Autocorrelation

class: center, middle, inverse, title-slide

.title[
# Autocorrelation
]
.subtitle[
## EC 421, Set 8
]
.author[
### Edward Rubin
]

---

class: inverse, middle

# Prologue

---
name: schedule

# Schedule

## Last Time

- Introduction to time series
- Midterm

## Today

Autocorrelation

---
layout: true
# .mono[R] showcase

---
name: r_showcase

## Functions

Writing your own functions.

---
layout: true
# Writing functions

---
name: functions
## Functions are everywhere

Everything you do in .mono[R] involves some sort of function, _e.g._,
- `mean()`
- `lm()`
- `summary()`
- `read_csv()`
- `ggplot()`
- `+`

The basic idea in .mono[R] is doing things to objects with functions.

---

## Functions can help

We write functions to make life easier. Instead of copying and pasting the same line of code a million times, you can write one function.

In .mono[R], you use the `function()` function to write functions..pink[†]

.footnote[
.pink[†] Meta.
]

``` r
# Our first function
the_name <- function(arg1, arg2) {
 # Insert code that involves arg1 and arg2 (this is where the magic happens)
}
```
- `the_name`: The name we are giving to our new function.
- `arg1`: The first argument of our function.
- `arg2`: The second argument of our function.

---

## Our first real function

Let's write a function that multiplies two numbers. (It needs two arguments.)

``` r
# Create our function
the_product <- function(x, y) {
 x * y
}
```
--

Did it work?

``` r
the_product(7, 15)
```

```
#> [1] 105
```

.bigger[💪]

---

## Functions can do anything

... that you tell them.

If you are going to repeat a task (_e.g._, a simulation), then you have a good situation for writing your own function.

.mono[R] offers many functions (via its many packages), but you will sometimes find a scenario for which no one has written a function.

Now you know how to write your own.

``` r
# An ad lib function
ad_lib <- function(noun1, verb1, noun2) {
 paste("The next", noun1, "of our lecture", verb1, noun2)
}
```

---

``` r
ad_lib(noun1 = "part", verb1 = "reviews", noun2 = "time series.")
```

```
#> [1] "The next part of our lecture reviews time series."
```

---
layout: true
# Time series
## Review

---
class: inverse, middle

---
name: review

Changes to our model/framework.

- Our model now has `$\color{#e64173}{t}$` subscripts for .hi[time periods].

- .hi[Dynamic models] allow .hi[lags] of explanatory and/or outcome variables.

- We changed our **exogeneity** assumption to .hi[contemporaneous exogeneity], _i.e._, `$\mathop{\boldsymbol{E}}\left[ u_t \middle| X_t \right] = 0$`

- Including .hi-orange[lags of outcome variables] can lead to .hi[biased coefficient estimates] from OLS.

- .hi-orange[Lagged explanatory variables] make .hi[OLS inefficient].

---
layout: false
class: inverse, middle

# Autocorrelation

---
layout: false
name: intro
# Autocorrelation
## What is it?

.hi[Autocorrelation] occurs when our disturbances are correlated over time, _i.e._, `$\mathop{\text{Cov}} \left( u_t,\, u_s \right) \neq 0$` for `$t\neq s$`.

Another way to think about: If the *shock* from disturbance `$t$` correlates with "nearby" shocks in `$t-1$` and `$t+1$`.

*Note:* **Serial correlation** and **autocorrelation** are the same thing.

Why is autocorrelation prevalent in time-series analyses?

---
class: clear

.hi-slate[Positive autocorrelation]: Disturbances `$\left( u_t \right)$` over time
<img src="slides_files/figure-html/positive auto u-1.svg" style="display: block; margin: auto;" />

---
class: clear

.hi-slate[Positive autocorrelation]: Outcomes `$\left( y_t \right)$` over time
<img src="slides_files/figure-html/positive auto y-1.svg" style="display: block; margin: auto;" />

---
class: clear

.hi-slate[Negative autocorrelation]: Disturbances `$\left( u_t \right)$` over time
<img src="slides_files/figure-html/negative auto u-1.svg" style="display: block; margin: auto;" />

