Topics in Econometrics - M2 ENS Lyon
Design: decisions of data collection and measurement
Modelling: define statistical models
Analysis: estimation and questions of statistical inference
Model: formal representation of a theory necessary simplified
Statistical model: a mathematical model of the DGP generating a random sample of which we observe a realization
Regression model:
Conditional distribution of :
Need to choose the functional form for
In particular, choosing which variables to include
Often informed by theory
Parametric:
Non-parametric: makes no assumptions about a parametric distribution
Semi-parametric: contains both a finite-dimensional set of parameters of interest and an infinite-dimensional nuisance component
Structural causal model framework, represented through Directed Acyclic Graphs (DAGs)
Developed mainly in Pearl (2000)
Potential outcomes framework and the structural causal model framework are complementary approaches to causal inference
A DAG is a picture of the assumptions we make. It makes explicit:
Nodes are random variables (here, in blue if observed, in beige if unobserved)
Arrows represent possible direct causal relationships
A path is any sequence of adjacent edges, regardless of the direction of the arrows
A path is open if it can generate an association, closed otherwise
Conditioning on a variable along a path may open or close it
Acyclic: no directed path leaving a variable comes back to it
causally affects through : the path is part of the effect
Not conditioning on identifies the total effect
Conditioning on identifies the direct effect only
causally affects both and : it creates a non-causal association between and
The path is open by default
Conditioning on blocks it and identifies the total causal effect
is caused by both and
The path through is blocked by default
Conditioning on , or a descendant, opens it and creates an association
| Structure | Path | Open by default? | Condition on ? |
|---|---|---|---|
| Mediating | Yes, it is part of the effect | Only for the direct effect | |
| Confounding | Yes | Yes, it blocks the path | |
| Colliding | No | No, it would open the path |
Four paths linking to :
The absence of an arrow is not an absence of assumption: it is a structural exclusion restriction
Effect of college on earnings (Cunningham 2021)
parental education, family income
unobserved background (ability, environment, networks)
Each arrow is a direct causal relationship we are willing to assume
There is no arrow: background matters only through , and
Four paths between and :
The last three are back-door paths
No collider on any of them all three are open
All end with conditioning on income alone closes them
If all back-door paths are blocked, the observed association identifies the total causal effect
Here, three back-door paths:
We block all of them by conditioning on:
Or use an instrument: is a candidate here, too after conditioning on
+ Make every assumption explicit, arrows and missing arrows
+ Non-parametric: no functional form assumed
+ Give a mechanical rule to pick the adjustment set
- Say nothing about functional form, monotonicity or continuity (ie nothing about LATE or RD)
- Do not display effect heterogeneity nor compliance types
- Cannot represent simultaneity or feedback loops
Take-away message
A DAG does not identify anything for us, it forces us to state what we assume and tells us what follows from it
| Assumption | Idea | If violated |
|---|---|---|
| 1. Linearity | Model linear in its parameters | Estimates misspecified biased/inconsistent |
| 2. No perfect collinearity | exists | Coefficients not identifiable |
| 3. Exogeneity | Estimator biased | |
| 4.a. Independent errors | for (eg no autocorrelation) | Invalid inference |
| 4.b. Homoskedasticity | Inefficient |
| Role of the control | In the DAG | Adjusting for it |
|---|---|---|
| Confounder | Removes bias | |
| Predictor of | only | Does not change bias, more precise |
| Mediator | Removes part of the effect | |
| Collider | Creates bias |
A model is saturated when it has one parameter per combination of the regressors
+ No restriction left on : nothing to misspecify
- Quickly needs a lot of controls (and continuous variables need to be binned first)
- Removes a lot of variation in the treatment
Modelling choices (eg which controls to include) can affect the design of our studies
They can affect amount of identifying variation available:
can affect statistical power and create exaggeration
| Limited | Example | Regression model |
|---|---|---|
| Binary | Probit, logit, … | |
| Categorical | Multinomial logit, … | |
| Count | Poisson, negative binomial, … | |
| Censored | eg | Censored regression models (eg tobit) |
Link function , the identity
This is an OLS on a binary outcome
Why not stick with this?
