Converting old nlfcn program to newer nl function evaluator program

[Simon Baldwin]

Dear Statalist community,

I am using StataSE 16 (64-bit) on Windows 10. I am wanting to convert nl hockey (a user-written command by Dr Mark Lunt) from the old nlfcn programming style to a newer nl function evaluator program. Dr Lunt's ado file can be viewed here: https://personalpages.manchester.ac....t/nlhockey.ado

I am struggling to get it right. Please see my attempt below to convert this.

capture program drop nlhockeynew
program define nlhockeynew 

{
  version 16.0

  syntax anything if [aw fw iw], at(name) [others]

  tokenize `anything'
  local first `1'
  macro shift
  local second `1'
  macro shift
  local others `*'
  local varcount 0
  if "`others'" ~= "" {
     local varcount     : word count `others'
  }
  local counter = 1
  while `counter' <= `varcount' {
    local thisvar : word `counter' of `others'
     local xbother `xbother' + \$`thisvar'*`thisvar'
    local counter = `counter' + 1
  }
//  if "`first'" == "?" {
    
     tempname breakpoint slope_l slope_r cons
//     global S_1 breakpoint
     scalar `breakpoint' = `at'[1, 1]
//     global S_1 $S_1 slope_l
     scalar `slope_l' = `at'[1, 2]
//     global S_1 $S_1 slope_r
     scalar `slope_r' = `at'[1, 3]
//     global S_1 $S_1 `others' cons 
     scalar `cons' = `at'[1, 4]
     tempvar min max
     egen `min' = min(`second') `if'
     egen `max' = max(`second') `if'
     scalar `breakpoint'    = (`max' + `min') / 2
    regress `e(depvar)' `second' [`weight'`exp'] `if' & `second' < `breakpoint'
     tempname b
     matrix `b' = e(b)
     scalar `cons' = `b'[1,2]
     scalar `slope_l' = `b'[1,1]
     regress `depvar' `second' [`weight'`exp'] `if' & `second' > `breakpoint'
     matrix `b' = e(b)
     scalar `slope_r' = `b'[1,1]
    if "`eps'" == "" {
        tempname eps
        scalar `eps'    = (`max' - `min') / 100
     }
     local counter = 1
     while `counter' <= `varcount' {
        local thisvar : word `counter' of `others'
        global `thisvar' = 1
      local counter = `counter' + 1
     }
//     exit
//  }
  local x1     = `breakpoint' - `eps'
  local x2     = `breakpoint' + `eps'
  local b      = (`x2' * `slope_l' - `x1' * `slope_r') / (`x2' - `x1')
  local cc     = (`slope_r' - `b') / (2*`x2')
  local a      = `cons' + `slope_l'*`x1' - `b'*`x1' - `cc'*(`x1'^2)
  local alpha2 = (`a'  + `b'*`x2' + `cc'*(`x2'^2))-`slope_r'*`x2'

  replace `first' = `cons'  + `slope_l'*`second' `if' & `second' < `x1'
  replace `first' = `alpha2' + `slope_r'*`second' `if' & `second' > `x2'
  replace `first' = `a' + `b'*`second' + `cc'*`second'^2 `if' & `second' >= `x1' & `second' <= `x2'
  replace `first' = `first' `xbother' `if'

}
end

There are several issues that crop up when I try to run this. The first is a seemingly endless run through the initial iteration (Iteration 0), which when I break it returns the error "nl returned 198". I am also unsure how what would have been the "query call" section of the code from the original program (if "`first'" == "?") fits into the newer nl programming style.

If anyone has any thoughts on how Dr Lunt's nl hockey may be converted to the newer nl programming style, I would be very grateful. Many thanks.

[Simon Baldwin]

Some code that can be used to generate an example dataset is given below:

use https://stats.idre.ucla.edu/stat/stata/faq/talk, clear
expand 2
sort id
by id: gen obs = _n
replace talk = talk + .8474287 if obs == 2 & age <= 11.35291
replace talk = talk + 3.974196 if obs == 2 & age > 11.35291
replace age = age + 1 if obs == 2

The original nl hockey program would then be ran as:

nl hockey talk age

The converted program I have tried to write (which won't run) is then:

nl hockeynew @ talk age, parameters(breakpoint slope_l slope_r cons)

For more information on the old nlfcn programming style, see:

rnethelp "https://www.stata.com/stb/stb7/sg1_2/nl.hlp"

... and for the newer nl:

help nl

For further information on nl hockey, it may be worth referring to a previous post of mine regarding a separate issue: https://www.statalist.org/forums/for...tandard-errors