---
class: clear

.hi-slate[Negative autocorrelation]: Outcomes `$\left( y_t \right)$` over time
<img src="slides_files/figure-html/negative auto y-1.svg" style="display: block; margin: auto;" />

---
layout: true
# Autocorrelation
## In static time-series models

---
name: static_models

Let's start with a very common model: a static time-series model whose disturbances exhibit .hi[first-order autocorrelation], *a.k.a.* .pink[AR(1)]:

$$
`\begin{align}
  \text{Births}_t &= \beta_0 + \beta_1 \text{Income}_t + u_t
\end{align}`
$$
where
$$
`\begin{align}
  \color{#e64173}{u_t} = \color{#e64173}{\rho \, u_{t-1}} + \varepsilon_t
\end{align}`
$$
and the `$\varepsilon_t$` are independently and identically distributed (*i.i.d.*).

.hi-purple[Second-order autocorrelation], or .purple[AR(2)], would be

$$
`\begin{align}
  \color{#6A5ACD}{u_t} = \color{#6A5ACD}{\rho_1 u_{t-1}} + \color{#6A5ACD}{\rho_2 u_{t-2}} + \varepsilon_t
\end{align}`
$$

---

An .turquoise[AR(p)] model/process has a disturbance structure of

$$
`\begin{align}
  u_t = \sum_{j=1}^\color{#20B2AA}{p} \rho_j u_{t-j} + \varepsilon_t
\end{align}`
$$

allowing the current disturbance to correlated with up to `$\color{#20B2AA}{p}$` of its lags.

---
layout: false
name: ols_static
# Autocorrelation
## OLS

For **static models** or **dynamic models with lagged explanatory variables**, in the presence of autocorrelation

1. OLS provides .pink[**unbiased** estimates for the coefficients.]

1. OLS creates .pink[**biased** estimates for the standard errors.]

1. OLS is .pink[**inefficient**.]

*Recall:* Same implications as heteroskedasticity.

Autocorrelation get trickier with lagged outcome variables.

---
layout: true
# Autocorrelation
## OLS and lagged outcome variables

---
name: ols_adl

Consider a model with one lag of the outcome variable—ADL(1, 0)—model with AR(1) disturbances

$$
`\begin{align}
  \text{Births}_t = \beta_0 + \beta_1 \text{Income}_t + \beta_2 \text{Births}_{t-1} + u_t
\end{align}`
$$
where
$$
`\begin{align}
  u_t = \rho u_{t-1} + \varepsilon_t
\end{align}`
$$

**Problem:**
--
 Both `$\text{Births}_{t-1}$` (a regressor in the model for time `$t$`) and `$u_{t}$` (the disturbance for time `$t$`) depend upon `$u_{t-1}$`. *I.e.*, a regressor is correlated with its contemporaneous disturbance.

**Q:** Why is this a problem?
--
 
**A:** It violates .pink[contemporaneous exogeneity]
--
, *i.e.*, `$\mathop{\text{Cov}} \left( x_t,\,u_t \right) \neq 0$`.

---

To see this problem, first write out the model for `$t$` and `$t-1$`:
$$
`\begin{align}
  \text{Births}_t &= \beta_0 + \beta_1 \text{Income}_t + \beta_2 \text{Births}_{t-1} + u_t \\
  \text{Births}_{t-1} &= \beta_0 + \beta_1 \text{Income}_{t-1} + \beta_2 \text{Births}_{t-2} + u_{t-1}
\end{align}`
$$
and now note that `$u_t = \rho u_{t-1} + \varepsilon_t$`. Substituting...
$$
`\begin{align}
  \text{Births}_t &= \beta_0 + \beta_1 \text{Income}_t + \beta_2 \color{#6A5ACD}{\text{Births}_{t-1}} + \overbrace{\left( \rho \color{#e64173}{u_{t-1}} + \varepsilon_t \right)}^{u_t} \tag{1} \\
  \color{#6A5ACD}{\text{Births}_{t-1}} &= \beta_0 + \beta_1 \text{Income}_{t-1} + \beta_2 \text{Births}_{t-2} + \color{#e64173}{u_{t-1}} \tag{2}
\end{align}`
$$

In `$(1)$`, we can see that `$u_t$` depends upon (covaries with) `$\color{#e64173}{u_{t-1}}$`.
 In `$(2)$`, we can see that `$\color{#6A5ACD}{\text{Births}_{t-1}}$`, a regressor in `$(1)$`, also covaries with `$u_{t-1}$`.

