Im currently trying to work through a question using synthetic control in r. link for context here: https://economics.mit.edu/files/17847
Ive found that the documentation is lacking and I have a syntax error. Has anyone put together a file to use synthetic control in julia?
I have the start of a package here https://github.com/nilshg/SynthControl.jl but I’ve shamefully neglected it over the last few months. It should work for getting a point estimate of the treatment effect out, but inference isn’t really operational (you can do permutation of placebos but have to process the results of the permutation yourself). Also it currently uses the pretreatment outcomes to find weights only, rather than Alfonso covariates like the originell Abadie et al paper.
Finally, the lady change I made was to spin out construction of the data object to go the estimation on into a separate TreatmentPanel type which I’m planning to use for other casual inference packages as well, and the readme doesn’t reflect this, so check the test file to see how you fit and plot the model.
First, thank you for putting this together.
but I’m having some trouble implementing it. I found the the test file and managed to get your example to work. However my code has an error that i dont understand. Could you help me?
julia> # format data so its easier to compare to example
formatdata = DataFrame(state = mydata.name, year = mydata.year , Y = mydata.Y )
809×3 DataFrame
Row │ state year Y
│ String Date Float64
─────┼────────────────────────────────
1 │ alabama 2000-01-01 5.27832
2 │ alabama 2001-01-01 7.63268
3 │ alabama 2002-01-01 6.51773
4 │ alabama 2003-01-01 6.55047
5 │ alabama 2004-01-01 5.58409
6 │ alabama 2005-01-01 7.15567
7 │ alabama 2006-01-01 7.73388
8 │ alabama 2007-01-01 8.3889
9 │ alabama 2008-01-01 7.24852
10 │ alabama 2009-01-01 6.68357
11 │ alabama 2010-01-01 3.99604
12 │ alabama 2012-01-01 0.041532
13 │ alabama 2013-01-01 0.0414039
14 │ alabama 2014-01-01 0.0206506
⋮ │ ⋮ ⋮ ⋮
797 │ wyoming 2005-01-01 3.30638
798 │ wyoming 2006-01-01 1.72194
799 │ wyoming 2007-01-01 3.36527
800 │ wyoming 2008-01-01 2.01449
801 │ wyoming 2009-01-01 2.14343
802 │ wyoming 2010-01-01 1.41938
803 │ wyoming 2011-01-01 3.52594
804 │ wyoming 2012-01-01 2.95001
805 │ wyoming 2013-01-01 2.92035
806 │ wyoming 2014-01-01 2.40324
807 │ wyoming 2015-01-01 3.07341
808 │ wyoming 2016-01-01 3.9364
809 │ wyoming 2017-01-01 2.41824
julia> # this is what i cant get to work
tp = TreatmentPanel("colorado" => Date(2014), formatdata; outcome = :Y, id_var = :state, t_var = :year)
ERROR: DimensionMismatch("new dimensions (44, 14) must be consistent with array size 615")
Stacktrace:
[1] (::Base.var"#throw_dmrsa#213")(::Tuple{Int64,Int64}, ::Int64) at ./reshapedarray.jl:41
[2] reshape at ./reshapedarray.jl:45 [inlined]
[3] reshape(::Array{Float64,1}, ::Int64, ::Int64) at ./reshapedarray.jl:116
[4] TreatmentPanel(::Pair{String,Date}, ::DataFrame; outcome::Symbol, id_var::Symbol, t_var::Symbol, predictors::Nothing) at /Users/modelt/.julia/packages/SynthControl/nHELX/src/treatmentpanel.jl:117
[5] top-level scope at REPL[10]:2
I wasnt sure if i needed to tag you in my reply
FWIW, I’ve never been able to use Synthetic Control in R, also encountering indecipherable errors, but I’ve found Stata’s package to work well, if you have access to Stata.
I do have access, i suppose it will be my last resort.
Can you share the data? I’d be happy to really a look tomorrow -the error suggests something is going wrong in forming the matrix of pretreatment outcomes, which uses reshape.
Off the top of my head, is your panel balanced? I haven’t looked at the code on a while but I think it currently assumes a balanced panel, mainly because I didn’t know how to treat missing observations for some i/t combinations in the minimization step. So one thing you can try is restricting your data to have the same number of pretreatment periods for all observations.
