I'm at a dead end!!

[lux aeterna]

Hello everyone I am an undergraduate psychology student and I'm facing lots of troubles with the statistic analysis.

Firstly, I've got 1 independent variable and 7 depended ones and my data is not normally distributed.

My main problem is that I don't know when and why to use a specific statistic test.
For example, I run a Kendall's tau for correlations. Most of my dependent-independent correlations are not significant (NS).
So, my first question is if I should run a linear regression analysis (is there a non parametric equivallent?) for NS correlations? Keep in mind that my grouping variables are not equal in size.

 

[hlsmith]

Well your data don't necessarily need to be normal, that assumption is on the residuals. So linear regression my be a possibility. Tell us more about your data and there formatting. Also to confirm, you have 7 different outcomes and 1 predictor, correct?

P.S., I love your post's title, though it isn't that informative beyond describing your existential crisis!

@Dason - please provide a link to you all's LR paper, please.

 

[lux aeterna]

Thank you for your answer @hlsmith!

Yes, I've got 1 predictor (religiosity), independent variable (IV) as we use to call it, and 7 dependent variables (DV), meaning 7 aspects that I expect to be affected by my predictor. Specifically, I'm testing if religiosity helps physical (4 variables) and psychological (3 variables) adjustment.

My hypothesis is that my IV will be helpful to specific groups of people, e.g. to those with lower education (LE) but not to those with higher (HE. Groups are binomial). My IV and DVs are all continuous, likert type, and not normally distributed (except one index of physical adjustment).

I started off with Kendall's tau (I concluded it's better than Spearman's rho) and I got some significant correlations. Then I got stuck because I don't know the next step I should take.

One of my questions is if I should run a linear regression for the whole sample (approximately 150 persons) or if I should use a selection variable. Look at the image. For example LE, since I found out a positive correlation between religiosity and stress for those people.

Before starting my analysis, my supervisor told me I should use stepwise as a method, in order to control all other variables, look at the images, but I get some crazy stuff. Check the SPSS output.

Am I doing something wrong?

 

[katxt]

Be wary of false positives when you are looking at several correlations. We commonly set significance at p = 0.05 to give us 95% protection against making a false positive claim. However, with several tests this protection gets eroded. One common solution is Bonferroni - in this case make the cutoff p = 0.05/7 tests, or 0.007.

 

[lux aeterna]

Thank you for your answer, @katxt!

I correlated my IV with my 7 DVs seperately. That is Religiosity - Physical adjustment 1, Religiosity - Physical adjustment 2, ..., Religiosity - Mental adjustment 3. Look here. I couldn't find a Bonferroni criteria in the SPSS analyse / correlate / bivariate menu.

 

[Karabiner]

Do you have single Likert-type items only (= ordinal scaled) for measuring your variables, or do you have Likert scales (= interval scaled variables)?

How are the 4 physical adjustment variables related to each other - are they (nearly) uncorrelated, and conceptually independent from each other? Or do they jointly operationalize one single construct? The same question with regard to the 3 psychological adjustment variables.

How many additional predictors do you have, beyond (lower) education?

With kind regards

Karabiner

 

[lux aeterna]

Thank you for your answer, @Karabiner!

All of them are Likert scales. Each one consists of 3 to 10 Likert-type items which were summed to provide a single score.

The 4 physical adjustment variables (internal, external, physical and general) come from a single questionnaire and I think they are relatively independent (approximately rt=0,4).

The 3 mental adjustment variables (depression, anxiety, stress) come from a single questionnaire too and they relate to each other at approximately rt=0,6.

The 4 and 3 variables relate to each other at about 0,3. You might want to see in detail here.

~
My hypotheses talk about nationality, religion, education and social support (all of them binomial). Specifically, I expect that religiosity will be helpful to Non Greeks, Non Christians, Lower educated, Insufficient socially supported).

For the regression my main predictor is Religiosity (Likert scale / 10 summed Likert-type items). There are other variables too, such as previous employment, previous arrest, priest visits etc. Most of them are binomial, but there is one with 3 and one with 6 levels, plus some continuous ones, like Age.

 

[Karabiner]

I would consider performing 2 MAN(C)OVAs, each using religiosity and the binary
sociodemographic variables as predictors.