∴ This model violates our contemporaneous exogeneity requirement.

---

*Implications:* For models with **lagged outcome variables** and **autocorrelated disturbances**

1. The models .pink[**violate contemporaneous exogeneity**.]

1. OLS is .pink[**biased and inconsistent** for the coefficients.]

---

*Intuition?* Why is OLS inconsistent and biased when we violate exogeneity?

Think back to omitted-variable bias...

$$
`\begin{align}
  y_t &= \beta_0 + \beta_1 x_t + u_t
\end{align}`
$$

When `$\mathop{\text{Cov}} \left( x_t,\, u_t \right)\neq0$`, we cannot separate the effect of `$u_t$` on `$y_t$` from the effect of `$x_t$` on `$y_t$`. Thus, we get inconsistent estimates for `$\beta_1$`.
--
 Similarly,

$$
`\begin{align}
  \text{Births}_t &= \beta_0 + \beta_1 \text{Income}_t + \beta_2 \text{Births}_{t-1} + \overbrace{\left( \rho u_{t-1} + \varepsilon_t \right)}^{u_t} \tag{1}
\end{align}`
$$

we cannot separate the effects of `$u_t$` on `$\text{Births}_t$` from `$\text{Births}_{t-1}$` on `$\text{Births}_{t}$`, because both `$u_t$` and `$\text{Births}_{t-1}$` depend upon `$u_{t-1}$`.
--
 `$\color{#e64173}{\hat{\beta}_2}$` .pink[is **biased** (w/ OLS).]

---
layout: true
# Autocorrelation and bias
## Simulation

---
name: ar2_simulation

To see how this bias can look, let's run a simulation.

$$
`\begin{align}
  y_t = 1 + 2 x_t + 0.5 y_{t-1} + u_t \\
  u_t &= 0.9 u_{t-1} + \varepsilon_t
\end{align}`
$$

One (easy) way generate 100 disturbances from AR(1), with `$\rho = 0.9$`:

``` r
arima.sim(model = list(ar = c(0.9)), n = 100)
```

We are going to run 10,000 iterations with `$T=100$`.

**Q:** Will this simulation tell us about *bias* or *consistency*?
--
 **A:** Bias. We would need to let `$T\rightarrow\infty$` to consider consistency.

---

Outline of our simulation:

.pseudocode-small[
1. Generate T=100 values of x

1. Generate T=100 values of u

- Generate T=100 values of ε

- Use ε and ρ=0.9 to calculate u.sub[t] = ρ u.sub[t-1] + ε.sub[t]

1. Calculate y.sub[t] = β.sub[0] + β.sub[1] x.sub[t] + β.sub[2] y.sub[t-1] + u.sub[t]

1. Regress y on x; record estimates

Repeat 1–4 10,000 times
]

---
layout: false
class: clear

.pull-left[
.hi-slate[Distribution of OLS estimates,] `$\hat{\beta}_2$`
] .pull.right[
`$y_t = 1 + 2 x_t + \color{#e64173}{0.5} y_{t-1} + u_t$`
]

---
layout: false
class: clear

.pull-left[
.hi-slate[Distribution of OLS estimates,] `$\hat{\beta}_1$`
] .pull.right[
`$y_t = 1 + \color{#FFA500}{2} x_t + 0.5 y_{t-1} + u_t$`
]

---
layout: false
class: inverse, middle
name: testing
# Testing for autocorrelation

---
layout: true
# Testing for autocorrelation
## Static models

---
name: testing_static

Suppose we have the **static model**,
$$
`\begin{align}
  \text{Births}_t = \beta_0 + \beta_1 \text{Income}_t + u_t \tag{A}
\end{align}`
$$
and we want to test for an AR(1) process in our disturbances `$u_t$`.

.hi[Test for autocorrelation:] Test for correlation in the lags of our residuals:

$$
`\begin{align}
  e_t = \color{#e64173}{\rho} e_{t-1} + v_t
\end{align}`
$$

Does `$\color{#e64173}{\hat{\rho}}$` differ significantly from zero?

