failing regression diagnostics and convenience sampling

[Kathryn Kearney]

Hi there,

I'm currently running a multiple regression model with an ordinal dependent variable. Initially, I used OLS and treated the dependent as effectively continuous, however, the model was found to be heteroskedasic and non-linear. Therefore, I ran an Ordinal Logistic Regression instead. However, this failed the proportional odds assumption (likelihood ratio test i followed from the UCLA stats page showed a prob>chi2 is 0.006). My last choice is to run a Generalised Ordered Logit model (Williams, 2006), which I'm just running now, although it's taking a long time to compute.

However - it's just struck me that maybe the reason my data is failing everything is because it's convenience sampled? I know using this type of sampling is bad but I'm just interpreting the p values heuristically and also just using it to talk in relation to my sample, not to generalise in any way.

Could convenience sampling be having an impact?

I'm pretty new to statistics and quant methods, so forgive me if this is a stupid question!

Thanks for your help and best wishes,
Kathryn.

[Carlo Lazzaro]

Kathryn:
some remarks about your query:
- if the residual distribution of an OLS turns out to be heteroskedastic, you can invoke -robust- standard error;
- the OLS framework allows non-lineraity between predictors and dependent variable; the coefficients should be linear instead.
After OLS you moved to (much) more demanding regression models, which require the researcher a sound theoretical knowledge.
That said, I do not think that convenience sample is to blame for what you're experiencing; I would rather focus on model specification (that is: does your regression model give a fair and true view of the data generating process?):
As an aside, please note that your chances of receiving helpful replies is conditional on posting what you typed and what Stata gave you back (via CODE delimiters). Thanks.