For example, you could analyse "mental adjustment", which is represented by the 3 variables
depression, anxiety, stress, and use religiosity together with your 4 binary sociodemographic
variables as predictors. You will probably have to adjust the model by kicking out all interactions
but the 4 interactions between religiosity and the sociodemographic variables. I'd guess that the
4 interactions (moderator effects) are the interesting part here, but the sample size is not
very large, and interactions are often difficult to demonstrate with non-large sample sizes.

With kind regards

Karabiner

 

[katxt]

[/USER]!

I correlated my IV with my 7 DVs seperately. That is Religiosity - Physical adjustment 1, Religiosity - Physical adjustment 2, ..., Religiosity - Mental adjustment 3. I couldn't find a Bonferroni criteria in the SPSS analyse / correlate / bivariate menu.

The Bonferroni correction probably won't be in a menu. You just have to use the idea to protect yourself. The correlations that have got p values in the 0.01 to 0.05 range may well indicate a connection but you cannot be 95% sure, so you run a more than 5% risk of making a false claim - the level of risk that is commonly thought acceptable in much research work.

 

[lux aeterna]

Thank you for your answers and sorry for taking this long to answer back, I was travelling yesterday.

@Karabiner, I am not familiar with MAN(C)OVA. Yet, I am familiar with ANCOVA and when I presented it to my Professor he replied that ANCOVA is "bad statistics" and he recommended that I use Linear Regression (LR). I don't understand the point of LR, since my correlations are neither strong nor significant (most of them), but anyway.
My professor asked me to do the stepwise method, which if I am not mistaken is Multiple Regression (MR). He said that through this method I'm controlling other variables that might be affecting my results. This makes some sense, but I was never taught MR and I get some super complicated tables that I cannot even understand.
Besides those, I've got several questions regarding LR / MR, whichever I finally use.

@katxt , are you suggesting that I should accept p values that are below 0.007? I probably don't need such accuracy in my paper. I just need to demonstrate that I am able to use the statistics I've been taught (sic) throughout my studies.

 

[Karabiner]

Yet, I am familiar with ANCOVA and when I presented it to my Professor he replied that ANCOVA is "bad statistics" and he recommended that I use Linear Regression (LR). I don't understand the point of LR, since my correlations are neither strong nor significant (most of them), but anyway.

ANCOVA and linear regression both are examples of the general linear model. They will
achieve exactely the same results, if the models are constructed accordingly.

My professor asked me to do the stepwise method, which if I am not mistaken is Multiple Regression (MR). He said that through this method I'm controlling other variables that might be affecting my results.

If by this s/he means stepwise variable elimination, then this would widely considered a silly advice.
I hope that your professor just meant hierarchical multiple regression. Or, with just 5 predictors,
you can just perform a multiple regression in one step.

With kind regards

Karabiner

 

[lux aeterna]

@Karabiner , he showed me the process described below.

Analyse / Regression / Linear / Add Dependent variable / Add the other predictors in block 1 [(they are 22 in total*, continuous and nominal, even non bivariate**)] / Choose method Stepwise / Add Religiosity (my main independent variable) in block 2 / Run the analysis.

*He didn't specifically tell me to add all 22 predictors but neither did he say to add only predictors that are relevant to my hypotheses (Nationality, Religion, Education level, Social support)
**He probably didn't remember I had non bivariate predictors.

He advised me to do that when I asked if I should run Linear Regression (LR) for the different groups of my hypotheses (i.e. LR Religiosity / Internal adjustment for Greeks / Non Greeks, Christians / Non Christians etc.)

He didn't say what kind of Regression it is, I just concluded its MR through reading about different Regression analyses. What he said was that through this method I'd be able to control all other predictors and see the effect of my main one (religiosity) only.

I have uploaded an output here. I'd appreciate it if you could tell me if the results make any sense to you. If so, I'll read more to get to understand it.

Thank you so much for your help so far.

 

[Karabiner]

I am sorry, I do not download files.

Normally, I'd suggest just to follow recommendations from one's instructor/professor/commander-in-chief, in order to avoid
problems with him/her, and to avoid a bad evaluation. But as far as I can understand, you have been given some advice which
was not fully comprehensibly for you. And for me, it is uncomprehensible even more. So I will regrettably have to drop out of this
discussion.

With kind regards

Karabiner

 

[lux aeterna]

I fully respect your decision. Thank you once again for your support so far. It helped me sort some things out.