RESEARCH METHOD COHEN ok
RESEARCH METHOD COHEN ok RESEARCH METHOD COHEN ok
EFFECT SIZE 523 Box 24.17 The Levene test for equality of variances Levene’s test for equality of variances Independent samples test t-test for equality of means Chapter 24 How well learners are cared for, guided and supported Equal variances assumed Equal variances not assumed 95 % confidence interval of the difference Sig. Mean SE F Sig. t df (2-tailed) difference difference Lower Upper 8.344 0.004 1.923 998 0.055 0.30 0.155 −0.006 0.603 2.022 811.922 0.044 0.30 0.148 0.009 0.589 The effect size can be worked out thus (using SPSS): t 2 t 2 + (N 1 − 1) = 16.588 2 16.588 2 + (1000 − 1) = 275.162 275.162 + 999 = 0.216 In this example the effect size is 0.216, a very large effect, i.e. there was a very substantial difference between the scores of the two groups. For analysis of variance (discussed later) the effect size is calculated thus: Eta squared = Sum of squares between groups Total sum of squares In SPSS this is given as ‘partial eta squared’. For example, let us imagine that we wish to compute the effect size of the difference between four groups of schools on mathematics performance in a public examination. The four groups of schools are: rural primary; rural secondary; urban primary; urban secondary. Analysis of variance yields the result shown in Box 24.20. Working through the formula yields the following: Sum of squares between groups = 7078.619 Total sum of squares 344344.8 =0.021 alternative equations to take account of Box 24.18 Mean and standard deviation in a paired sample test Paired samples statistics Mean N SD SE mean Pair Maths 81.71 1,000 23.412 0.740 1 Science 67.26 1,000 27.369 0.865 The figure of 0.021 indicates a small effect size, i.e. that there is a small difference between the four groups in their mathematics performance (note that this is a much smaller difference than that indicated by the significance level of 0.006, which suggests a statistically highly significant difference between the four groups of schools. In regression analysis (discussed later) the effect size of the predictor variables is given by the beta weightings. In interpreting effect size here Muijs (2004: 194) gives the following guidance: 0–0.1 weak effect 0.1–0.3 modest effect 0.3–0.5 moderate effect >0.5 strong effect Hedges (1981) and Hunter et al. (1982) suggest
524 QUANTITATIVE DATA ANALYSIS Box 24.19 Difference test for a paired sample Paired samples test Paired differences 95 % confidence interval of the difference Mean SD SE mean Lower Upper t df Sig. (2-tailed) Pair 1 Maths-Science 14.45 27.547 0.871 12.74 16.16 16.588 999 0.000 Box 24.20 Effect size in analysis of variance Maths ANOVA Sum of squares df Mean square F Sig. Between groups 7078.619 3 2359.540 4.205 0.006 Within groups 337266.2 601 561.175 Total 344344.8 604 differential weightings due to sample size variations. The two most frequently used indices of effect sizes are standardized mean differences and correlations (Hunter et al. 1982:373),although non-parametric statistics, e.g. the median, can be used. Lipsey (1992: 93–100) sets out a series of statistical tests for working on effect sizes, effect size means and homogeneity. Muijs (2004: 126) indicates that a measure of effect size for cross-tabulations, instead of chisquare, should be phi, whichisthesquarerootof the calculated value of chi-square divided by the overall valid sample size. He gives an example: ‘if chi-square = 14.810 and the sample size is 885 then phi = 14.810/885 = 0.0167 and then take the square root of this = 0.129’. Effect sizes are susceptible to a range of influences. These include (Coe 2000): Restricted range: the smaller the range of scores, the greater is the possibility of a higher effect size, therefore it is important to use the standard deviation of the whole population (and not just one group), i.e. a pooled standard deviation, in calculating the effect size. It is important to report the possible restricted range or sampling here (e.g. a group of highly able students rather than, for example, the whole ability range). Non-normal distributions: effect size usually assumes a normal distribution, so any non-normal distributions would have to be reported. Measurement reliability: the reliability (accuracy, stability and robustness) of the instrument being used (e.g. the longer the test, or the more items that are used to measure a factor, the more reliable it could be). There are downloadable software programs available that will calculate effect size simply by the researcher keying in minimal amounts of data, for example: The Effect Size Generator by Grant Devilly: http://www.swin.edu.au/victims
- Page 492 and 493: METHODOLOGICAL TOOLS FOR ANALYSING
- Page 494 and 495: 23 Content analysis and grounded th
- Page 496 and 497: HOW DOES CONTENT ANALYSIS WORK 477
- Page 498 and 499: HOW DOES CONTENT ANALYSIS WORK 479
- Page 500 and 501: HOW DOES CONTENT ANALYSIS WORK 481
- Page 502 and 503: A WORKED EXAMPLE OF CONTENT ANALYSI
- Page 504 and 505: A WORKED EXAMPLE OF CONTENT ANALYSI
