DID model for non-panel data

[Brahim KHOUILED]

Hi everyone,

I want to apply the DID model to non-panel data, given the following issue:
What is the contribution of agricultural policy to achieving food security in the country?

I have a dependent variable representing food security y_t,
and independent food variables x_ti,
and other non-food variables z_ti that are not affected by agricultural policies.

I encountered a problem in structuring the data, which does not include cross-section. On the other hand, I have the policy and a set of variables affected by the policy and a second set that is not affected by it.

Please help, and thank you in advance.

[George Ford]

pooled cross section?

are some units receiving the policy treatment and others not?

[Brahim KHOUILED]

Thank you for response,

No, there are no units,

Yes, Yi and Xi is affected by politics, and Zi is not,

The data structure below, is there a command in STATA for this structure? Or do you estimate the DID manually?

YEARS	POLICY	Y	X1	X2	X3	X4	X5	X6	Z1	Z2	Z3
1990	0	216	2,6	1391	5334	9,02	116777	7700	34,944	26,76994	1,6
1991	0	184	3,08	1096	4647	9,8	191500	7950	34,944	24,30406	1,464423
1992	0	210	4,41	1128	5441	12,88	123000	8116	33,389	22,2882	1,337898
1993	0	203,68	4,15	1193	5414	15,96	106335	8095,07	32,836	20,93612	1,229477
1994	0	148,76	3,31	1023	4045	22,16	102381	8069,26	32,382	21,85831	1,148209
1995	0	153,99	3,26	1084	4280	23,2	98995	8070	30,587	22,0412	1,103148
1996	0	173,63	4,26	1154	4729	28	100902	8081	32,199	21,24599	1,103345
1997	0	148,6	4,2	1180	4519	28,51	101897	8202,72	33,295	20,42303	1,157851
1998	0	179,93	2,31	1200	5267	36,02	100886	8215	32,461	24,68631	1,275718
1999	0	164,7	3,92	1250	5075	42,34	101472	8227	31,52	24,17799	1,465998
2000	1	140,41	4	1288	4599	43,53	102550	8226	30,789	25,49091	1,737741
2001	1	162,57	5,02	2591	5488	56,76	101578	8169,23	31,677	27,09684	2,1
2002	1	167,78	4,3	2660	5209	75,45	105398	8205,05	32,087	29,40007	2,557947
2003	1	208,51	3,64	2729	6589	94,21	108142	8160	32,827	30,84067	3,101237
2004	1	248,58	3,22	1617	8032	85,11	106166	8196,82	35,219	33,52381	3,715646
2005	1	242,25	3,13	1381	7866	90,05	112474	8389,64	36,01	36,17219	4,386951
2006	1	261,31	3,04	1610	8812	112,92	114781	8414,67	37,051	38,66807	5,100928
2007	1	290,37	2,97	2220	10105	201,04	116112	8414,67	36,179	42,57355	5,843352
2008	1	321,75	3,27	2244	11197	308,56	117179	8414,67	36,975	47,76752	6,6
2009	1	365,24	2,86	2358	12775	393,75	120306	8423,34	37,221	53,92953	6,84593
2010	1	382,6	2,66	2420	13644	335,59	113993	8435,03	37,556	58,18996	6,858818
2011	1	442,43	2,66	2443	16111	392,44	111498	8445,49	37,686	59,29557	6,6
2012	1	488,97	3,02	2477	18334	301,26	113156	9032,7	37,393	64,69147	6,059339
2013	1	537,24	4,76	2529	20573	129,61	115408	8461,87	38,958	70,31963	5,340795
2014	1	556,12	5,26	2551	21967	203,52	118336	8465,04	36,58	76,29679	4,576853
2015	1	493,41	6,33	4960	19718	315,96	120046	8488,03	37,033	80,26423	3,9
2016	1	478,78	8,49	2545	19477	271,43	121133	8494,57	37,6	82,51162	3,41454
2017	1	492,92	7,32	2609	20565	151,65	122343	8534,6	36,805	83,83181	3,112058
2018	1	487,55	12,02	2649	20769	157,77	122089	5522,41	36,706	85,59228	2,955959
2019	1	588,16	13,42	2694	25291	235,6	123069	5724,78	36,642	85,07873	2,909647
2020	1	469,07	16,6	2650	20756,16	209,53	123864	7240,49	33,893	78,95306	2,936526
2021	1	450,54	14,2	2650	19903,89	214,97	124659	8509,57	34,39	79,26887	3

[Clyde Schechter]

Based on what you show, there is no DID estimation for this data, neither in Stata, nor manually. What you show here is nothing more than a single time series of data with onset of a policy in year 2000. The best you can do with this data is compare the pre- and post- distributions of the outcome variable, with adjustment for the X's and Z's. This is a very weak model for estimating causal effects, though I suppose it is better than nothing.

In order to do a DID estimation you would need data that included multiple countries, some of which never adopted the policy in question. Or multiple countries, all of which adopted the policy in question but at different times. But without that, you are left with what you have: a single time series and a pre-post comparison. In the end, you will be unable to defend any findings you make from this data against the alternative that something other than the policy adoption happened in this country also happened in 2000 and accounts for any observed "effect" of the policy.

[Brahim KHOUILED]

Thank you Clyde!

In fact, I have a good understanding of the DID model and have already applied it using Stata and other software. The data I presented can be used in a regression model with interactive variables to reveal the impact of policy on various types of variables.

So, I raised the issue to see if there might be a way to apply DID to non-panel data. As an initial idea, I considered longitudinally aggregating the data and classifying it into two groups: affected and unaffected by the policy (treatement). Additionally, I thought about coding the variables as if they were units, provided they have the same units of measurement, such as percentages.

What do you think of this method? Or do you have any similar suggestions? Always within the framework of DID models.

[Clyde Schechter]

No, there is no way to apply the DID framework to this data. DID is the initialism for Difference In Differences, which is the short form of Between Group Difference of Pre-Post Differences. That the last Differences is plural is not some syntactic oddity of English. It really means it: you have to have more than one Group, and then you calculate the pre-post difference in each group, and then calculate the difference between those differences. A simple time series has only one group--there are no differences to take the difference of.

[Brahim KHOUILED]

Understood!

Meaning there is no hope with this data, a second group as a control group is necessary..

Thanks Clyde.