---
title: "Exercise template for 'Natural experiments'"
author: Your name
format: html
bibliography: book.bib
---

```{r}
#| message: false
library(farr)
library(ggplot2)
library(dplyr, warn.conflicts = FALSE)
```

```{r}
#| label: fig-days-to-filing
#| fig-height: 3
#| fig-cap: Distribution of days to filing in @Michels:2017uc sample
#| fig-alt: "Histogram of days to next filing in Michels (2017) sample shows a cluster of observations around 50-75 days, which generally fall into the 'recognize is TRUE' category because they occur after the filing for the previous quarter has occurred. Most observations with 'recognize is FALSE' category occur when fewer than 30 days remaining to the next filing, likely the filing for the previous quarter."
michels_2017 |> 
  mutate(days_to_filing = as.integer(date_filed - eventdate)) |>
  ggplot(aes(x = days_to_filing, fill = recognize)) + 
  geom_histogram(binwidth = 1) 
```

```{r}
#| label: fig-days-to-period-end
#| fig-height: 3
#| fig-cap: Distribution of days to period-end in @Michels:2017uc sample
#| fig-alt: "Histogram of days to period-end in Michels (2017) sample shows a spike (i.e., concentration of observations) at 32 days. There are also spikes at 14 and 17 days. These spikes likely reflect common events across observations. Observations with more than 50 days to the end of the period generally fall into the 'recognize is FALSE' category because they occur before the filing for the previous quarter has occurred. Most observations with 'recognize is TRUE' category occur within 50 days of the end of the period, as the next filing is the filing for the current period."
michels_2017 |> 
  mutate(days_to_period_end = as.integer(next_period_end - eventdate)) |>
  ggplot(aes(x = days_to_period_end, fill = recognize)) + 
  geom_histogram(binwidth = 1)
```

```{r}
#| label: tbl-top5
#| message: false
#| tbl-cap: Number of observations for top 5 natural disasters
#| render: !expr function(x, ...) knitr::knit_print(knitr::kable(x))
michels_2017 |> 
  count(eventdate) |>
  arrange(desc(n)) |>
  top_n(5)
```

```{r}
#| include: false
top5 <-
  michels_2017 |> 
  count(eventdate) |>
  arrange(desc(n)) |>
  top_n(5) |>
  pull() |>
  sum()
```

## Discussion questions

1. One assumption in @Michels:2017uc is that whether a natural disaster occurs before or after the balance sheet date of the next filing is random.
Do the inherent properties of natural disasters ensure that they are random?
Why?
If not, how would you evaluate the randomness of natural disasters in the sample of @Michels:2017uc?
Do the analyses above help this evaluation?

2. Describe what you imagine to be the process from the occurrence of a natural disaster to reporting on that event in the subsequent filing?
Do you think this process differs for recognized versus disclosed events?

3. From the analysis above, it appears that five natural disasters account for `r top5` observations.
A simple Google search for each date and the word "disaster" reveals that these events are Hurricane Katrina (`2005-08-29`), Hurricane Ike (`2008-09-13`), Hurricane Ivan (`2004-09-16`), Hurricane Charley (`2004-08-13`), and Hurricane Wilma (`2005-10-24`).
Is it problematic that a small number of disasters accounts for a significant portion of the sample?

4. Where does @Michels:2017uc get data on natural disasters from?
Is there anything that is problematic about this data source?
Would it be possible to use another approach to data collection?
What challenges would that approach face?

5. A recurring question in accounting is if it matters whether information is disclosed or recognized.
One view is that, if markets are efficient, it should not matter where the information is disclosed, so recognition should not matter relative to disclosure.
What assumptions underlie this view?
Are there any reasons to believe that they do or do not hold in the setting of @Michels:2017uc?
What are the implications for the ability of @Michels:2017uc to deliver clean causal inferences?
How might materiality criteria differ for disclosed and recognized events?
Would differences in these criteria affect the empirical analysis of @Michels:2017uc?

6. What causal inferences does @Michels:2017uc draw?
What (if any) issues do you see with regard to these?

7. Choose a paper that you have seen recently that uses empirical analysis of non-experimental data.
(If you cannot choose such a paper, @Hopkins:2018tg provide one option.)
Looking at the abstract of the paper, can you determine whether this paper seeks to draw causal inferences?

8. Choose what you think the authors regard to be the most important causal inference they draw (or would like to draw) in your chosen paper.
Which table or tables provide the relevant empirical analyses?
Sketch a rough causal diagram for this causal inference using either discussion in the paper or your own background knowledge to identify important variables.
How credible do you find the reported causal inferences to be in light of your causal diagram?