- Page 506 and 507: COMPUTER USAGE IN CONTENT ANALYSIS
- Page 508 and 509: COMPUTER USAGE IN CONTENT ANALYSIS
- Page 510 and 511: GROUNDED THEORY 491 data, thereby c
- Page 512 and 513: GROUNDED THEORY 493 fragments are t
- Page 514 and 515: INTERPRETATION IN QUALITATIVE DATA
- Page 516 and 517: INTERPRETATION IN QUALITATIVE DATA
- Page 518 and 519: INTERPRETATION IN QUALITATIVE DATA
- Page 520 and 521: 24 Quantitative data analysis Intro
- Page 522 and 523: DESCRIPTIVE AND INFERENTIAL STATIST
- Page 524 and 525: DEPENDENT AND INDEPENDENT VARIABLES
- Page 526 and 527: EXPLORATORY DATA ANALYSIS: FREQUENC
- Page 528 and 529: EXPLORATORY DATA ANALYSIS: FREQUENC
- Page 530 and 531: EXPLORATORY DATA ANALYSIS: FREQUENC
- Page 532 and 533: EXPLORATORY DATA ANALYSIS: FREQUENC
- Page 534 and 535: STATISTICAL SIGNIFICANCE 515 alongt
- Page 536 and 537: STATISTICAL SIGNIFICANCE 517 hands
- Page 538 and 539: HYPOTHESIS TESTING 519 selection fr
- Page 540 and 541: EFFECT SIZE 521 differential measur
- Page 544 and 545: THE CHI-SQUARE TEST 525 The Effect
- Page 546 and 547: DEGREES OF FREEDOM 527 Box 24.22 A2
- Page 548 and 549: MEASURING ASSOCIATION 529 Box 24.23
- Page 550 and 551: MEASURING ASSOCIATION 531 found and
- Page 552 and 553: MEASURING ASSOCIATION 533 Box 24.26
- Page 554 and 555: MEASURING ASSOCIATION 535 Many usef
- Page 556 and 557: REGRESSION ANALYSIS 537 we know or
- Page 558 and 559: REGRESSION ANALYSIS 539 Box 24.32 S
- Page 560 and 561: REGRESSION ANALYSIS 541 Box 24.35 S
- Page 562 and 563: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 564 and 565: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 566 and 567: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 568 and 569: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 570 and 571: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 572 and 573: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 574 and 575: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 576 and 577: MEASURES OF DIFFERENCE BETWEEN GROU
- Page 578 and 579: 25 Multidimensional measurement and
- Page 580 and 581: FACTOR ANALYSIS 561 Box 25.1 Rank o
- Page 582 and 583: FACTOR ANALYSIS 563 Box 25.3 The st
- Page 584 and 585: FACTOR ANALYSIS 565 Box 25.5 Ascree
- Page 586 and 587: FACTOR ANALYSIS 567 Box 25.7 The ro
- Page 588 and 589: FACTOR ANALYSIS 569 The school
- Page 590 and 591: FACTOR ANALYSIS: AN EXAMPLE 571 sco
524 QUANTITATIVE DATA ANALYSIS<br />
Box 24.19<br />
Difference test for a paired sample<br />
Paired samples test<br />
Paired differences<br />
95 % confidence<br />
interval of the difference<br />
Mean SD SE mean Lower Upper t df Sig. (2-tailed)<br />
Pair 1 Maths-Science 14.45 27.547 0.871 12.74 16.16 16.588 999 0.000<br />
Box 24.20<br />
Effect size in analysis of variance<br />
Maths<br />
ANOVA<br />
Sum of squares df Mean square F Sig.<br />
Between groups 7078.619 3 2359.540 4.205 0.006<br />
Within groups 337266.2 601 561.175<br />
Total 344344.8 604<br />
differential weightings due to sample size<br />
variations. The two most frequently used indices of<br />
effect sizes are standardized mean differences and<br />
correlations (Hunter et al. 1982:373),although<br />
non-parametric statistics, e.g. the median, can be<br />
used. Lipsey (1992: 93–100) sets out a series of<br />
statistical tests for working on effect sizes, effect<br />
size means and homogeneity.<br />
Muijs (2004: 126) indicates that a measure of<br />
effect size for cross-tabulations, instead of chisquare,<br />
should be phi, whichisthesquarerootof<br />
the calculated value of chi-square divided by the<br />
overall valid sample size. He gives an example:<br />
‘if chi-square = 14.810 and the sample size is 885<br />
then phi = 14.810/885 = 0.0167 and then take<br />
the square root of this = 0.129’.<br />
Effect sizes are susceptible to a range of<br />
influences. These include (Coe 2000):<br />
<br />
Restricted range: the smaller the range of scores,<br />
the greater is the possibility of a higher effect<br />
size, therefore it is important to use the standard<br />
<br />
<br />
deviation of the whole population (and not just<br />
one group), i.e. a pooled standard deviation,<br />
in calculating the effect size. It is important to<br />
report the possible restricted range or sampling<br />
here (e.g. a group of highly able students rather<br />
than, for example, the whole ability range).<br />
Non-normal distributions: effect size usually<br />
assumes a normal distribution, so any<br />
non-normal distributions would have to be<br />
reported.<br />
Measurement reliability: the reliability (accuracy,<br />
stability and robustness) of the instrument<br />
being used (e.g. the longer the test, or the more<br />
items that are used to measure a factor, the<br />
more reliable it could be).<br />
There are downloadable software programs<br />
available that will calculate effect size simply by<br />
the researcher keying in minimal amounts of data,<br />
for example:<br />
<br />
The Effect Size Generator by Grant Devilly:<br />
http://www.swin.edu.au/victims