Year 2 SPSS Cheat Sheet - A years worth of notes on \"How To\" work SPSS for the coming exam. PDF

Preview

Summary

A years worth of notes on "How To" work SPSS for the coming exam....

Description

* P value must be less than or equal to 0.05 = significant * If the “observed” value is greater than the “critical” value, then the result is statistically significant * Never report p as .000 always .001 For questions were you aren’t allowed to use SPSS * when need to find F= mean squares/mean square error * when need to find the tabled critical F value. Start with the critical values of F at the 5% level. If the observed value (the calculated F) is larger than the table value, look at the 1%, if still bigger look at the 0.1%. If from the beginning, the calculated F is smaller than the table/critical value, then always use the 5% table. If the calculated F is always smaller than table value = not significant * within participants = repeated measure → same people repeat in each condition * between participants = non-repeated measure → different people in each condition Key Terms: P value = Has to be less than 0.05 (this is the main one we look at) Sig = This is what the P value is called in the table F = F – Ratio, this is calculated by the total average score of the conditions and is then divided by the error score. K = Means the number of conditions Error = This looks at how the scores between the groups vary DF = Degrees of Freedom DF1 = Numerator DF for F ratio DF2 = Denominator DF for F ratio (this is the error DF) X > Y = X is greater than Y X < Y = X is less than Y ≥ = Equal to or Greater than ≤ = Equal to or Less Than

Week 1: Setting up SPSS ● ● ● ●

Number of rows must equal number of ss. Once set up, go to analyse, general linear model, univariate. IV into fixed factors, DV into DV. F is the test statistic, calculated/ observed value by looking up critical value. Observed value must be greater than critical value to be significant.

AVOVA - able to look for polynomial contracts; linear, quadratic, cubic - When 2 conditions/levels of the IV, able to judge whether those scores are significantly different

-

Able to tell you whether the influence of 1 variable is modulated by changes in other variables → if where are interaction effects - can have more than one sig trend. We can compare main effects by plotting both onto the graph and seeing which is steeper. EXAMPLE: - IV: 4 levels/categories of driving conditions (day/night + clear/foggy) - DV: accuracy of distance estimate (1-10) Week 2: Planned (priori - designed before the experimenter see data for experiment) Unplanned (post hoc - devised after experimenter looks at data) - To determine where the difference shows from an overall effect, we need to do contrast tests. Two-way ANOVA 1. Set up data like it looks on the table provided 2. Analyse → general linear model → univariate 3. Insert the DV, and insert the IV as fixed factor, by using the arrows 4. Optional: plots → add IV into horizontal axis → add → continue * if looking at a 2 or 3 way interaction, the factor with most levels eg. DURATION will go in the ‘horizontal axis’ box 5. OK 6. Look at the ‘test of between-subjects effects’ output box

1st column = Alcohol; 24 1’s, 24 2’s 2nd column = Task; 2+ 2- 2>, 2+ 2- 2>.... 3rd column = Duration; 12 1’, 12 2’ … 4th column = score; 2,0,3,4,6,8,4,2… Columns 1-3 are the ‘factors’ that divide the sample up

Planned Contrast 1. Set up ANOVA normally 2. Paste 3. Edit syntax by adding: /contrast(x)=special(weighting) 4. Look for sig and F value. Post hoc/ Unplanned Contrast → Scheffe Test - Reduce the likelihood of type 1 error by increasing the critical value of the test statistic. Type 1 error is a false positive. Type 2 error is a false negative.

-

Prevent risking saying you’ve found an effect when you have not.

1. 2. 3. 4. 5.

Set up ANOVA normally Paste Edit syntax by adding: /contrast(x)=special(weighting) Look at the F value, ignore the sig. column Scheffé criterion aka Critical F value = (k - 1) * Tabled value of Fk-1, error df - K = number of levels/ conditions - Error DF (‘test results’ output - source, error, DF) = use this to look at the y axis on the F table Look at F value table of the 5% significance level; K= x axis, Error DF= y axis → table value (k-1) * table value = critical F-ratio If the calculated F-ratio (in the original ANOVA/’test results’ table) is greater than the critical F-ratio (scheffe) in the 5% table, look at the 1% table, if still greater, look at the 0.1% table Significant = calculated (original) is bigger than critical (scheffe)

6. 7. 8.

9.

