See these links, please:

https://statistics.laerd.com/spss-tutorials/one-way-manova-using-spss-statistics.php

https://ezspss.com/test-for-normality-in-spss/

https://stats.oarc.ucla.edu/spss/dae/one-way-manova/

https://www.youtube.com/watch?v=rCgeWeXRtDs

https://www.youtube.com/watch?v=AtHZLMWZyYo

You may need to also do a test for homogeneity of variances before proceeding to ANOVA. The following link will take you to a very nice tutorial for checking both Assumptions of ANOVA (normality of distribution and homogeneity of variance) using SPSS. Good luck!

https://www.youtube.com/watch?v=xxcE5Kyo3vU

Hi Emily doris Barbuscia

May see if there arebany potential outliers impacting normality of data distribution.

There are several issues here. First, I'd argue MANOVA is rarely a sensible approach and should probably be avoided. See:

http://psychologicalstatistics.blogspot.com/2021/08/i-will-not-ever-never-run-manova.html

Data don't need to be normally distributed for ANOVA - the assumption is that the residuals are sampled from a normal distribution and hence should be approximately normal. I would therefore use. residual plot rather than a significance test of departures from normality. As data are never exactly normally distributed tests of assumptions are a bad idea (lacking power to detect problems in small samples and detecting very minor departures in large samples).

Thom Baguley is right about MANOVA, but I'm curious : Thom – have you ever encountered a case where MANOVA was the analysis of choice, tested a definable hypothesis and produced an interpretable measure of effect size?

Neither have I!

Thank you so much for all your answers and links! Really appreciated it !

Ronán Michael Conroy: I haven't encountered a sensible use of MANOVA and I think that "a case where MANOVA was the analysis of choice, tested a definable hypothesis and produced an interpretable measure of effect size" is probably impossible because of the last bit as the effect size is always going to be on the variance accounted for in the linear transformed function of the DVs and that unlikely to be interpretable.

I can conceive of situations where the MANOVA test is appropriate and somewhat useful. For example imagine you have a dozen DVs and you want to test whether they are unrelated to a factor. The omnibus test of MANOVA gives you a single quick test of the alt hypothesis that the factor is related to the DVs. Not necessarily the best approach but I think it is a reasonable to report that.

Thom Baguley – I have been trying to conceive of situations in which you have a dozen DVs – for example a dozen health states – and you are wondering if a factor such as gender is related to them. But I cannot see the rationale for pooling them into an unknowable mashup variable whose weighting of each disease is solely to optimise prediction. The whole thing does not correspond to any sensible hypothesis I can come up with.

Huang FL. MANOVA: A procedure whose time has passed?. Gifted Child Quarterly. 2020 Jan;64(1):56-60.

I would go further and argue that it's a procedure that never had a time, like one of those baffling things on the swiss army knife that no-one has ever found a use for. (I'm an Opinel person, out-and-out!)

@Ronan Michael Conroy I have found multivariate random effects analysis which has several response variables without the 'mashup' of MANOVA to be useful. You get the variance of each outcome separately and the correlation between each pair of outcomes before and after conditioning on 'explanatory' variables. It can also bee useful when there are missing observations on the response under the fairly forgiving MAR assumption; for example you have asked one question of everybody but more detailed questions of random subsets - the estimation of the effects for the models of the smaller subsets can potentially borrow strength from the larger sample.

Modern implementations allow simultaneous estimation of discrete and continuous response variables. As an example, see

Article Investigating the Macro Determinants of Self-Rated Health an...

I would suggest transforming your data to logarithmic value if the continuous data is non-parametric.

If a sample size is large enough, particularly in relation to the number of variables, and data has no extreme outliers, ANOVA models can be robust to normality violations. Assessing their normality can be attained numerically (e.g., employing Shapiro-Wilk Test) or graphically (e.g., using boxplots of residuals). However, as pointed out by the gentlemen above, classical MANOVA could be error-prone, especially when analyzing multiple dependent variables (see Huang, 2020; Smith et al., 2020). Therefore, as an option, you could check out the modified ANOVA-type statistic (MATS) by Friedrich and Pauly (2018). Here are the full citations.

Friedrich, S., & Pauly, M. (2018). MATS: Inference for potentially singular and heteroscedastic MANOVA. Journal of Multivariate Analysis, 165, 166–179. https://doi.org/10.1016/j.jmva.2017.12.008

Huang, F. L. (2020). MANOVA: A procedure whose time has passed? Gifted Child Quarterly, 64(1), 56–60. https://doi.org/10.1177/0016986219887200

Smith, K. N., Lamb, K. N., & Henson, R. K. (2020). Making meaning out of MANOVA: The need for multivariate post hoc testing in gifted education research. Gifted Child Quarterly, 64(1), 41–55. https://doi.org/10.1177/0016986219890352

Good luck,

Ronán Michael Conroy My thought was that the fact that the MANOVA DV is a mashup of the other DVs isn't much of an issue if they are accounting for negligible variation and positively correlated, however, I would not want to advocate this with any conviction. There certainly would be more principled approaches.