Dialysis

clustercov(x::Array{Real,2}, clusterid)

Compute clustered, heteroskedasticity robust covariance matrix estimate for Var(∑ x). Assumes that observations with different clusterid are independent, and observations with the same clusterid may be arbitrarily correlated. Uses number of observations - 1 as the degrees of freedom.

Arguments

• x number of observations by dimension of x matrix
• clusterid number of observations length vector

Returns

• V dimension of x by dimension of x matrix

downloadDFR(;redownload=false)

Downloads Dialysis Facility Reports. Saves zipfiles in Dialysis/data/.

errors_gm(y::Symbol, k::Symbol, l::Symbol, Φ::Symbol,
               id::Symbol, t::Symbol, data::DataFrame;
               npregress::Function=polyreg)

Returns functions that given β calculate ω(β) and η(β).

Arguments

• y Symbol for y variable in data
• k Symbol for k variable in data
• l Symbol for l variable in data
• q Symbol for q variable in data
• Φ Symbol for Φ variable in data
• id Symbol for id variable in data
• t Symbol for t variable in data
• data DataFrame containing variables
• α estimate of α

Returns

• ωfunc(β) computes ω given β for the data and α passed in as input. length(ωfunc(β)) == nrow(data) ωfunc(β) will contain missings if the data does.
• ηfunc(β) computes η given β. for the data and α passed in as input. length(ηfunc(β)) == nrow(data) ηfunc(β) will contain missings.

Warning: this function is not thread safe!

loadDFR(;recreate=false)

If Dialysis/data/dfr.zip exists, load it from disk. Otherwise, create Dialysis/data/dfr.zip exists by loading Dialysis Facility Reports from zipfiles in Dialysis/data/.

loaddata_old()

Loads "dialysisFacilityReports.rda". Returns a DataFrame.

locallinear(xpred::AbstractMatrix,
            xdata::AbstractMatrix,
            ydata::AbstractMatrix)

Computes local linear regression of ydata on xdata. Returns predicted y at x=xpred. Uses Scott's rule of thumb for the bandwidth and a Gaussian kernel. xdata should not include an intercept.

Arguments

• xpred x values to compute fitted y
• xdata observed x
• ydata observed y, must have size(y)[1] == size(xdata)[1]
• bandwidth_multiplier multiply Scott's rule of thumb bandwidth by this number

Returns

• Estimates of f(xpred)

objective_gm(y::Symbol,  k::Symbol, l::Symbol, q::Symbol,
                  Φ::Symbol, id::Symbol, t::Symbol,
                  instruments::Array{Symbol,1}, data::DataFrame
                  ; W=UniformScaling(1.),
                  npregress::Function=(xp,xd,yd)->polyreg(xp,xd,yd,degree=1))

panellag(x::Symbol, data::AbstractDataFrame, id::Symbol, t::Symbol,
          lags::Integer=1)

Create lags of variables in panel data.

Arguments

• x variable to create lag of
• data DataFrame containing x, id, and t
• id cross-section identifier
• t time variable
• lags number of lags. Can be negative, in which cause leads will be created

Returns

• A vector containing lags of data[x]. Will be missing for id and t combinations where the lag is not contained in data.

function partiallinear(y::Symbol, x::Array{Symbol, 1}, controls::Array{Symbol,1}, data::DataFrame; npregress::Function=polyreg, clustervar::Symbol=Symbol())

Estimates a partially linear model. That is, estimate

\[ y = xβ + f(controls) + ϵ\]

Assuming that E[ϵ|x, controls] = 0.

Arguments

• y symbol specificying y variable
• x
• controls list of control variables entering f
• data DataFrame where all variables are found
• npregress function for estimating E[w|x] nonparametrically. Used to partial out E[y|controls] and E[x|controls] and E[q|controls].
• clustervar symbol specifying categorical variable on which to cluster when calculating standard errors

Returns

• regression output from FixedEffectModels.jl

Details

Uses orthogonal (with respect to f) moments to estimate β. In particular, it uses

\[0 = E[(y - E[y|controls]) - (x - E[x|controls])β)*(x - E[x|controls])]\]

to estimate β. In practice this can be done by regressing (y - E[y|controls]) on (x - E[x|controls]). FixedEffectModels is used for this regression. Due to the orthogonality of the moment condition the standard errors on β will be the same as if E[y|controls] and E[x|controls] were observed (i.e. FixedEffectModels will report valid standard errors)

  partiallinearIV(y::Symbol, q::Symbol, z::Symbol,
                  controls::Array{Symbol,1}, data::DataFrame;
                  npregress::Function)

Estimates a partially linear model using IV. That is, estimate

y = αq + Φ(controls) + ϵ

using z as an instrument for q with first stage

q = h(z,controls) + u

It assumes that E[ϵ|z, controls] = 0.

Uses orthogonal (wrt Φ and other nuisance functions) moments for estimating α. In particular, it uses

0 = E[(y - E[y|controls] - α(q - E[q|controls]))*(E[q|z,controls] - E[q|controls])]

See section 4.2 (in particular footnote 8) of Chernozhukov, Chetverikov, Demirer, Duflo, Hansen, Newey, and Robins (2018) for more information.

In practice α can be estimated by an iv regression of (y - E[y|controls]) on (q - E[q|controls]) using (E[q|z,controls] - E[q|controls]) as an instrument. FixedEffectModels is used for this regression. Due to the orthogonality of the moment condition, the standard error on α will be the same as if E[y|controls] and E[q|controls] were observed (i.e. FixedEffectModels will report valid standard errors)

Arguments

• y symbol specificying y variable
• q
• z list of instruments
• controls list of control variables entering Φ
• data DataFrame where all variables are found
• npregress function for estimating E[w|x] nonparametrically. Used to partial out E[y|controls], E[q|z,controls], and E[q|controls]. Syntax should be the same as locallinear or polyreg

Returns

• α estimate of α
• Φ estimate of Φ(controls)
• regest regression output with standard error for α

  polyreg(xpred::AbstractMatrix,
          xdata::AbstractMatrix,
          ydata::AbstractMatrix; degree=1)

Computes polynomial regression of ydata on xdata. Returns predicted y at x=xpred.

Arguments

• xpred x values to compute fitted y
• xdata observed x
• ydata observed y, must have size(y)[1] == size(xdata)[1]
• degree
• deriv whether to also return df(xpred). Only implemented when xdata is one dimentional

Returns

• Estimates of f(xpred)

parse data from file in new format (2020 or newer)

parse data from file in old format (2019 or older)

parse data from file in new format (2020 or newer)