Predicted probabilities are not constrained to
is highly non-linear in X: the effect of an increase of 1 in X is not the same if
LPM is thus biased and inconsistent
LPM fine for an ATE: with a saturated model it is the difference in means
In a logit model,
In a probit, = quantile function of the standard normal (the inverse of the CDF of the normal distribution)
These models are non-linear regression models
Probit and logit are very similar, except in the tails, ie for very rare events
Rule of thumb: probit coefs logit coefs divided by 1.6
A logit/probit point estimate gives the direction, not the magnitude
The marginal effect depends on : effects largest in the middle of the distribution
The inverse-logit curve is steepest at , where its slope is
is an upper bound of the relation, in percentage points
Ok approximation when
Outcome: votes Republican
|
|||
|---|---|---|---|
| LPM | Logit | Probit | |
| Income (in 1000 USD) | 0.01261 | 0.08684 | 0.04939 |
| (0.00037) | (0.00508) | (0.00253) | |
| Observations | 1000 | 1000 | 1000 |
Standard errors in parentheses. Intercept estimated but not reported.
Only the LPM is already in percentage points
The others are not directly interpretable, they need to be “converted”
Three ways to turn 0.087 into a probability:
The Average Marginal Effect averages the flat parts of the curve
The LPM slope is approximately equal to the AME the LPM is often fine
The odds of an event compare how likely it is to happen to how likely it is not:
Odds map to and log odds to while probabilities only take values in
A logit assumes that the log odds are linear in :
Why use this? A probability cannot be linear in without leaving , the log odds can
A one unit larger is associated with larger log odds, ie odds are multiplied by
is the odds ratio
Why use this? Summary of that does not depend on where we start from
The odds ratio is the same everywhere, the change in probability is not
Yet, odds ratios are not straightforward to interpret
participation on income using glm, with a logit then a probit linkSometimes is categorical but unordered: mode of transport, occupation, brand bought
Use a multinomial logit: essentially a logit for each category, modeled against a reference category
one set of coefficients per category
Why use this? The probabilities of all categories must sum to one, separate logits would not impose it
Relies on Independence of Irrelevant Alternatives (IIA): the relative odds of two categories do not depend on the others available
nnet::multinom(), or MASS::polr() when the categories are ordered
job on education using nnet::multinom()job is ordered. Run MASS::polr() too: what do you get instead?Many outcomes are counts: patents, deaths, hospital admissions, crimes, trips
We may want to model them
A Poisson regression assumes
It assumes that the link function is the logarithm , ie
The expected count is therefore multiplicative:
OLS fits a straight line to a multiplicative relationship: predictions go below zero
It reports a single slope, when the change in depends on the level of
Counts are heteroskedastic by construction
is not defined at
Outcome: patents filed
|
||
|---|---|---|
| OLS | Poisson | |
| R&D (in million USD) | 1.006 | 0.298 |
| (0.024) | (0.007) | |
| Observations | 1000 | 1000 |
Standard errors in parentheses. Intercept estimated but not reported.
0.3 one million more of R&D multiplies the expected count by 1.35
ie 35% more patents, whether the firm expects 1 or 10 of them
Rule of thumb: for a small ,
In levels, OLS reports the same 1.01 patents per million, wherever we stand
A Poisson model has a single parameter and imposes
Counts are usually more spread out than that: this is overdispersion
Thus, better to use a negative binomial model
It adds a parameter and allows
Interpreting: “a 1-unit difference in is associated with a -unit difference in the log of the expected count”
These models are non-linear (except LPM) they cannot be estimated by OLS
Estimated by maximum likelihood: we assume a distribution for , then pick the that makes our data the most likely
A logit assumes , a Poisson
The distribution choice affects and actually defines the standard errors
PPML: fitting by maximising a Poisson likelihood, without assuming is Poisson
Ok because maximising it = solving
Only depends on the mean is consistent as soon as the mean is right
Why use it rather than a log? drops the zeros, depends on the units of
then reads directly as a percentage difference in
Since s.e. computed from an incorrect likelihood, we need to report robust s.e.
DriversKilled on law and PetrolPrice using glm, with a log linklaw?DriversKilled. Do Poisson assumptions make sense?MASS::glm.nb() and compare the resultsMany choices when building regression models
Choices not straightforward and can affect the credibility of our estimates
The Causal Graph Framework can be particularly helpful to think about the relations between variables and causality
We often deal with limited outcomes. That needs to (and can) be accounted for using several types of models.