Update README.md

README.md

@@ -1,3 +1,76 @@
----
-license: mit
----
+---
+license: mit
+---
+
+# Survey on Social Issues in the United States (2016 Election Study)
+
+## Overview
+This data product contains individual-level responses to an online survey experiment conducted by **Connor Jerzak and co-authors** in the run-up to—and immediately after—the 2016 U.S. presidential election. Fieldwork began on **12 September 2016** and continued through **mid-November 2016**, giving researchers a rare before/after snapshot of attitudes shaped by a highly salient national campaign.
+
+**Key design features**
+
+* **Crime-framing experiment.** Respondents read a mock police-blotter story with experimentally varied details (suspect race, number of break-ins, presence/absence of racial information) before answering questions about policy, crime perceptions, and social spending.  
+* **Rich demographics & ideology.** Over 50 items capture party identification, vote choice, income, employment security, family status, education, racial identity, and core value trade-offs.  
+* **Panel structure.** A subset of respondents was re-contacted after Election Day, enabling within-person analyses of opinion change.
+
+## File Manifest
+
+| File | Description |
+|------|-------------|
+| `survey_results.csv` | Clean, respondent-level dataset (wide format). Each column corresponds to a survey variable prefixed by its original Qualtrics question ID. |
+| `Oct21_survey.pdf`  | Archived survey instrument, including consent form and full questionnaire. |
+
+## Quick Start (R)
+
+```r
+library(tidyverse)
+df <- read_csv("survey_results.csv")
+
+# Recode experimental treatment
+#   Q42 == "No" → Control
+#   Q42 == "Yes" & Q43 gives race
+df <- df %>% 
+  mutate(treat = case_when(
+    Q42 == "No" ~ "Control",
+    Q43 == "Black" ~ "Black",
+    Q43 == "White" ~ "White"
+  ))
+
+# Estimate effect of racial cue on support for longer sentences
+lm(long_sentences ~ treat + party_id + age, data = df)
+```
+
+## Variable Highlights
+
+* **Safety perceptions:** `Q2`–`Q4`, `Q37`, `Q39`  
+* **Crime policy preferences:** `Q11`, `Q12`  
+* **Redistribution & welfare attitudes:** `Q8`, `Q9`, `Q46`–`Q51`  
+* **2016 vote intention & choice:** `Q41`, `Q44`, `Q45`  
+* **Economic security:** `Q29`–`Q32`  
+* **Child-rearing values:** `Q33`–`Q36`
+
+See `Oct21_survey.pdf` for exact wording and response options.
+
+## Possible Use Cases
+
+1. **Election-season opinion dynamics** – exploit the before/after panel to examine how campaign events (debates, the Comey letter, Election Day) shifted perceptions of crime, policing, or redistribution.  
+2. **Stereotype activation & policy support** – estimate causal effects of suspect-race cues on punitive crime policies or welfare attitudes.  
+3. **Replication exercises** – reproduce classic findings from ANES or GSS items using a contemporary MTurk sample; ideal for teaching regression, causal inference, or text analysis (e.g., coding open-ended crime causes in `Q10`).  
+4. **Value trade-off scaling** – model latent moral or parenting value dimensions with the paired choice items (`Q33`–`Q36`).  
+5. **Small-N machine-learning demos** – demonstrate text classification, topic modeling, or mixed-effects models on a manageable survey.
+
+## Sampling & Fieldwork
+
+Respondents were recruited via **Amazon Mechanical Turk**. Each wave paid \$0.25 and took ~5 minutes. The instrument included an informed-consent screen and was approved by the Harvard CUHS IRB. IP geo-coordinates (rounded to 3 decimals) were recorded for coarse location checks; no personally identifying information is included.
+
+| Wave | Dates | N (unique) | Notes |
+|------|-------|------------|-------|
+| Pre-Election | 12 Sep – 04 Nov 2016 | ~1,200 | Prior to Election Day |
+| Post-Election | 09 Nov – 15 Nov 2016 | 420 | Post Election Dady |
+
+## Data Quality Notes
+
+* **Non-probability sample.** MTurk respondents skew younger, more educated, and more politically engaged than the general U.S. adult population.  
+* **Attention checks.** Various items (e.g., number of break-ins retention check) facilitate quality screening.  
+* **Missing values.** Skipped or invalid responses are coded `NA`.
+