**Familiar idea:** Use residuals to learn about disturbances.

---

Specifically, to test for AR(1) disturbances in the static model
$$
`\begin{align}
  \text{Births}_t = \beta_0 + \beta_1 \text{Income}_t + u_t \tag{A}
\end{align}`
$$

.pseudocode-small[

1\. Estimate `$(\text{A})$` via OLS.

2\. Calculate residuals from the OLS regression in step 1.

3\. Regress the residuals on their lags (without an intercept).

4\. Use a _t_ test to determine whether there is statistically significant evidence that ρ differs from zero.

5\. Rejecting H.sub[0] implies significant evidence of autocorrelation.
]

---
layout: false
class: clear, middle

For an example, let's return to our plot of negative autocorrelation.

---
class: clear

.hi-slate[Negative autocorrelation]: Disturbances `$\left( u_t \right)$` over time
<img src="slides_files/figure-html/negative auto u 2-1.svg" style="display: block; margin: auto;" />

---
layout: true
# Testing for autocorrelation
## Example: Static model and AR(1)

---

**Step 1:** Estimate the static model `$\left( y_t = \beta_0 + \beta_1 x_t + u_t \right)$` with OLS

``` r
reg_est <- lm(y ~ x, data = ar_df)
```

**Step 2:** Add the residuals to our dataset

``` r
ar_df$e <- residuals(reg_est)
```

**Step 3:** Regress the residual on its lag (**no intercept**)

``` r
reg_resid <- lm(e ~ -1 + lag(e), data = ar_df)
```

---

**Step 4:** _t_ test for the estimated `$(\hat{\rho})$` coefficient in step 3.

``` r
tidy(reg_resid)
```

```
#> # A tibble: 1 × 5
#> term estimate std.error statistic p.value
#> <chr> <dbl> <dbl> <dbl> <dbl>
#> 1 lag(e) -0.851 0.0535 -15.9 6.88e-29
```

That's a very small *p*-value—much smaller than 0.05.

.hi[Reject H.sub[0]] (H.sub[0] was `$\rho=0$`, _i.e._, no autocorrelation).

**Step 5:** Conclude. Statistically significant evidence of autocorrelation.

---
layout: true
# Testing for autocorrelation
## Example: Static model and AR(3)

---

What if we wanted to test for AR(3)?

- We add more lags of residuals to the regression in *Step 3*.

- We **jointly** test the significance of the coefficients (_i.e._, `$\text{LM}$` or `$F$`).

Let's do it.

---

**Step 1:** Estimate the static model `$\left( y_t = \beta_0 + \beta_1 x_t + u_t \right)$` with OLS

``` r
reg_est <- lm(y ~ x, data = ar_df)
```

**Step 2:** Add the residuals to our dataset

``` r
ar_df$e <- residuals(reg_est)
```

**Step 3:** Regress the residual on its lag (**no intercept**)

``` r
reg_ar3 <- lm(e ~ -1 + lag(e) + lag(e, 2) + lag(e, 3), data = ar_df)
```
--

*Note:* `lag(v, n)` from `dplyr` takes the .mono[n].super[th] lag of the variable .mono[v].

---

**Step 4:** Calculate the `$\text{LM} = n\times R_e^2$` test statistic—distributed `$\chi_k^2$`.
 
`$k$` is the number of regressors in the regression in *Step 3* (here, `$k=3$`).

``` r
# Grab R squared
r2_e <- summary(reg_ar3)$r.squared
# Calculate the LM test statistic: n times r2_e
(lm_stat <- 100 * r2_e)
```

```
#> [1] 72.38204
```

``` r
# Calculate the p-value
(pchisq(q = lm_stat, df = 3, lower.tail = F))
```

```
#> [1] 1.318485e-15
```

---

**Step 5:** Conclude.

*Recall:* Our hypotheses consider the model
$$
`\begin{align}
  e_t = \rho_1 e_{t-1} + \rho_2 e_{t-2} + \rho_3 e_{t-3}
\end{align}`
$$

which we are actually using to learn about the model
$$
`\begin{align}
  u_t = \rho_1 u_{t-1} + \rho_2 u_{t-2} + \rho_3 u_{t-3}
\end{align}`
$$

H.sub[0]: `$\rho_1 = \rho_2 = \rho_3 = 0$`
 *vs.* H.sub[A]: `$\rho_j \neq 0$` for at least one `$j$` in `$\{1,\, 2,\, 3\}$`

Our p-value is less than 0.05.
--
 .hi[Reject H.sub[0].]

