rsquared for lmfit

[Anselmoo]

Is there actually still interest in implementing rsquared as formally proposed #81?

This still might fit there?

lmfit-py/lmfit/minimizer.py

Lines 358 to 380 in e54a44c

	def _calculate_statistics(self):
	"""Calculate the fitting statistics."""
	self.nvarys = len(self.init_vals)
	if not hasattr(self, 'residual'):
	self.residual = -np.inf
	if isinstance(self.residual, np.ndarray):
	self.chisqr = (self.residual**2).sum()
	self.ndata = len(self.residual)
	self.nfree = self.ndata - self.nvarys
	else:
	self.chisqr = self.residual
	self.ndata = 1
	self.nfree = 1
	self.redchi = self.chisqr / max(1, self.nfree)
	# this is -2*loglikelihood
	self.chisqr = max(self.chisqr, 1.e-250*self.ndata)
	_neg2_log_likel = self.ndata * np.log(self.chisqr / self.ndata)
	self.aic = _neg2_log_likel + 2 * self.nvarys
	self.bic = _neg2_log_likel + np.log(self.ndata) * self.nvarys
	def _repr_html_(self, show_correl=True, min_correl=0.1):
	"""Return a HTML representation of parameters data."""
	return fitreport_html_table(self, show_correl=show_correl,

See also: python-lmfit-how-to-calculate-r-squared

[newville]

PR #81 was merged more than 8 years ago. Are you talking about a comment that was both a) made on a closed PR many months after it was merged and b) made more than 7 years ago? If so, then I would say that implementing rsquared was not ever "formally proposed". It was sort of suggested in passing in a way that could not possibly be useful 7 years ago and basically never brought up again.

You can certainly bring up the idea of adding some more fit statistics - I don't know that anyone would object to that.

[Anselmoo]

@newville thx for quick response.

I was referring to #81 (comment)

I am still thinking that an R2 analysis can help to further add confidence in the fitting results. However, I don't wanna duplicate a closed Issue and PR, so I was asking, what your thoughts are?

thx again

[newville]

@Anselmoo it would be fine to open an Issue on that and/or make a PR to add that.

[Anselmoo]

@newville

I was taking a deeper look into the source code of lmfit-minimizer, and I was wondering if it is possible to get somehow the initial data and not just the residual?

The reason is that for calculating R2, we need the average value of data as proposed by sklearn.metrics.r2_score

    numerator = (weight * (y_true - y_pred) ** 2).sum(axis=0, dtype=np.float64)
    denominator = (
        weight * (y_true - np.average(y_true, axis=0, weights=sample_weight)) ** 2
    ).sum(axis=0, dtype=np.float64)

https://github.com/scikit-learn/scikit-learn/blob/36958fb240fbe435673a9e3c52e769f01f36bec0/sklearn/metrics/_regression.py#L927-L930

Is there a chance to reconstruct data by using self.userfcn? I am a little bit lost.

thx for feedback

[newville]

Well, no it is not generally possible to get the "data" in the minimizer interface. That simply does not have a concept of "data". It has an objective function that converts a set of parameters into a residual whose sum-of-squares is minimized. That residual could be (data-model), but it might not be. It might include multiple sets of different kinds of "data" (in the sense of "observations" or "measurements"). It could simply not include any data at all -- sort of a "non-linear root finding". In any case, it does not separately think about "data".

The lmfit.Model interface does include a clearer distinction between "data" and "model" for the common case of "curve fitting", where the residual is weight*(data-model). Any statistic such as R2 that separates "data" and "model" would need to be done there.

[Anselmoo]

The lmfit.Model interface does include a clearer distinction between "data" and "model" for the common case of "curve fitting", where the residual is weight*(data-model). Any statistic such as R2 that separates "data" and "model" would need to be done there.

thx for the input, this sounds very promising. I will take a look