## (urgent) Sample Size

misterhussein
PLS Junior User
Posts: 1
Joined: Thu Jun 17, 2010 9:25 pm
Real name and title:

### (urgent) Sample Size

Hi All,

Please advise me how to determine the appropriate sample size for running 2x2 factorial design by using PLS.

Kind Regard
allenjia
PLS Senior User
Posts: 22
Joined: Sat Mar 26, 2011 6:43 am
Real name and title:
above 1000, 5000 is recommended by Dr. Chingle
Do the formative by using PLS!!
Diogenes
PLS Super-Expert
Posts: 905
Joined: Sat Oct 15, 2005 5:13 pm
Real name and title:
Location: São Paulo - BRAZIL
Contact:
Hi,

1000 or 5000 is recommended to the bootstrap procedure.

To the 2x2 factorial design I thought:

1) ANOVA to compare 4 groups (mean) = 180 cases [alpha = 5%, power = 80%, effect size = medium = 0.25]

2) Multigroup comparison (two independent path coefficients) = 356 cases [alpha = 5%, power = 80%, effect size = medium = 0.3]

See (Cohen, 1977) and use G*Power 3 (Faul, Erdfelder, Lang, & Buchner, 2007), it is free.

Cohen, J. (1977). Statistical power analysis for the behavioral sciences (Revised.). New York: Academic Press.
Faul, F., Erdfelder, E., Lang, A.-G., & Buchner, A. (2007). G*Power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior research methods, 39(2), 175-91.

Best regards,

Bido
anugrah
PLS User
Posts: 14
Joined: Thu Apr 28, 2011 9:05 am
Real name and title:
Hi All,

I hv the same problem too..
Please advise me how to determine the appropriate sample size for running 91 observations by using PLS?
husseins
PLS User
Posts: 13
Joined: Mon Dec 07, 2009 9:08 pm
Real name and title:
Halo Pak Anugrah,

Maybe you can try this:

Chin (1998) the sample required for PLS can be determined in the following ways: 1) the construct with the largest number of formative indicators or 2) the dependent latent variable with the largest number of relationship. The sample size required would be 5 to 10 times of either 1) or 2) whichever is higher

Semoga membantu...

Salam,
Diogenes
PLS Super-Expert
Posts: 905
Joined: Sat Oct 15, 2005 5:13 pm
Real name and title:
Location: São Paulo - BRAZIL
Contact:
Hi

The best way is using power analysis:
1) How many arrows arrive in each LV?
2) Choose the biggest number
3) Use this number as number of predictors in the multiple regression procedure in the G*Power to have the sample size.

Software G*Power 3 available at
http://www.psycho.uni-duesseldorf.de/ab ... p/gpower3/

Best regards,
Bido
anugrah
PLS User
Posts: 14
Joined: Thu Apr 28, 2011 9:05 am
Real name and title:
Thank you for reply my question Pak Hussein and Mr Bido,

I owned a maximum of 7 LVs, will that mean my sample size is 70?
Can a sample size be smaller that the total number of cases?
(cases=91)

If i use G*Power3, i find 2 types.
There are 2 linear multiple regression
1) Fixed model, R2 deviation from zero (I have 138 sample sizes)
2) Fixed model, R2 increase (I have 107 sample sizes)

Which one should i use?
Is that ok from that results? or should be above 1000?

NB:
My model is hierarchical component model (repeated indicator approach)

Kind Regard
Diogenes
PLS Super-Expert
Posts: 905
Joined: Sat Oct 15, 2005 5:13 pm
Real name and title:
Location: São Paulo - BRAZIL
Contact:
f2 = .15 --> R2 = 13% medium
alfa = 5%
Power = 80%
Predictors = 2 (KPR & KPS) = maximum number of the arrows arriving in a LV.
Sample size = 68 cases

Best regardss,

Bido
anugrah
PLS User
Posts: 14
Joined: Thu Apr 28, 2011 9:05 am
Real name and title:
Test> Correlation and regression>multiple regression:fixed model, R2 deviation from zero.(is it true?)

Why should i use 80% not 95%? [80% = 68 samples; 95% = 107 samples]
and where can i find/calculate R2 from G*Power?
Diogenes
PLS Super-Expert
Posts: 905
Joined: Sat Oct 15, 2005 5:13 pm
Real name and title:
Location: São Paulo - BRAZIL
Contact:
Hi,

1) Yes.

2) 0.8 is the minimum value recommended by Hair Jr. et al., but it depends of the field of research, in fact the default value of the G*Power is .95.

3) G*Power gives us the effect size (f2), and we need to compute the R2, see:

FAUL, F.; ERDFELDER, E.; LANG, A.-G.; BUCHNER, A. G*Power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior research methods, v. 39, n. 2, p. 175-91, 2007.

f2 = R2/(1 – R2) [Faul et al., 2007, p.181]

R2 = f2/(1 + f2)

Best regards,

Bido
mazi
PLS User
Posts: 18
Joined: Sat Apr 23, 2011 2:34 pm
Real name and title:
i have 8 construct. for loyalty construct there are 4 direct predictor, s for using the mentioned software.
i can go to the
test> correlation and regression> linear multiple R2 zero
and for estimating the sample size
how i can acquire the effect size?
as you said
power=0.8
alpha=0.05
number of predictor is 4
mazi
PLS User
Posts: 18
Joined: Sat Apr 23, 2011 2:34 pm
Real name and title:

### value of f2 in Gpower

is it true for my sample size 85? i dont know which vaue insert in effect size box?. for powe i insert 0.8. for alpha 0.05, and because 4 paths arrived ate loyalty which is the maximum i set 4 for number of predictor
Thank you
Nazila Babakhani
Diogenes
PLS Super-Expert
Posts: 905
Joined: Sat Oct 15, 2005 5:13 pm
Real name and title:
Location: São Paulo - BRAZIL
Contact:
Hi,

the classification given by Cohen and used in G*Power are:
.02 = small
.15 = medium
.35 = large

With .02 the sample will be able to detect small effects as significant.

With .36 the sample will be able to detect just large effects (R2 greater than .26) as significant.

Best regards,

Bido
mazi
PLS User
Posts: 18
Joined: Sat Apr 23, 2011 2:34 pm
Real name and title:
so you mean that small effect size give more sample size which is better? which of them should i use?
Best regards
Nazila Babakhani
Diogenes
PLS Super-Expert
Posts: 905
Joined: Sat Oct 15, 2005 5:13 pm
Real name and title:
Location: São Paulo - BRAZIL
Contact:
With “small effects” your sample will be larger, and more sensible, it will detect small effects as significant.

Usually, to collect data is a problem (cost, time..) and we use the “medium effect”.

Remembering that with small samples we could finish with a model where nothing is significant.

Best regards,

Bido