Conclude there is statistically significant evidence of autocorrelation.

---
layout: true
# Testing for autocorrelation
## Dynamic models with lagged outcome variables

---
name: testing_adl

**Recall:** OLS is biased and inconsistent when our model has *both*

1. a lagged dependent variable

1. autocorrelated disturbances

**Problem:** If OLS is biased for `$\beta$`, then it is also biased for `$u_t$`.

∴ We can't apply our nice trick of *just* using `$e_t$` to learn about `$u_t$`.

**Solution:** .hi-purple[Breusch-Godfrey] test includes the other explanatory variables,

$$
`\begin{align}
  e_t = \color{#6A5ACD}{\underbrace{\gamma_0 + \gamma_1 x_{1t} + \gamma_2 x_{2t} + \cdots}_{\text{Explanatory variables (RHS)}}} + \color{#e64173}{\underbrace{\rho_1 e_{t-1} + \rho_2 e_{t-2} + \cdots}_{\text{Lagged residuals}}} + \varepsilon_t
\end{align}`
$$

---

Specifically, to test for AR(2) disturbances in the ADL(1, 0) model
$$
`\begin{align}
  \text{Births}_t = \beta_0 + \beta_1 \text{Income}_t + \beta_2 \text{Births}_{t-1} + u_t \tag{B}
\end{align}`
$$

.pseudocode-small[
1\. Estimate `$(\text{B})$` via OLS.

2\. Calculate residuals (e.sub[t]) from the OLS regression in step 1.

3\. Regress residuals on an intercept, explanatory variables, and lagged residuals.

<center>
 e.sub[t] = γ.sub[0] + γ.sub[1] Income.sub[t] + γ.sub[3] Births.sub[t-1] + ρ.sub[1] e.sub[t-1] + ρ.sub[2] e.sub[t-2] + v.sub[t]
</center>

4\. Conduct LM or F test for ρ.sub[1] = ρ.sub[2] = 0.

5\. Rejecting H.sub[0] implies significant evidence of AR(2).
]

---

For an example, let's consider the relationship between monthly presidential approval ratings and oil prices during President George W. Bush's.pink[†] presidency.

We will specify the process as ADL(1, 0) and test for an AR(2) process in our disturbances.

$$
`\begin{align}
  \text{Approval}_t = \beta_0 + \beta_1 \text{Approval}_{t-1} + \beta_2 \text{Price}_t + u_t
\end{align}`
$$

.footnote[
.pink[†] [Fun with approval ratings](https://projects.fivethirtyeight.com/trump-approval-ratings/).
]

*Note:* We're ignoring any other violations of exogeneity for the moment.

---
exclude: true

---
layout: false
class: clear

.hi-slate[Monthly presidential approval ratings, 2001–2006]
<img src="slides_files/figure-html/approval 1-1.svg" style="display: block; margin: auto;" />

---
class: clear

.hi-slate[Approval rating vs. its one-month lag, 2001–2006]
<img src="slides_files/figure-html/approval 2-1.svg" style="display: block; margin: auto;" />

---
class: clear

.hi-slate[Approval rating vs. its two-month lag, 2001–2006]
<img src="slides_files/figure-html/approval 3-1.svg" style="display: block; margin: auto;" />

---
class: clear

.hi-slate[Oil prices, 2001–2006]
<img src="slides_files/figure-html/approval 4-1.svg" style="display: block; margin: auto;" />

---
class: clear

.hi-slate[Approval rating vs. oil prices, 2001–2006]
<img src="slides_files/figure-html/approval 5-1.svg" style="display: block; margin: auto;" />

---
layout: true
# Testing for autocorrelation
## Example: Approval ratings and oil prices

