categorical response options (y).
What evidence from the European Community Household Panel (ECHP 2001) illustrated reporting heterogeneity in self-reported health?
Respondents in Denmark reported much higher health ratings than respondents in Portugal, even though objective epidemiological measures showed no such large health discrepancy.
What are several potential sources of reporting heterogeneity identified in the lecture?
Differences in standards, awareness, expectations, adaptation to chronic illness, reference groups, and cultural norms.
Why is the standard ordered probit model problematic when reporting heterogeneity is present?
It assumes constant cut-points across all individuals, causing differences in reporting standards to be falsely attributed to true differences in the latent index.
In the SHARE France breathing ability study, what was the dependent variable y?
Self-rated shortness of breath in the last 30 days, categorized into five ordered options: Extreme, Severe, Moderate, Mild, and None (N=837).
What was the result of testing the effect of education on breathing problems in the standard ordered probit model?
The joint Wald test on the education dummies gave χ2(3)=2.41 (p=0.4918), indicating no statistically significant effect of education on breathing ability.
What is the Hierarchical Ordered Probit (HOPIT) model?
An extended ordered probit model that parameterizes the cut-point thresholds τj as linear functions of individual characteristics to account for response heterogeneity.
How does the HOPIT model specify the individual-specific cut-points τj mathematically?
τj=γ0,j+γ1,jx1+γ2,jx2 for j=1,…,J−1.
Why can the HOPIT model not be identified using self-assessed health data (y) alone?
Because without an external benchmark, shifts in the cut-points (γ) cannot be mathematically distinguished from shifts in the latent index (β).
What are anchoring vignettes, and how do they resolve the identification problem in HOPIT models?
Hypothetical descriptions of fixed health states evaluated by all respondents; because the described state is identical, any systematic variation in vignette ratings reflects differing cut-points rather than differences in health.
What is the latent equation for a vignette rating vl in the HOPIT model?
vl∗=αl+ul, where ul∼N(0,1) and αl is a constant fixed health state, meaning vl∗ is not a function of individual covariates.
What is the 'Vignette Equivalence' assumption in the HOPIT framework?
The assumption that all respondents perceive and interpret the hypothetical health state described in a vignette in the same way, up to a random i.i.d. error term.
What is the 'Response Consistency' assumption in the HOPIT framework?
The assumption that an individual applies the exact same threshold cut-points (τj) when evaluating their own health as they do when evaluating anchoring vignettes.
In the French breathing study, what happened to the education coefficients on latent health (y∗) in HOPIT compared to ordered probit?
Education became positive and highly significant (level 3: β^=0.3283, p=0.010; level 4: β^=0.4990, p=0.001), revealing that higher education increases true breathing ability.
What did the HOPIT cut-point estimates reveal about the reporting standards of higher-educated individuals in France?
Higher-educated individuals have significantly higher cut-points (τ3 and τ4), meaning they have stricter health standards and are less likely to report having 'no problem' for a given breathing state.
What was the result of testing for reporting heterogeneity by education across all cut-points in the French breathing HOPIT model?
The joint Wald test across the 12 education threshold coefficients yielded χ2(12)=31.85 (p=0.0015), rejecting the null of reporting homogeneity at the 1% level.
What was the result of testing for reporting heterogeneity jointly across education, gender, and age in the HOPIT model?
χ2(20)=49.87 (p=0.0002), confirming that cut-points vary significantly across demographic characteristics.
Why are raw predicted probabilities P(y=j) not directly comparable across individuals in the presence of response heterogeneity?
Because each individual evaluates their latent state against a different personal set of threshold cut-points (τj).
How are predicted probabilities 'cleaned' of reporting heterogeneity in the HOPIT model?
By calculating probabilities using the estimated latent index coefficients (β^) while fixing the cut-point thresholds (τj) to those of a chosen reference individual for everyone in the sample.
What reference profile was used to fix cut-point τ4 when evaluating average marginal effects on P(y=5) in the French breathing example?
A male individual, aged 64, with the lowest education level (educ1).
How did the marginal effect of highest education (educ4) on P(y=5) (no breathing problems) change after HOPIT correction?
It increased from an insignificant +0.0676 in ordered probit to a statistically significant +0.1811 (p<0.001) in HOPIT.
What was the major substantive conclusion reached by Bago d'Uva, O'Donnell, and van Doorslaer (2008)?
Education-related health inequalities are often severely underestimated or go undetected in uncorrected survey data because higher-educated individuals evaluate health with stricter standards; correcting for reporting heterogeneity reveals significant inequalities.