Week 3 : Repeated measures ANOVA & Sphericity - If ps are tested twice or more. - A regular ANOVA assumes the 2 samples are independent, so must do a repeated measures ANOVA “t” and “F”… they are always significant whenever the other is. They are different versions of the same test! - If p = 0.00 that means; “The probability of this results happening is zero” “This result can never happen” hence, impossible Sphericity assumption → reliability of F ratios depend on the extent to which we are not affected by our individual differences. We assume the manipulation is the same for all P. Standard deviations of columns will be equal. - If p-values and DF are the same for all sphericity tests = there is sphericity - If p-values and DF are different for sphericity = there is a violation of sphericity report the Huynh-Feldt *sphericity is only for REPEATED MEASURES AND HAS 3 or MORE levels -

To assess sphericity: Mauchly test, but is inaccurate So use Greenhouse/ lower bound instead; the worse case scenario is the assumption that each participant is affected entirely differently by the manipulation Can see the difference between tests in the DF Best case scenario: assume sphericity → All ps equally affected by levels of IVs. Worst case scenario: Lower bound or Greenhouse-Geisser Conservative test. Not usually this bad tho. Increase chance of type 2 error. Intermediate scenario: Use Huynh-Feldt as offers the best compromise between likelihood of making type 1 or 2 error.

Only true for non-repeated measures: - DF1* is the same as DF1 - DF2* is the same as DF 2 (remember 2 = the denominator aka the error) Repeated measures: - (DF1 + DF2)/k-1 - K = number of levels of the repeated measures variable Repeated Measures ANOVA: (finds F) 1. Columns are the IV (eg. 3 different delays), rows are the participant performances 2. Analyse → general linear model → repeated measures 3. Rename within subjects factor name 4. add number of levels (amount of columns) → Add 5. Click Define 6. Move variables from right box, to the left hand box that corresponds 7. Options → tick descriptive statistics → continue 8. (optional) Pots → move the factor into the horizontal axis → add → continue 9. OK 10. Look at ‘tests of within subjects effects’ table, and look at the Huynh-Feldt Paired sample T test: (finds T, T squared is F) 1. 2. 3. 4. 5.

Paired samples t test- only 2 groups (do not put participants in a row). Not more. Analyse → compare means → paired samples t test Add the one pair OK Describe the nature of this relationship if sig… report lower and upper confidence intervals.

*One way repeated measures anova is the same as a paired sample t test *don't put a column for participants _______________ Week 4: Repeated measures, ANOVA mixed measures * number of rows = number of ss. - Repeated measure is when the ss provides data for each of the levels. Mixed measures ANOVA (within and between) 1. Set up data (when adding data, the non-repeated measure should only take up 1 column) 2. Analyse → general linear model → repeated measures 3. Type name in capitals into the ‘within subjects factor name’ box (usually participant/gender) 4. Type how many levels there are (columns not including the within subjects factor aka participant/gender)

5. Move the variables from left hand box into the appropriate one in the right hand box, and move the between subject factor into its box 6. Plots → add each main effect into the horizontal axis individually → add & do the same for the interaction between the two main effects (repeated measure goes in H axis as it has the most number of levels, non repeated in seperate lines) 7. Continue → OK 8. Look at the ‘within-subjects effect’ *non-repeated = between *repeated = within Within subjects contrasts Polynomial contests are meaningless for non-quantitative data. This may be harmful was the order if important. Not understand if the trends are linear or quadratic are significant because of the order. May be misleading information. Look at the nature to see if they are ordered or not. Contrasts…. 1. Add data → general linear model → repeated measures 2. Change within subject factor name in caps → add number of levels → define 3. Move the variables from left hand box into the appropriate one in the right hand box, and move the between subject factor into its box 4. Delete ‘polynomial’ from automatic generated syntax 5. add in ‘special’ and in brackets, add your contrasts. They are squared of the number of levels you have - eg. if you have 3 levels, you’ll have 3 sets of 3 for these contrasts. 6. This will then look like: /WSFACTOR=IV 3 special(1,1,1,-2,1,1,0,0,0) 7. For the 1st set, ALWAYS all 1s (1,1,1…..) 8. For the first contrast weights, put this in the middle of the 9 weights. So it will be (1,1,1,1,0,-1,0,0,0) if want to compare the first to last. 9. If you want to do an additional contrast, use final 3 eg. (1,1,1,-2,1,1,0,-1,1) * If you do not want a second contrast fill with 000. 1. Run the analysis 2. Look at the “tests of within subject contrast” Reading the analysis L1 is SPSS label given to the first contrast. F values and significance is given. L2 is the second contrast information. (the end numbers) Sheffé for mixed Post hoc can be calculated the same way as for non-repeated contrasts. Non repeated contrasts can be carried out using previously learnt method. Error

● ● ● ● ●

Is this the only thing that effected the scores? Is the IV valid? There is always a difference in scores (variance) that can be explained by the factor differences between groups. There will also be differences between scores (variance) that cannot be explained by the factor eg. differences within groups “within group” variance is what we call error F ratio is the between group variance divided by the within group variance.

