OPTIONAL / ADVANCED
Not part of basic masteryGo Further
Advanced tools for students who want to go beyond the required course minimum. You do not need these topics for the basic task. They explain ideas that already appeared in some homework submissions and questions.
Progress
Advanced exploration
No extra grade. Core course mastery remains separate.
Optional / advanced
From omnibus χ² to cell-level interpretation
You do not need this for the basic task. This section explains what some of your classmates already explored.
| Age | Outcome: Yes | Outcome: No |
|---|---|---|
| 18–29 | O = 42 E = 31.0 Residual = 11.0 ASR = 3.48 | O = 18 E = 29.0 Residual = -11.0 ASR = -3.48 |
| 30–49 | O = 31 E = 31.0 Residual = 0.0 ASR = 0.00 | O = 29 E = 29.0 Residual = 0.0 ASR = 0.00 |
| 50+ | O = 20 E = 31.0 Residual = -11.0 ASR = -3.48 | O = 40 E = 29.0 Residual = 11.0 ASR = 3.48 |
Read the workflow in order
- χ²: something differs somewhere in the table.
- Cramer’s V: how strong is the overall association?
- Percentages: what is the descriptive pattern?
- Residuals: which cells depart most from independence?
p < .001
Cramer’s V = 0.30
Optional / advanced
Pairwise / 2×2 follow-up comparisons
You do not need this for the basic task. This section explains what some of your classmates already explored.
Start with the 5×2 table
Select two categories
Teaching assumption: each age group has 100 respondents, so the percentages are also the counts used for χ².
χ²(1) = 29.53
p < .001
A significant larger table does not tell you which pairs differ. A focused 2×2 follow-up can help, but repeated testing raises false-positive risk. Prefer hypothesis-driven comparisons over “test every pair until something is significant.” Multiple-testing correction is an optional next step, not a basic requirement here.
Optional / advanced
Trend vs association
You do not need this for the basic task. This section explains what some of your classmates already explored.
Optional / advanced
Survey weights: who does each case represent?
You do not need this for the basic task. This section explains what some of your classmates already explored.
| Case | Group | Outcome | Weight |
|---|---|---|---|
| A | Urban | Yes | 0.7 |
| B | Urban | Yes | 0.8 |
| C | Urban | No | 0.8 |
| D | Rural | No | 1.6 |
| E | Rural | No | 1.5 |
| F | Rural | Yes | 1.6 |
Ask before weighting
- What population are we trying to describe?
- What does the weight represent?
- Was weighting intended for this analysis?
WEIGHT BY dweight. CROSSTABS ... WEIGHT OFF.
In SPSS: Data → Weight Cases. Weight status can remain active and silently affect later N, percentages and estimates. Reproducibility rule: always know whether weighting is active.
Optional / advanced
Lambda, PRE and Gamma
You do not need this for the basic task. This section explains what some of your classmates already explored.
Lambda = proportional reduction in error
First predict Y without knowing X. Then predict Y using X. Count how many classification errors remain.
Knowing X reduces prediction error for Y by 25%.
Cramer’s V describes overall association strength. Lambda describes prediction-error reduction. They are not interchangeable.
Gamma = direction among ordered pairs
For ordinal variables, concordant pairs move in the same direction; discordant pairs move in opposite directions.
Higher X tends to occur with higher Y.
A variable is not ordinal merely because its categories happen to be coded 1 and 2. Ask: are the categories substantively ordered?
Optional / advanced
Choosing association measures
You do not need this for the basic task. This section explains what some of your classmates already explored.
Method chooser
Lambda can be useful if prediction is substantively meaningful.
“Why did you use this?” challenge
/STATISTICS=CHISQ PHI LAMBDA GAMMA
Suppose both variables are nominal and prediction is not part of the research question. Which requested statistics are substantively justified?
Optional / advanced
p-value: test statistic, reference distribution, simulation
You do not need this for the basic task. This section explains what some of your classmates already explored.
Hold two things constant, change the third
t = 5.0 / 3.16 = 1.58
df = 39
p = .122
Test statistic = observed deviation from H₀ / standard error. The numerator is the observed effect in original units. The denominator is expected sampling variability. The ratio says how many standard errors the observation lies from H₀.
Simulation view: assume H₀ is true
Draw 500 simulated samples under H₀ and count how often |tsim| ≥ |tobserved|.
Larger effect usually lowers p; larger N usually lowers p; larger noise usually raises p. But p is not effect size. Pair χ² with Cramer’s V, t-tests with Cohen’s d / Hedges’ g, while correlation r already carries strength information.
Optional / advanced
Reproducible SPSS: paste, inspect, run
You do not need this for the basic task. This section explains what some of your classmates already explored.
Prefer this workflow
FREQUENCIES VARIABLES=age outcome. MISSING VALUES age outcome (99). RECODE age (18 THRU 29=1) (30 THRU 49=2) (50 THRU HI=3) INTO age3. COMPUTE eligible = (age >= 18). TEMPORARY. SELECT IF eligible = 1. CROSSTABS /TABLES=age3 BY outcome /STATISTICS=CHISQ PHI /CELLS=COUNT EXPECTED ROW ASRESID. WEIGHT BY dweight. CROSSTABS /TABLES=age3 BY outcome. WEIGHT OFF.
Useful building blocks: FREQUENCIES, MISSING VALUES, RECODE, COMPUTE, TEMPORARY, SELECT IF, WEIGHT BY, WEIGHT OFF, CROSSTABS.