---

**Step 1:** Estimate our ADL(1, 0) model with OLS.

``` r
# Estimate the model
ols_est <- lm(
 approve ~ lag(approve) + price_oil,
 data = approval_df
)
# Summary
tidy(ols_est)
```

```
#> # A tibble: 3 × 5
#> term estimate std.error statistic p.value
#> <chr> <dbl> <dbl> <dbl> <dbl>
#> 1 (Intercept) 16.2 7.86 2.06 4.40e- 2
#> 2 lag(approve) 0.841 0.0752 11.2 2.17e-16
#> 3 price_oil -0.0410 0.0215 -1.90 6.15e- 2
```

---

**Step 2:** Record residuals from the OLS regression.

``` r
# Grab residuals
approval_df$e <- c(NA, residuals(ols_est))
```
--

**Note:** We add an `NA` because we use a lag—the first element is missing.

*E.g.*,
 .mono[{1, 2, 3, 4, 5, 6, 7, 8, 9} = x]
--
 .mono[{.pink[?], 1, 2, 3, 4, 5, 6, 7, 8} = lag(x)]
--
 .mono[{.pink[?], .pink[?], 1, 2, 3, 4, 5, 6, 7} = lag(x, 2)]
--
 .mono[{.pink[?], .pink[?], .pink[?], 1, 2, 3, 4, 5, 6} = lag(x, 3)]

---

---

**Step 3:** Regress residuals on an intercept, the explanatory variables, and lagged residuals.

``` r
# BG regression
bg_reg <- lm(
 e ~ lag(approve) + price_oil + lag(e) + lag(e, 2),
 data = approval_df
)
```

```
#>              Estimate Std. Error t value Pr(>|t|)  
#>  (Intercept)   7.92474    9.30455   0.852   0.3979  
#>  lag(approve) -0.08503    0.09192  -0.925   0.3589  
#>  price_oil    -0.01690    0.02407  -0.702   0.4854  
#>  lag(e)        0.25236    0.14648   1.723   0.0903 .
#>  lag(e, 2)     0.07865    0.14471   0.544   0.5889
```

---

**Step 4:** `$F$` (or `$\text{LM}$`) test for `$\rho_1=\rho_2=0$`.

*Recall:* We can test joint significance using an `$F$` test that compares the restricted (here: `$\rho_1=\rho_2=0$`) and unrestricted models.
$$
`\begin{align}
  F_{q,\,n-p} = \dfrac{\left(\text{SSE}_r - \text{SSE}_u\right)\big/q}{\text{SSE}_u\big/\left( n-p \right)}
\end{align}`
$$
where `$q$` is the number of restrictions and `$p$` is the number of parameters in our unrestricted model (include the intercept).

We can use the `waldtest()` function from the `lmtest` package for this test.

---

**Step 4:** `$F$` (or `$\text{LM}$`) test for `$\rho_1=\rho_2=0$`.

``` r
# BG regression
bg_reg <- lm(
 e ~ lag(approve) + price_oil + lag(e) + lag(e, 2),
 data = approval_df
)
# Test significance of the lags using 'waldtest' from 'lmtest' package
p_load(lmtest)
waldtest(bg_reg, c("lag(e)", "lag(e, 2)"))
```
Here, we're telling `waldtest` to test
- the model we specified in `bg_reg` (our .hi[unrestricted model])
- against a model without `lag(e)` and `lag(e, 2)` (our .hi[restricted model])

---
count: false

**Step 4:** `$F$` (or `$\text{LM}$`) test for `$\rho_1=\rho_2=0$`.

```
#> Wald test
#> 
#> Model 1: e ~ lag(approve) + price_oil + lag(e) + lag(e, 2)
#> Model 2: e ~ lag(approve) + price_oil
#>   Res.Df Df      F Pr(>F)
#> 1     57                 
#> 2     59 -2 1.6153 0.2078
```

---

**Step 5:** Conclusion of hypothesis test

With a *p*-value of ∼0.208, .hi[we fail to reject the null hypothesis.]

- We cannot reject `$\rho_1=\rho_2=0$`.

- We cannot reject "no autocorrelation".

