0
$\begingroup$

I'm going to analyse suicide rates for a time series, and I'd like to use robust tests, but I don't know which would be a good one. My purpose is to compare the variation of the suicide rates through the months and years and see if there are significant changes in respect to GDP. Then I also want to check differences among sex, age, etc. What would you recommend to me?

I've heard about robust tests in Introduction to Robust Estimation and Hypothesis Testing by Rand Wilcox. They are the best choice when normality and other assumptions are violated. Thanks.

$\endgroup$
4
  • 2
    $\begingroup$ I don't understand your question. What does it mean to "test... for a time series"? Are you wondering if there is autocorrelation? something else? $\endgroup$ Commented Nov 25, 2013 at 21:49
  • 2
    $\begingroup$ For that matter, please say more on what "robust" means to you. $\endgroup$
    – Nick Cox
    Commented Nov 25, 2013 at 21:58
  • 1
    $\begingroup$ Gung, I meant analyse or study, don't know why I wrote test, I guess I was thinking of the key question xD. By robust I mean that they don't assume normal distribution and homocedasticity. I know there are non parametric tests, but the newest thing are the robust tests. Rand Wilcox has a book about them (Introduction to Robust Estimation and Hypothesis Testing). It seems they are better than classic tests when normality is violated. $\endgroup$
    – user35205
    Commented Nov 26, 2013 at 8:18
  • 2
    $\begingroup$ Your question remains ambiguous as you give no details about what "suicide rates" mean precisely and about how you intend to analyse them. What's predictable is that rates, however defined, cannot be negative, but hardly more. If your time and space mesh is fine enough, there will be zeros in the data. Whether you can get away with an assumption that error structure is Gaussian remains an open question, but it will be a real stretch if zeros are present. $\endgroup$
    – Nick Cox
    Commented Nov 26, 2013 at 12:17

1 Answer 1

1
$\begingroup$

Gaussian violations (untreated) often lead one to non-parametric tests. I would suggest that you consider building an ARMAX model properly dressed up for any needed remedies to render the error structure Gaussian . Analysis of the model residuals might then suggest the need for alternative strategies BUT might not. Predicting percentages is in my opinion not the way to go ...rather predict the count using a model that incorporates the total. This leads nicely to converting the predicted count to a percentage based upon the predicted total.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.