It depends on what your dependent variable is.. I can't tell from the question. If the DV is continuous try ols regression, if categorical try logistic regression. See the attached file.

Best regards, David Booth

For specifying these hypotheses as regression models, you need to translate the term "moderated" into into interaction effects. So, look up how to create and interpret interaction effects in regression.

thanks! wouldn't be easier to exclude completely any dummy coding and use ANCOVA instead and enter the levels of social media usage as covariate?

ANCOVA will not answer your questions about moderating (interaction) effects. Your first hypothesis requires dummy coding to assess how age categories interact with level of media usage to predict attractiveness. Your second hypothesis does not seem to involve the age groups, so there is no need for dummy variables; instead, you are asking about an interaction effect between two continuous variables -- perceived age and media usage.

thanks again David! this is exactly what I had in mind however there has been a bit of a confusion on how to approach this and advice varied between hierarchical multiple regression, ANCOVA and others. the original approach was to measure effects on two dependent variables separately (attractiveness and perceived age) with interaction between age groups and social media usage as predictors. However that does not make sense re hypothesis 2

Not everyone means the same thing when they say ANCOVA. What I was taught (and what David is thinking, I think) is that ANCOVA models do not include any interactions between group variables and (quantitative) covariates.

But I also wondered if what Katarina meant when she suggested using ANCOVA was that she could use the UNIANOVA command in SPSS rather than the REGRESSION command. REGRESSION treats all explanatory variables as quantitative, and therefore, one must compute k-1 indicator (or dummy) variables for each categorical explanatory variable they want to include; and if that categorical variable is involved in any interactions, more variables have to be computed. UNIANOVA, on the other hand, allows one to include both quantitative and categorical explanatory variables. The former are called "covariates" in the GUI, and the latter "fixed factors". (In UNIANOVA syntax, keyword BY is followed by a list of fixed factors, and keyword WITH is followed by a list of covariates.) Nor is there any need to compute product terms for interactions if you use UNIANOVA--you can include product terms on the DESIGN sub-command. Here's a (relatively) simple example to illustrate the differences between REGRESSION and UNIANOVA for a model with one quantitative explanatory variable, one categorical explanatory variable, and their interaction. HTH.

NEW FILE.

DATASET CLOSE ALL.

* Change path on next line to folder with SPSS sample datasets.

GET FILE='C:\SPSSdata\survey_sample.sav'.

FREQUENCIES marital.

DESCRIPTIVES age.

* Create marital status indicator variables for

* use with REGRESSION command below.

NUMERIC m1 to m5 (F1) / AxM1 to AxM5(F5.2).

DO REPEAT i = 1 to 5 / dummy = m1 to m5 /int = AxM1 to AxM5.

- COMPUTE dummy = marital EQ i.

- COMPUTE int = dummy*age.

END REPEAT.

* Confirm that indicator variables were computed correctly.

MULT RESPONSE GROUPS=$mstatus 'Marital Status' (m1 m2 m3 m4 m5 (1))

/VARIABLES=marital(1 5)

/TABLES=marital BY $mstatus

/CELLS=COUNT

/BASE=CASES.

* First estimate the model via REGRESSION.

REGRESSION

/STATISTICS COEFF OUTS CI(95) R ANOVA

/DEPENDENT educ

/METHOD=TEST (age) (m1 to m4) (AxM1 to AxM4).

* Now estimate the same model via UNIANOVA.

UNIANOVA educ BY marital WITH age

/EMMEANS=TABLES(marital) WITH(age=25) COMPARE

/EMMEANS=TABLES(marital) WITH(age=50) COMPARE

/EMMEANS=TABLES(marital) WITH(age=75) COMPARE

/PRINT=PARAMETER

/DESIGN=age marital age*marital.

* UNIANOVA allows one to generate fitted values of Y

* at selected combinations of Age and Marital Status

* via the EMMEANS sub-command. You can also use

* OMS to write those fitted values to another dataset,

* and then graph them etc.

thanks Bruce, yes I definitely meant UNIANOVA command that allows to enter categorical and continuous variable, one as fixed factor and one as covariate, from what I checked, that does not require any additional dummy coding. I have looked at interactions in linear regression as suggested by David. Unfortunately, I have very little time left to decide which way to go as I need to get the writing done in less than a week now

Regression analysis is just a special case of the general linear model. If you use the General Linear Model (GLM) in SPSS, and enter the continuous independent variables as Covariates and the categorical independent variables as Factors, SPSS will create the dummy variables to replace the factors in the regression analysis, each time choosing one of the categories as reference category. You have to specify under "options" that you want the regression coefficients displayed. Under the "model" you may select which main effects and interactions you want to include.

Many thanks Francois and many thanks to everybody who contributed into discussion. I have now closed this as I only have 2 days left to finish the write up. In the end, we decided to go for 2x2 between subjects ANOVA and split main IV and moderator into two binary categorical variables in order to observe interactions

Unnecessarily coarsening variables results in needless loss of power and can cause other problems. Here are a few good articles to look at when you have time.

Article Royston P, Altman DG, Saurbrei WDichotomizing continuous pre...

Article Statistics Notes: The cost of dichotomising continuous variables

Article Breaking up is hard to do: The heartbreak of dichotomizing c...

One of the classic references on how to handle interactions involving continuous variables (in regression models) is this book:

Aiken LS, West SG, Reno RR. Multiple regression: Testing and interpreting interactions. Sage; 1991.

HTH.

many thanks Bruce. we are aware of this. unfortunately there are already various issues with the data and social media scores in particular are affected by ceiling effect when majority of tested individuals are a heavy social media users hence this is the best option under circumstances. main IV was originally measured as categorical so instead of 4 unequal sized groups we decided to go for 2: younger and older adults. we have considered numerous parametric and non parametric options.