5
$\begingroup$

When this cell is executed:

data = {1.67655096, 1.73234376, 1.8965016, 2.06723664, 
   2.19692088, 2.20885056, 2.21420304, 2.2199184, 2.25389304, 
   2.26954224, 2.42848368, 2.62902024, 2.6572795, 
   2.76596208, 2.95570296, 3.15574056, 3.25290168, 
   3.5498736, 3.81037608, 3.88707984, 4.00737456,
    4.1241312, 4.34081592, 4.56303456, 4.57950024};
FindDistribution[data, TargetFunctions -> {LogNormalDistribution}]
FindDistribution[data, TargetFunctions -> {LogNormalDistribution}]
FindDistribution[data, TargetFunctions -> {LogNormalDistribution}]

the fitted values for mu and sigma are different for each FindDistribution. Why? Can one get the estimated uncertainties of the fitted values from a large number of these different fitted values?

$\endgroup$
  • 3
    $\begingroup$ use the option RandomSeeding->1 to get the same results in all three runs. $\endgroup$ – kglr Jan 22 '18 at 23:42
  • 1
    $\begingroup$ On Mathematica 11.1.1 (Windows 7) I get the same results for each. $\endgroup$ – JimB Jan 23 '18 at 0:26
  • $\begingroup$ I'm running Mathematica 11.2.0 on mac. $\endgroup$ – Michael B. Heaney Jan 23 '18 at 1:24
  • 1
    $\begingroup$ Does anyone know exactly how FindDistribution works? $\endgroup$ – Michael B. Heaney Jan 23 '18 at 1:28
  • 1
    $\begingroup$ Giving the different fitted values might be helpful. (If this is a bug, then when reporting that to Wolfram, you'll certainly be more convincing with those values.) $\endgroup$ – JimB Jan 23 '18 at 2:03
11
$\begingroup$

I don't know what FindDistribution is doing because it doesn't produce estimates that match any of those produced by FindDistributionParameters. I avoid FindDistribution because it is much more of a black box with minimal documentation. (A Bayesian prior is involved. From the documentation: The internal information criterion uses a Bayesian information criterion together with priors over TargetFunctions. There's nothing wrong with assuming some Bayesian prior but the documentation is lacking in any details about that. I'm more of a conservative statistician so I think its options are more along the line of "wishful thinking" given that it allows a hodgepodge of fits, finds the best fit among the candidate models but then does not account for or warn about the fishing expedition it has done.)

FindDistributionParameters will get you the maximum likelihood estimates but not measures of precision:

mle = FindDistributionParameters[data, LogNormalDistribution[μ, σ]]
(* {μ -> 1.0309, σ -> 0.30526} *)

One can use the LogLikelihood function to get estimates of the standard errors:

logL = LogLikelihood[LogNormalDistribution[μ, σ], data];
cov = -Inverse[(D[logL, {{μ, σ}, 2}]) /. mle];
seμ = cov[[1, 1]]^0.5
(* 0.0610519 *)
seσ = cov[[2, 2]]^0.5
(* 0.0431702 *)
$\endgroup$
  • 1
    $\begingroup$ Did you mean to say "FindDistribution" instead of "FittedDistribution?" $\endgroup$ – Michael B. Heaney Jan 23 '18 at 2:32
  • $\begingroup$ Oops! Thanks for catching that. I'll change it. $\endgroup$ – JimB Jan 23 '18 at 3:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.