Dear Jelle,

first, you have to decide if the variables are nominal/categorial or metric. if the variables are nominal or categorial you can take the chi-square test for unpaired samples. If the variables are paired you can use the McNemar test for 2 groups or the sign test for more than 2 ordinal variables.

For metric variables you have to calculate the normal distribution and after this, you can choose the right test. If you fulfill the first steps, I will give you some answers for the next steps. Ok?

Best regards

Tanja

Dear Tanja,

The variables are:

- Induvidual exposure time;

- # Injuries.

So, the variables are nominal/categorial I think.

(There are 6 age categories)

Thanks

Dear Jelle,

if you want to compare the 6 age categories for example with regard to the individual exposure time or the number of injuries per exposure hours, you should start with testing the distribution. The best test here seems to be the Kolmogorow-Smirnow-Test. I suppose the number of injuries will not be normally distributed, but the exposure time might be normally distributed. Thus, you can calculate with following tests:

For normally distributed parameters:

- t-test for the comparison of two groups (for example men vs. women)

- ANOVA for more than 2 groups (for example age groups)

For not normally distributed parameters:

- Mann-Whitney-U-Test for the comparison of 2 groups

- Kruskal-Wallis-test for comparison of more than 2 groups.

Best regards

Tanja

Beste Jelle,

Je kan ook non-parametrische analyse-methoden (zoals Mann-Whitney-U-Test) gebruiken voor normale distributies hoewel het niet vaak wordt gedaan is het statistisch verdedigbaar gezien het vaak zelf sterkere resultaten geeft en je sowieso geen 'randomised'-sample gaat hebben. Als je op zoek gaat naar correlaties moet je hiervoor ook rekening houden met normale of niet-normale distributie (Pearson of Spearman-test).

Vind de hoeveelheid aan te onderzoeken gegevens wel zeer veel, wordt een uitdaging. Kan je op de VUB niet terecht voor statistische ondersteuning? Altijd handig als iemand je set-up is kan bekijken.

Veel succes,

Sam

Dear Sam,

I would be interested in your answer,too.... .

Best regards

Tanja

Hello Jelle,

Texts (and statistical packages) sometimes treat the evaluation of assumptions of homogeneity and normality as preliminary to statistical analysis of data. This points to evaluation of the response variable. While this can be a useful guide to initial choice of error distribution (normal, something else) it is short of sufficient. The assumptions for the general linear model (including regression, ANOVA, and ANCOVA) are that the errors (residuals) are normal and homogeneous (Eisenhart 1947 Biometrics 3:1-21; Quinn and Keogh 2002 Experimental Design and Data Analysis for Biologists. p 110, 280). Evaluation of assumptions in advanced texts, where it occurs, usually entails a residual versus fit plot (for homogeneity) and QQ or normal score plot (for normality of the residuals).

The statistical literature warns against statistical tests to

evaluate assumptions and advocates graphical tools (Montgomery & Peck 1992; Draper & Smith 1998, Quinn & Keough 2002). La¨a¨ ra¨ (2009) gives several reasons for not applying preliminary tests for normality, including: most statistical techniques based on normal errors are robust against violation; for larger data sets the central limit theory

implies approximate normality; for small samples the power of the tests is low; and for larger data sets the tests are sensitive to small deviations (contradicting the central limit

theory).

That preliminary evaluation is insufficient can come as a surprise, as can advice against statistical tests to evaluate assumptions. But I think best practice, as in the literature, needs to be said.

Good luck with your research,

~David S

Hello Jelle,

I agree that factorial set-up is appropriate. You have 5 explanatory variables, all categorical from what you describe ---

· The differences between training, match, training and match combined (total);

· The differences between age categories;

· The differences between severity of injuries;

· The differences between game position;

· (The differences between nature of injuries).

The simplest model is

Incidence ~ TrType + Age+Severity+Pos+InjType

A factorial design makes better use of the information by including interaction terms. For example TrType*Age tells you whether differences in Incidence among the TrType categories depends on Age. You have 5*4/2 = 10 possible interaction terms. The factorial design with interaction terms can tell you more than the simple design, but you need to have enough data for each pair of explanatory variables. For example do you have data in all the cells of the TrType X Age table? For 3 TrTypes and 3 age groups, that would be 9 cells. If you have 2way tables with data in all the cells, you could add the term to the model.

For example, if data in all cells of table of TrType*Age then

Incidence ~ TrType + Age+Severity+Pos+InjType +TrType*Age

This factorial approach is more complicated than the simple model with 5 explanatory variables, but it extracts more information about your results. It's a trade-off.

A model like this can be run in any of the widely used packages. In R of course, but also menu packages such as SPSS.

B.t.w. with that many factors in you model, there is a good chance the residuals will be normal and homogeneous, even though the response variable does not look normal. So check your residuals, not the response variable.

Good luck with your research,

David S

I think I differ from the above suggestions, you had earlier said incidence rate, the outcome is number of injuries (thats count data), so that appropriate statistical test will be to either continue with epidemiological incidence rate testing between various categorical variables or combine the variable into a Poisson regression model where the outcome is count of injuries and the explanatory variables are categorical and continuous. I am just not sure if you can approximate the count variable to a continuous when the numbers of injuries within one league season might not be large enough to allowing for approximation of count data to continuous to warrant parametric or non parametric test.

Hello Jelle,

Your data might well prove to violate assumption of homogeneous residuals when using a normal error, if you

use counts. Check the residual vs fit plot. If not homogeneous,

then Poisson regression is the next step. The assumptions are about the *residuals* not about the response variable. In my experience count data is almost always "overdispersed" so a Poisson regression almost always violates homogeneity when you look at the residuals. There are no "free passes" --the residuals need to be examined, whatever error model you use. Quasi-poisson or negative binomial error address the problem.

SPSS is a good package to use for non-normal errors.

Good luck with your research,

David S