1
$\begingroup$

I am trying to run a linear regression model (ideally) to see whether age (continuous variable) affects levels of stress hormone (also continuous, dependent variable), i.e. hypothesis testing. My hormone level results are right-skewed and non-normal, and the residuals of a simple linear model have a similar distribution (see plots of residuals below). Age is normally distributed.

Scatter of raw data (has a couple outliers): datapoints Histogram of residuals: hist_nontransformed QQplot_nontransformed

I have also tried various transformations of my dependent variable (hormone levels) to improve the outcome of the model. The best normality plots came from the Tukey transformation. See plots below.

Scatter of transformed hormone level on age: transformed_scatter Histogram of residuals: hist_transformed QQplot_transformed

However, the transformation yields a lower multiple R-squared and adjusted R-sqaured value than the original.

Further to this post (What if a transformed variable yields more normal and less heteroskedastic residuals but lower $R^2$?), I tried a model where the hormone level is log transformed and fitted with the exponent of age to correct for the log transformation. This is because I didn't know what the opposite of Tukey transformation would be and log seems to be the next best transformation for my data. The normality of residuals is good but the R2 values are even lower.

Given this information, which model is best - the one with best normality of residuals and lowest R2 or worst normality of residuals and highest R2?

Alternatively, is there another form of regression better suited to this data, given that it is continuous and has non-normal residuals?

$\endgroup$
  • 2
    $\begingroup$ I am strong on residual plots, but this is Hamlet with Ophelia as central character and no Prince of Denmark. More plainly put, what would be direct here are plots of hormone versus age and of transformed hormone versus age, as I read your question as implying that you have one outcome, hormone level and one predictor, age. $\endgroup$ – Nick Cox Oct 28 '19 at 19:12
  • 1
    $\begingroup$ What's "the Tukey transformation"? I've been reading Tukey's works assiduously for nearly 50 years and have no idea what you're implying there. $\endgroup$ – Nick Cox Oct 28 '19 at 19:14
  • $\begingroup$ @NickCox , likely Tukey's Ladder of Powers transformation. $\endgroup$ – Sal Mangiafico Oct 28 '19 at 22:08
  • $\begingroup$ @Sal Mangiafico Thanks; I am aware of that, but the point is which transform is chosen, and why and how, that makes it the transformation. $\endgroup$ – Nick Cox Oct 28 '19 at 22:23
  • $\begingroup$ Getting normal residuals is nice but not the main goal of regression. Providing a good handle on the relationship between variables comes first. Just 23 data points are in this sample, so listing the data would be easy as well as helpful. $\endgroup$ – Nick Cox Oct 28 '19 at 22:28
0
$\begingroup$

In most cases, residuals will never be perfectly normal. In the first QQ plot you posted, the residuals (assuming that's what the plot if of) follow normality except for your two outlying points at the end. I really wouldn't sweat that too much, outliers happen, and it's not the worst departure from normality I've ever seen.

I'd check a scatter plot of your residuals in the first case (in all cases, really) and see if you notice any patterns. I strongly suspect that there exists an explanatory variable you haven't accounted for yet. Adding it to the model may help your residuals conform.

Another thing to consider is that you may wish to add a squared term, something like

$Y = B_0 + B_1*age + B_2*age^2$. Giving your model that quadratic flexibility might help as well.

And if none of the above works, you may need to try nonlinear regression.

And if all else fails and it comes down to choosing between a higher $R^2$ or normal residuals... normal residuals get violated a lot in practice. Personally, I'd find a model which can explain the more variation the most useful. But there are instances where that's not the case. Use your best judgement.

Best of luck.

| cite | improve this answer | |
$\endgroup$
0
$\begingroup$

Let's look at your final question:

Alternatively, is there another form of regression better suited to this data, given that it is continuous and has non-normal residuals?

Yes, there is. You could use quantile regression. This makes no assumptions about the distribution of the residuals. It also allows you to look at more questions than regular OLS regression does. You can do it in R with the quantreg package or in SAS with PROC QUANTREG.

| cite | improve this answer | |
$\endgroup$

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.