Explore model with state by year-month fixed effects

[santiagohermo]

Follow-up to https://github.com/diegogentilepassaro/min_wage_rent/issues/250#issuecomment-1500303715.

In this issue we will try to understand what drives the null results in the model that controls for state by year-month fixed effects. Things we want to look at controlling for state by year-month FE:

• How does the unbalanced model look like?
• How do pre-trends look like in this model?
• What happens if we include or drop controls?
• What happens if we drop states (CA)?
• What happens if we only use workplace MW?
• What if we create a dummy for having a statutory local MW in month t, and use this dummy interacted with time instead of the MW measures? Do we see that these places are in a different trend?

We need a good response on why this model looks different for the new submission.

[santiagohermo]

Some replies to the points in the opening comment

---

• How does the unbalanced model look like?

Results of estimating unbalanced models:

	baseline	state-time	cbsa-time	county-time
D.mw_wkp_tot_17	0.0479	-0.00750	0.0723	0.0645
	(1.58)	(-0.32)	(1.25)	(1.35)
D.mw_res	0.0263	-0.00271	-0.0408	-0.0304
	(-1.23)	(-0.14)	(-1.65)	(-0.94)
N	193292	192863	188048	183103

Note: t-statistics are in parentheses. Models include controls.

Code

```STATA cd C:\Users\shermo\Documents\GitHub\min_wage_rent\analysis\fd_test\code local instub "../../../drive/derived_large/estimation_samples" #delimit ; local keep_vars "zipcode_num year_month fullbal_sample_SFCC county_num cbsa_num statefips_num ln_rents mw_wkp_tot_17 mw_res ln_emp* ln_estc* ln_avg* "; #delimit cr use `keep_vars' using `instub'/zipcode_months.dta, clear xtset zipcode_num year_month reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_avg* ln_est*), /// absorb(year_month) nocons /// cluster(statefips_num) eststo baseline reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_avg* ln_est*), /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo state_time reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_avg* ln_est*), /// absorb(cbsa_num##year_month) nocons /// cluster(statefips_num) eststo cbsa_time reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_avg* ln_est*), /// absorb(county_num##year_month) nocons /// cluster(statefips_num) eststo county_time esttab *, keep(D.mw_wkp_tot_17 D.mw_res) ```

---

• How do pre-trends look like in this model?

Figures from https://github.com/diegogentilepassaro/min_wage_rent/commit/17c240c2e8b1a3ade5c7534714e661d75b358cbf

See figures

| baseline | state-time | | --- | --- | |  |  | | cbsa-time | county-time | | --- | --- | |  |  |

---

• What happens if we include or drop controls?

Not much changes.

Results of estimating unbalanced models:

	No ctrls	Empl. ctrls	Est. count ctrls.	Avg. wag ctrls	All controls
D.mw_wkp_tot_17	-0.00757	-0.0114	-0.00711	-0.0104	-0.0103
	(-0.14)	(-0.22)	(-0.13)	(-0.20)	(-0.20)
D.mw_res	0.0142	0.0152	0.0137	0.0148	0.0141
	(0.59)	(0.64)	(0.58)	(0.60)	(0.58)
N	80393	79928	80393	79928	79928

Note: t-statistics are in parentheses.

Code

```STATA clear all cd C:\Users\shermo\Documents\GitHub\min_wage_rent\analysis\fd_test\code local instub "../../../drive/derived_large/estimation_samples" #delimit ; local keep_vars "zipcode_num year_month fullbal_sample_SFCC county_num cbsa_num statefips_num ln_rents mw_wkp_tot_17 mw_res ln_emp* ln_estc* ln_avg* "; #delimit cr use `keep_vars' using `instub'/zipcode_months.dta, clear xtset zipcode_num year_month reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo no_controls reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo emp_controls reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_est*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo estcount_controls reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_avg*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo avgwage_controls reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo all_controls esttab *, keep(D.mw_wkp_tot_17 D.mw_res) ```

---

• What happens if we drop states (CA)?

The states with the most observations in the fullbalanced sample are California (fips: 6, % of zip-months: 11.6), Florida (fips: 12, % of zip-momths: 20.59), Texas (fips: 48, % of zip-months: 12.94).

Results look like this:

	Baseline	Excl. CA	Excl. FL	Excl. TX
D.mw_wkp_tot_17	-0.0103	0.0696	-0.0298	-0.00985
	(-0.20)	(0.85)	(-0.58)	(-0.19)
D.mw_res	0.0141	-0.00438	0.0284	0.0130
	(0.58)	(-0.12)	(1.11)	(0.54)
N	79928	70581	63316	69488

Note: t-statistics are in parentheses.

Code

```stata clear all cd C:\Users\shermo\Documents\GitHub\min_wage_rent\analysis\fd_test\code local instub "../../../drive/derived_large/estimation_samples" #delimit ; local keep_vars "zipcode_num year_month fullbal_sample_SFCC county_num cbsa_num statefips_num ln_rents mw_wkp_tot_17 mw_res ln_emp* ln_estc* ln_avg* "; #delimit cr use `keep_vars' using `instub'/zipcode_months.dta, clear xtset zipcode_num year_month reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo baseline reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1 & statefips_num != 6, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo no_CA reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1 & statefips_num != 12, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo no_FL reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1 & statefips_num != 48, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo no_TX esttab *, keep(D.mw_wkp_tot_17 D.mw_res) ```

---

• What happens if we only use workplace MW?

Things remain mostly insignificant:

	Both	Only wkp	Only res
D.mwwkp~17	-0.0103	0.00961
	(-0.20)	(0.40)
D.mw_res	0.0141		0.00853
	(0.58)		(0.78)
N	79928	79928	79928

Note: t-statistics are in parentheses.

Code

```stata clear all cd C:\Users\shermo\Documents\GitHub\min_wage_rent\analysis\fd_test\code local instub "../../../drive/derived_large/estimation_samples" #delimit ; local keep_vars "zipcode_num year_month fullbal_sample_SFCC county_num cbsa_num statefips_num ln_rents mw_wkp_tot_17 mw_res ln_emp* ln_estc* ln_avg* "; #delimit cr use `keep_vars' using `instub'/zipcode_months.dta, clear xtset zipcode_num year_month reghdfe D.ln_rents D.mw_wkp_tot_17 D.mw_res /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo baseline reghdfe D.ln_rents D.mw_wkp_tot_17 /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo only_wkp reghdfe D.ln_rents D.mw_res /// D.(ln_emp* ln_est* ln_avg*) /// if fullbal_sample_SFCC == 1, /// absorb(statefips_num##year_month) nocons /// cluster(statefips_num) eststo only_res esttab *, keep(D.mw_wkp_tot_17 D.mw_res) ```

[santiagohermo]

I responded the questions in the opening comment @diegogentilepassaro @gabrieleborg. See https://github.com/diegogentilepassaro/min_wage_rent/issues/257#issuecomment-1509863148

I also formally tested whether the model with state-time FE is inconsistent with the baseline. The final run with the results is in commit https://github.com/diegogentilepassaro/min_wage_rent/commit/c37befef9622cff59a66ee0d7bf4bafa0e740bb1. In particular, see this table.

[santiagohermo]

• Add coefficients of models to test table
• Search for a cite to back up the estimation method

[santiagohermo]

Continues in its PR #259