When you notice that there are various degrees of freedom, assume that sphericity is violated. For its adjustments, you need appropriate ate numerator and denominate DF, so you can look up F ratio critical values to make the adjustment. - Type 1 error: when you say it is significant when it is not ___________________________________ Week 5: Fixed vs Random effects -

●

●

● ●

Differences between group CAN be explained by the factor Differences within groups CANNOT be explained by the factor → this within groups variance = error F ratio is the between groups variance divided by the within groups variance

Fixed effects analyses are used when you want to know whether the individual and precise conditions tested have an effect on performance Eg. gender; male vs female Eg. treatment; CBT vs Mindfulness vs No therapy Eg. primed condition; aggressive vs friendly vs no prime Eg. location; london vs manchester vs exeter Random effects analyses are used when the values of the experimental conditions can be regard as having been sampled at random from a wider population of different -- generalise the findings cause interested in the dimension eg. delay as a whole --- usually has more levels Eg. time Eg. not interested in just the delay of 20 and 40 sec, but might also be interested in delay in between eg. 10, 30, 50 sec = random In almost all circumstances treat variables and fixed effects Subjects are always treated as a random effect ~ SPSS does this automatically in repeated measures designs

Effect size and statistical power: - Expressed in terms of a probability value such that: - Power = 1.0 implies there is 100% chance the test will detect the effect

-

- Power = 0.0 means there is no chance of detecting the effect. The stronger the influence of IV on DV, the larger is the effect size: - The bigger differences between groups - Means more powerful analyses = more likely you are to find a statistically significant, genuine effect - The more data points (eg participants) in each group, the more powerful analysis

1. 2. 3. 4. 5.

After inputting data, and getting the box with arrows Options → tick ‘estimates of effect size’ and tick ‘observed power’ Continue OK Report the: ‘Partial eta squared’ (ƞ2) after the p-value for each test statistic - ‘Estimated power’ in the text of your Results section ________________________ Week 6: self tests Week 7: report writing _________________________ Week 8: Regression -

A correlation measures the “degree of association” between two variables (interval or ordinal) - Associations can be: - positive (an increase in one variable is associated with an increase in the other) - negative (an increase in one variable is associated with a decrease in the other)

-

Regression is able to investigate as continuous variables, whereas ANOVA sees the IV’s as category variables only

Line of best fit - allows us to describe relationship between variables more accurately. - We can now predict specific values of one variable from knowledge of the other - All points are close to the line Y = bx + a Y= DV b = line of best fit X = IV a = y-intercept -

“Best-fit” line same as “Regression” line m is the “regression coefficient” for x

-

x is the “predictor” or “regressor” variable for y

-

2-valued categorical variables (dichotomies) can be used directly as regressors (e.g. gender) k-valued categorical variables (“k” being a whole number greater than 2) are dealt with by using dummy variables (we’ll deal with this later in the course)

-

Regression Line 1. Analyse → regression → linear _____________________________ Week 9: Simple and multiple regression Simple regression R2 - “Goodness of fit” R2 is the square of the correlation coefficient - Reflects variance accounted for in data by the best-fit line - Takes values between 0 (0%) and 1 (100%) - Frequently expressed as percentage, rather than decimal - High values show good fit, low values show poor fit - If R2 is 0 - implies that a best-fit line will be a very poor description of data - Less variance explained - If R2 is 1 - Implies that a best-fit line will be a very good description of data - Lots of variance explained - uses a t-test to establish whether or not the model describes a significant proportion of the variance in the data (reported in the SPSS output) T-test 1. 2. 3. 4. 5.