Here is a link to my data. GitHub - efhart4/Syntheticdata
If my treatment period is 2014, should i restrict data from 2010 to 2017 to balance the dataset?
Sorry I wasn’t exactly clear maybe - by balanced panel I mean that each unit of observation is observed in the same time periods. From your data I see that Alabama is missing one observation (all other states have 18 years in the data, Alabama only 17). The package internally builds a matrix of size (n_control_units, n_pretreatment_periods), which doesn’t work if control units don’t have the same number of pretreatment outcomes.
If you exclude Alabama it works:
julia> tp = TreatmentPanel("colorado" => 2014, df[df.name .!= "alabama", :], outcome = :Y, id_var = :name, t_var = :year)
TreatmentPanel
data: DataFrame
outcome: Symbol Y
id_var: Symbol name
t_var: Symbol year
treatment: Pair{Union{String, Symbol}, Union{Int64, Date}}
y₁₀: Array{Float64}((14,)) [3.09689037539627, 3.41189966665062, 4.03081592176743, 4.19543483694774, 4.7212980596995, 3.82133592176667, 3.72847941805215, 3.35146594369371, 3.27216431173091, 3.3184539222617, 2.88316462512099, 3.26059091669659, 2.98437286099889, 3.52908899034282]
y₁₁: Array{Float64}((4,)) [3.08341016942311, 3.4665497210528, 4.02460168625396, 4.14893279832875]
yⱼ₀: Array{Float64}((43, 14)) [7.46039306388848 3.64459424983568 … 1.75225436285053 2.78619300619598; 8.36260402766524 3.72130409380663 … 1.75987869949795 3.20970259891833; … ; 2.58698745311085 1.30826593198556 … 2.60933747015937 2.95000607354192; 3.43348512728692 2.50007197176888 … 3.11916245120324 2.92034501299554]
yⱼ₁: Array{Float64}((43, 4)) [4.64816508856759 4.63807710133236 2.25612353902884 8.56456115413491; 5.03395284122737 3.80565515825976 2.3988631474649 7.82536527615169; … ; 4.36500816978301 3.03268597015624 1.89306815234655 3.93640144448818; 3.4692846369428 3.33067760256769 6.86182014858432 2.41823765748773]
x₁₀: Nothing nothing
xⱼ₀: Nothing nothing
predictors: Nothing nothing
J: Int64 43
T₁: Int64 4
T₀: Int64 14
comp_labels: Array{Union{String, Symbol}}((43,))
julia> s = SynthControlModel(tp)
Synthetic Control Model
Outcome variable: Y
Time dimension: year with 18 unique values
Treatment period: 2014
ID variable: name with 44 unique values
Treatment ID: colorado
Model is not fitted
julia> SynthControl.fit!(s)
Synthetic Control Model
Outcome variable: Y
Time dimension: year with 18 unique values
Treatment period: 2014
ID variable: name with 44 unique values
Treatment ID: colorado
Model is fitted
Impact estimates: [-2.211, -2.506, -0.674, -0.451]
julia> using Plots
julia> plot(s)
produces:
Now just based on gut feel that looks like a slightly surprising result, the jaggedness of the pre-treatment fit suggests to me that this might not survive inference if I got it to work 
I also realize that the weigthplot that I thought I’d implemented seems not to work properly, but you can do
julia> weightdf = DataFrame(State = s.treatment_panel.comp_labels, Weight = s.w);
julia> bar(weightdf[weightdf.Weight .> 0.01, :State], weightdf[weightdf.Weight .> 0.01, :Weight], label = "Weight", xrot = 45)
to get
So the weights aren’t exactly sparse, which can be a problem if one fits purely based on pre-trends I believe. I really should get covariates to work to help with this, but I probably won’t have time over the next weeks unfortunately…
thank you
I wrote it up for r here
Hi @Evan, I wrote up Synthetic Control code in Julia to study the impact of foreign buyer taxes (15% in 2016, 20% in 2018) on house price growth in Vancouver.
I use Elastic Net (as argued in Doudchenko & Imbens), but you can easily modify my code to use your favorite ML model or ensemble…
I just created a Github repo w/ data & parsimonious code:
https://github.com/azev77/Synthetic_Control_in_Julia
Goodluck!
Below are the four figures produced by my code.
PS: I actually have code for at least 8 inference methods for SCM (all in Julia) stay tuned…