However, .hi[we tested for a specific type of autocorrelation]: AR(2).

We might get different answers with different tests.

The *p*-value for AR(1) is 0.0896—suggestive of first-order autocorrelation.

---
layout: false
class: inverse, middle
# Living with autocorrelation

---
layout: true
# Autocorrelation
## Working with it

---
name: fixes

Suppose we believe autocorrelation is present. What do we do?

I'll give you three options..pink[†]

1. .hi[Misspecification]

1. .hi[Serial-correlation robust standard errors] (a.k.a. *Newey-West*)

1. .hi[FGLS]

.footnote[
.pink[†] You should take .mono[EC 422] to go much deeper into time-series analysis/forecasting.
]

---
layout: true
# Autocorrelation
## Option 1: Misspecification

---
name: misspecification

.hi[Misspecification] with autocorrelation is very similar to our discussion with heteroskedasticity.

By incorrectly specifying your model, you can create autocorrelation.

Omitting variables that are correlated through time will cause your disturbances to be correlated through time.

---

**Example:** Suppose births depend upon income and previous births
$$
`\begin{align}
  \text{Births}_t = \beta_0 + \beta_1 \text{Births}_{t-1} + \beta_2 \text{Income}_t + u_t
\end{align}`
$$
--
but we write down the model as only depending upon previous births, _i.e._,
$$
`\begin{align}
  \text{Births}_t = \beta_0 + \beta_1 \text{Births}_{t-1} + v_t
\end{align}`
$$
--
Then our disturbance `$v_t$` is
$$
`\begin{align}
  v_t = \beta_2 \text{Income}_t + u_t
\end{align}`
$$
--
which is likely autocorrelated, since income is correlated in time.

*Note:* This autocorrelation has nothing to do with `$u_t$`.

---

**"Proof"**

$$
`\begin{align}
  v_{t} &= \beta_2 \text{Income}_{t} + u_{t} \\
  v_{t-1} &= \beta_2 \text{Income}_{t-1} + u_{t-1}
\end{align}`
$$

`$\mathop{\text{Cov}} \left( v_{t},\,v_{t-1} \right)$`

.pad-left[
`$= \mathop{\text{Cov}} \left( \beta_2 \text{Income}_{t} + u_{t},\, \beta_2 \text{Income}_{t-1} + u_{t-1} \right)$`
]

.pad-left[
`$= \color{#e64173}{\mathop{\text{Cov}} \left( \beta_2 \text{Income}_{t},\, \beta_2 \text{Income}_{t-1} \right)} + \mathop{\text{Cov}} \left( \beta_2 \text{Income}_{t},\, u_t \right)$`
 `$\color{#ffffff}{=} + \mathop{\text{Cov}} \left(u_{t},\, \beta_2 \text{Income}_{t-1}\right) + \mathop{\text{Cov}} \left( u_{t},\, u_{t-1} \right)$`
]

.pad-left[
`$\neq 0$` (in general) even if `$u_t$` is exogenous and without autocorrelation.
]

---
layout: true
# Autocorrelation
## Option 2: Newey-West standard errors

---
name: newey

As was also the case with heteroskedasticity, you can still estimate consistent standard errors (and inference) in the presence of autocorrelation.

These standard errors are called *serial-correlation robust standard errors* (or *Newey-West standard errors*).

We are not going to derive the estimator for these standard errors.

---
layout: true
# Autocorrelation
## Option 3: FGLS

---
name: gls

If we do not have a lagged outcome variable, then .pink[feasible generalized least squares] (.hi[FGLS]) can give us efficient and consistent standard errors.

