Suppose I know the mean of a normal variable and I want to estimate its standard deviation using MLE in Stata. My data looks similar to the following simulated data (it is messy panel data with missing values, etc.):

Code:
clear all

*** Generate panel data
set seed 1339
set obs 100
g id =_n
expand 20
bys id: g t = _n 
xtset id t 

g eps = rnormal(0,1)
g x = rnormal(0,1)
replace x = . if x<0 
g y = 5 + 3*L1.x + eps
I then tried to estimate the standard deviation using:

Code:
capture program drop myMLprog
program myMLprog
  version 14
  args lnf sig
  local y "$ML_y1"

  tempvar mu
  qui gen `mu' = 5 +3*L1.x
  qui replace `lnf' = ln(normalden(`y',`mu',sqrt(exp(2*`sig')))) 
end 

ml model lf myMLprog (sig: y=) 
ml maximize
display exp([sig]_cons)
which gives the error that the log likelihood could not be evaluated. I then put the calculation of the mean outside of the program and it worked (estimated sd of .98790158):


Code:
g mu= 5 +3*L1.x

capture program drop myMLprog2
program myMLprog2
  version 14
  args lnf sig
  local y "$ML_y1"

  //tempvar mu
  //qui gen `mu' = 5 +3*L1.x
  qui replace `lnf' = ln(normalden(`y',mu,sqrt(exp(2*`sig')))) 
end 

ml model lf myMLprog2 (sig: y=) 
ml maximize
display exp([sig]_cons)
This is a simplified problem of the actual model that I am trying to estimate. I think I need to understand why the solution with the tempvar does not work in this case to tackle the more complicated problem. I think it has to do with some of the missing values in the data and/or am I missing something obvious?