Analyse → compare means → independent samples t-test ‘Test variable’ = DV you want to test for differences, use arrow to move over ‘Grouping variable’ = IV eg. gender, use arrow to move over Define groups Type number that you use to identify the ‘grouping variable Eg. male = 1, female = 2 SO, group 1 = 1, group 2 = 2 6. Continue → OK

Regression equation y=Bx+c - B is shown in the ‘coefficients output box’ after doing a regression, substitute this into the equation - C is also shown in the ‘coefficients output box’ after doing a regression, substitute

SO; y = 844.670 X - 4239.600 If X = time, and question is asking at 9:45 am… regression equation would be: Y = 844.670 (9.75) - 4239.600 Multiple regression - uses an Analysis of Variance to discover if the proportion of variance in the data explained by the model is significant

1. 2. 3. 4.

Analyse → regression → linear Move DV into DV box using arrow, more all IV’s into the IV box using arrow OK Look at ‘adjusted R2’ value (bc it takes into account the number of regressors (IV’s) VS. the R2 gets bigger with every regressor added, so not so good

___________________________ Week 10: Reporting multiple regression and Model checking; outliers and residuals -

Would ideally want the residual value to = 0, indicating the model perfectly predicts the outcome. The larger the residual value, the more inaccurate the model is

-

If R2’ = 0.520 aka 52% = means that 52% of the variance is explained by the number of variables/regressors together VS. is ‘adjusted R2’ = 0.399 aka 39.9% = means there is a 12% difference between R2 and adj R2, showing that the R2 is less accurate, as there are 1 or more variables in the R2 model that doesn’t explain the outcome well

-

-

In the ‘coefficients’ outcome table; Column headed Unstandardised coefficients - B - Gives regression coefficient for each regressor variable (IV) - “With all the other variables held constant” - Units of coefficient are same as those for regressor (IV) - For a one unit increase in eg. age the change in score (Y) would be the value of B eg. 8.42 - For one unit increase in gender (0=male, 1=female, and the gender unstandardized B is -8.42. As gender increases (female) their salary wil decrease by 8.42. - Eg. gender coded as 1 (male) 2 (female) - will show how much y changes when there is this increase in one value - ** for every increase of 1 (hour aka the X), we would predict that score (Y) will increase by B

R2’ - report all decimals but when reporting as a percentage, report to 1 decimal place Adj R2’ - report all decimals but when reporting as a percentage, report to 1 decimal place F - report to 2 decimal places

Sig (P) - report to 3 decimal places Coefficients - report to 2 decimal places - 1.770-E04 = 1.770x10^-4 = 0.0001770 = 0.0002 - Coefficient can never be 0, double click and true value will appear Residual = actual/observed value - predicted/calculated value (line of best fit) - Residuals should be normally distributed; most small and close to zero Outliers - To be termed an outlier it must be a qualitative difference - it shows a different model - Large residual value - Are NOT extreme scores - extreme cases only show a quantitative difference eg. change in significance Removing outliers 1. Run regression w ‘casewise diagnostics’ 2. Delete DV score (1 cell) for these cases, so the whole row won’t disappear 3. Run the regression again 4. If there is a big difference in the output aka a qualitative difference = those cases that you deletes are outliers BUT if only demonstrate a quantitative change (eg. slight change in significance or correlations) = not an outlier, just extreme cases Distribution of residual values - If our line of best fit does not fit too well, this will be revealed in the ‘distribution of the residuals’ - So better to show a histogram; it checks that the frequency distribution for the residuals resembles a normal distribution 1. Plots → histogram - If the bars approximately match the line = all is well, if not, investigate ‘casewise diagnostics’ Residuals analysis - “An ex-employee claims that the firm breaches sex-discrimination legislation by favouring male employees, this discrimination being reflected in the relative salaries of men and women within the company.” - “Is there any truth behind these allegations?” - Run a regression with casewise diagnostics and plot a histogram The circled case is isolated from the main distribution of residuals, so may be an outlier - A residual value of zero means that the model perfectly predicts the actual salary -A positive residual means that the actual salary is higher than we might expect

A consistent relationship between residual values and IV/DV - is a sign of a problem.. They are outliers.

“Residual analysis” 1. Run a regression 2. Statistics → casewise diagnostics → change SD to 2 → tick ‘standardized’ under residuals 3. Plots → tic both histogram and normal probability plot with ZRESID in Y axis and ZPRED in X 4. Continue → OK _____________________ Week 11: model choice What makes the best model when: 1. Samle # of regressors - Chose model w (2nd) highest R2’ adjusted value - Highest F value 2. Different...