Let's start with a simple static model that includes an AR(1) disturbance `$u_t$`.

$$
`\begin{align}
  \text{Births}_t &= \beta_0 + \beta_1 \text{Income}_t + u_t \tag{1} \\
  u_t &= \rho u_{t-1} + \varepsilon_t \tag{2}
\end{align}`
$$

Now our old trick: Write out `$(1)$` for period `$t-1$` (and then multiple by `$\rho$`)
$$
`\begin{align}
  \text{Births}_{t-1} &= \beta_0 + \beta_1 \text{Income}_{t-1} + u_{t-1} \tag{3} \\
  \rho \text{Births}_{t-1} &= \rho \beta_0 + \rho \beta_1 \text{Income}_{t-1} + \rho u_{t-1} \tag{4}
\end{align}`
$$

And now subtract `$(4)$` from `$(1)$`…

---

$$
`\begin{align}
  \text{Births}_t - \rho \text{Births}_{t-1} =& \beta_0 \left( 1 - \rho \right) + \\
  & \beta_1 \text{Income}_t - \rho \beta_1 \text{Income}_{t-1} + \\
  & u_t - \rho u_{t-1}
\end{align}`
$$
--
which gives us a very specific dynamic model
$$
`\begin{align}
  \text{Births}_t =& \beta_0 \left( 1 - \rho \right) + \rho \text{Births}_{t-1} +\\
  & \beta_1 \text{Income}_t - \rho \beta_1 \text{Income}_{t-1} + \\
  & \color{#e64173}{\underbrace{u_t - \rho u_{t-1}}_{=\varepsilon_t}} \\
  \color{#ffffff}{=}& \color{#ffffff}{\beta_0 \left( 1 - \rho \right) + \rho \text{Births}_{t-1} +}\\
  &\color{#ffffff}{ \beta_1 \text{Income}_t - \rho \beta_1 \text{Income}_{t-1} + \varepsilon_t}
\end{align}`
$$

---
count: false

$$
`\begin{align}
  \text{Births}_t - \rho \text{Births}_{t-1} =& \beta_0 \left( 1 - \rho \right) + \\
  & \beta_1 \text{Income}_t - \rho \beta_1 \text{Income}_{t-1} + \\
  & u_t - \rho u_{t-1}
\end{align}`
$$
which gives us a very specific dynamic model
$$
`\begin{align}
  \text{Births}_t =& \beta_0 \left( 1 - \rho \right) + \rho \text{Births}_{t-1} +\\
  & \beta_1 \text{Income}_t - \rho \beta_1 \text{Income}_{t-1} + \\
  & \color{#e64173}{\underbrace{u_t - \rho u_{t-1}}_{=\varepsilon_t}} \\
  =& \beta_0 \left( 1 - \rho \right) + \rho \text{Births}_{t-1} +\\
  & \beta_1 \text{Income}_t - \rho \beta_1 \text{Income}_{t-1} + \color{#e64173}{\varepsilon_t}
\end{align}`
$$
that happens to be .hi[free of autocorrelation].

---

This .hi[transformed model] is free of autocorrelation.
$$
`\begin{align}
  \text{Births}_t =& \beta_0 \left( 1 - \rho \right) + \rho \text{Births}_{t-1} +\\
  & \beta_1 \text{Income}_t - \rho \beta_1 \text{Income}_{t-1} + \color{#e64173}{\varepsilon_t}
\end{align}`
$$

**Q:** How do we actually estimate this model? (We don't know `$\rho$`.)
--
 **A:** FGLS (of course)…

.pseudocode-small[
1. Estimate the original (untransformed) model; save residuals.
2. Estimate `$\rho$`: Regress residuals on their lags (no intercept).
3. Estimate the **transformed model**, plugging in `$\hat{\rho}$` for `$\rho$`.
]

---
layout: false
# Table of contents

.pull-left[
### Admin
.smallest[

1. [Schedule](#schedule)
1. [.mono[R] showcase](#r_showcase)
  - [.mono[ggplot2]](#ggplot2)
  - [Writing functions](#functions)
1. [Review: Time series](#review)
]
]

.pull-right[
### Autocorrelation
.smallest[

1. [Introduction](#intro)
1. [In static models](#static_models)
1. [OLS and bias/consistency](#ols_static)
  - [Static models](#ols_static)
  - [Dynamic models with lagged `$y$`](#ols_adl)
1. [Simulation: Bias](#ar2_simulation)
1. [Testing for autocorrelation](#testing)
  - [Static models](#testing_static)
  - [Dynamic models with lagged `$y$`](#testing_adl)
1. [Working with autocorrelation](#fixes)
  - [Misspecification](#misspecification)
  - [Newey-West standard errors](#newey)
  - [FGLS](#gls)
]
]

---
exclude: true