t-test: Development of propaganda and military before and after events, excluding event day

[yulinHAO1995]

categorize the events on November 13 and 14 as signaling weakness, and then we should test whether they are related to less activity (or at least awareness of activity), as the data suggest so for. o The t-tests as they are, exclude the day of the event itself.

[yulinHAO1995]

finished

[KaliGi]

What is the ethnicity control category? What do you do there exactly?

[KaliGi]

Any idea why the province FE currently turns results around? Does a district FE not work here?

[KaliGi]

We have to think whether we need a district FE here. I think we don't, as we do not make a cross-district comparison. But maybe we should add more controls, like initial IS presence in 2015 (the aggregated presence variable)

[KaliGi]

Can we do the same test for the two strength events pooled together? And then see whether we find a similar result when we use all strength events, even when there was no google search spike. When we pool, we need to add wave FE. Same we could do, in addition to existing tables, for weakness events.

[yulinHAO1995]

What is the ethnicity control category? What do you do there exactly?

There is a survey question about respondents' ethnic group. There are over 30 ethnic groups but around 85% people fall into 4 groups: Pashtun,Tajik,Uzbek, Hazara. I categorized other minorities as other groups, so ethnicity dummy variable have 5 categories: Pashtun,Tajik,Uzbek, Hazara, other groups.

[yulinHAO1995]

Any idea why the province FE currently turns results around? Does a district FE not work here?

I did not use district FE and I will check whether there would be different patterns.

[KaliGi]

Do we have results on that?

[yulinHAO1995]

I have tried the following specifications:

1. Use during event as reference group, include district FE insead of province FE, getting similar results as before.
2. Use during event as reference group, include district FE, pooled events signaling strength(wave 27 and wave 32), getting inconsistent and undesiable results(the coefficients of before event and after events are both negative).
3. Only include after event(exclude before event), include initial level of IS presence, getting relatively interpretable results. I present the result in the pdf.

[yulinHAO1995]

If needed, I can present all the other results.

[KaliGi]

One question: In t-test_before_after_event, exclude event day.pdf and the other one, how do we restrict the sample? How many days/month before and after the event do we include? Are the observations individuals? We should vary that, to make sure the events do not overlap with other events before or after. We could also run it at the district level, and just compute the change between 30 days before and 30 days after (varying wether to include event date or not). Would be good to find a specification that shows difference between weakness and strength event.

[yulinHAO1995]

1. In the test, the observations are individuals. I include obs in one survey wave. For example, if the event took place in wave 27, I only kept obs in wave 27.
2. I will run it at district level and compute the change between 30 days before and 30 days after.

[KaliGi]

Thanks, I will reply on Monday.

Am 11. Oktober 2019 05:01:01 GMT-07:00 schrieb Yulin Hao notifications@github.com:

1. In the test, the observations are individuals. I include obs in one survey wave. For example, if the event took place in wave 27, I only kept obs in wave 27.
2. I will run it at district level and compute the change between 30 days before and 30 days after.

-- You are receiving this because you commented. Reply to this email directly or view it on GitHub: https://github.com/yulinHAO1995/Afghanistan-Insurgent-competition/issues/1#issuecomment-541035111

-- Diese Nachricht wurde von meinem Android-Gerät mit K-9 Mail gesendet.

[yulinHAO1995]

I runned regressions at district level and computed the change between 30 days before and 30 days after.

1. Without controlling for Initial IS presence in 2015, weakness and strength event show similar pattern. Less propaganda and military after strength or weakness event(undesirable results for strength event, maybe we should find more strength events?).
2. After including Initial IS presence in 2015, weakness and strength event show different pattern. Less propaganda and more military after strength event, and the opposite after weakness event.

[KaliGi]

Thank you. I think these regressions should include a district FE. We only want to compare the before and after (so far), no difference between districts. So we should erase the differences. Once we do that, we should check again. But generally, I agree with you. We have results that are quite diverging and are hard to interpret. Let's wait whether district FE make a difference. The other idea I have is to use the regressions that control for initital IS presence, and interact initial presence with before/after dummy. The idea is to see whether their tactics change differentially in districts in which they are already active and those that they do not control.

[KaliGi]

I might have suggested this before, but to make things a bit easier to overlook, can you please aggregate the most interesting results in one PDF file in the main "\Dropbox\Afghanistan Insurgent competition\Paper insurgent Competition" folder? You can use the template there, just create a file called "appendix 2019_10_14" which we can then relabel every day to keep track of versions. We have an important skype group talk this week, and would like to discuss your interesting results. For now I would like you to include the results from:

• [ ] Correlation between social-economic needs and IS presence: But please add district and time FE, and cluster at district level.

• [ ] IS_between and within variation

• [ ] ISIS_Sum_Stats

• [ ] The new version of the event tables as described above

• [ ] Table 1 from reg_How government military presence affects IS actions (2), and Table 2 but controlling for province times wave in all regressions please

• [ ] Same tables and changes from "reg_How Taliban presence affects IS intensity (2).pdf"

• [ ] Two new tables that control for governmetn and Taliban presence in t-1 in one regressions. With all three sets of FE again please, incl provXwave.

• [ ] Question: Why are in "stat_IS_presence.pdf" the number of obs 87000 and in "ISIS_Sum_Stats.pdf" sometimes over 117000? Generally both seems useful, but please explain how they differ or which one we should rely on.

• [ ] Opium profitability seems extremly dependant on province-times-wave FE, even change signs. Any idea why? How do the results look like when you run the same regressions aggregated at the province level? Still negative? THen it would mean within provinces they select the ones more suitable for opium, but generally those provinces less suitable. Can you run those with also including government and military presence? And please check with Sarah so she can tell you which time-invariant control variables you can include as controls (interacted with period FE).

• [ ] "propaganda2_cont.pdf" , "propaganda2_75_all.pdf",military_cont.pdf, military2_75.pdf,gov_military_presence.pdf,

• [ ] Check if there are old maps and old map folders we might want to delete.

Thank you!

[yulinHAO1995]

1. Regarding events regression: most of individuals were interviewed before or after 1 week of the event, taking the event took place in 27feb2015 (wave 27) as example, . If treating 1 month before the event as reference category and analyzing at district level, few observations left (< num of district, unable include district FE). I therefore regress at individual level and treat 3 months before the event as reference category(including some obs from wave 26 and 28), incluing district FE and district FE* initial IS level. The event signaling strength now has positive and significant effect. But events signaling weakness has undesiable positive sign, although not signicant overall. Using the same specification, regress at district level doesnot help.

[yulinHAO1995]

2. ISIS_Sum_Stats.pdf was not created by me and I haven't found related tex file so I incuded stat_IS_presence instead ISIS_Sum_Stats in the aggregate file. I have no ideal why there are more obs in ISIS_Sum_Stats. Stat_IS_presence was created using ANQAR_panel_districtwave.dta and ANQAR_panel_ind.dta, the two main datasets I work on.

[yulinHAO1995]

3. why Opium profitability seems extremly dependant on province-times-wave FE? I am not sure whether it is related with the variation pattern of opoium profitability. It shows little within variation. Taking its value in district 101 as example, if we regress IS presence on lag_opoium profitability, only information from wave 27-wave 31 can be ued for estimation. No variation from wave 28-wave 31. I don't know what would happen when including Province*wave FE?

[yulinHAO1995]

The following graph shows the variation patterns of our variables. Taking district 102 as example, Phone from NRVA, rural from AF. propganda from ANQAR. suitability_rw_opium is not indicated.

[yulinHAO1995]

Combined results are saved in: ..\Paper insurgent Competition\combine_results_2019_10_18