The Empathy
Spectrum
How empathetic are people, and does it vary by who they are? This study applied the Toronto Empathy Questionnaire to 260 participants (from 277 collected) to map empathy as a measurable construct — using advanced statistical methods to find patterns, profiles, and design opportunities hidden in the data.
Overview
What does it mean to measure empathy scientifically — and why does it matter for design?
Empathy is the ability to sense and share what others feel. This study measured it using a validated 16-question survey — the Toronto Empathy Questionnaire (TEQ) — with 277 people. We then applied 14 different statistical techniques to understand: how empathetic people are, whether there are distinct empathy "types," and what that means for designing products that work for real humans.
The TEQ, developed by Spreng et al. (2009), is a 16-item self-report measure that captures empathy as a unified construct. Responses span a 5-point scale from Never (0) to Always (4), with 8 items reverse-scored to control for acquiescence bias. Total scores range from 0 to 64, with higher scores indicating greater empathic tendency.
This research is part of a broader UX research initiative at Dhirubhai Ambani University, under the Human Factors in Interaction Design course. The goal is to derive empirically-grounded design directions for emotionally responsive digital products.
The analysis pipeline covers 14 analytical methods — from descriptive statistics to network psychometrics, IRT, and latent profile analysis — producing 20 interactive charts and 10 design directions. Key finding: the TEQ captures two empathy dimensions in this sample rather than one, and the population splits into two empirically-distinct profiles with a moderate gender effect (r=0.287). HDBSCAN on UMAP confirms a third low-reactive subgroup (N=31, ~12%, mean TEQ≈32) providing empirical basis for a fourth Minimal design mode.
Methodology
The instrument, scoring procedure, and the full analytical pipeline used in this study.
We used a 16-question survey where participants rated how often they experience empathic thoughts and feelings. Some questions are phrased negatively (e.g., "I am not interested in how others feel") — these are reverse-scored so that a higher score always means more empathy. The data was then run through 14 different statistical tests, each answering a different question about the nature of empathy in this sample.
The Instrument
| # | Item | Direction |
|---|---|---|
| Q1 | When someone else is feeling excited, I tend to get excited too | Forward |
| Q2 | Other people's misfortunes do not disturb me a great deal | Reverse |
| Q3 | It upsets me to see someone being treated disrespectfully | Forward |
| Q4 | I remain unaffected when someone close to me is happy | Reverse |
| Q5 | I enjoy making other people feel better | Forward |
| Q6 | I have tender, concerned feelings for people less fortunate than me | Forward |
| Q7 | When a friend talks about problems, I steer the conversation elsewhere | Reverse |
| Q8 | I can tell when others are sad even when they do not say anything | Forward |
| Q9 | I find that I am "in tune" with other people's moods | Forward |
| Q10 | I do not feel sympathy for people who cause their own serious illnesses | Reverse |
| Q11 | I become irritated when someone cries | Reverse |
| Q12 | I am not really interested in how other people feel | Reverse |
| Q13 | I get a strong urge to help when I see someone who is upset | Forward |
| Q14 | When I see someone treated unfairly, I do not feel very much pity for them | Reverse |
| Q15 | I find it silly for people to cry out of happiness | Reverse |
| Q16 | When I see someone being taken advantage of, I feel kind of protective | Forward |
Analytical Pipeline
Data Cleaning & Preprocessing
Encoding responses (Never=0 → Always=4), reverse scoring of 8 negatively-worded items, removal of incomplete responses, and computation of total and mean TEQ scores.
Exploratory Data Analysis (EDA)
Descriptive statistics, item-level frequency heatmaps, inter-item correlation matrices, and demographic breakdowns.
Reliability Analysis
Cronbach's α, McDonald's ω, split-half reliability (Spearman-Brown corrected), item-total correlations, and α if item deleted.
Normality Testing
Shapiro-Wilk test, Kolmogorov-Smirnov test, Jarque-Bera test, and Q-Q plots for each item and the total score.
Exploratory & Confirmatory Factor Analysis
KMO, Bartlett's test, scree plot, parallel analysis, oblimin-rotated EFA; followed by CFA to confirm the 1-factor TEQ structure with fit indices (CFI, TLI, RMSEA, SRMR).
Item Response Theory — Graded Response Model
GRM parameter estimation (discrimination a, difficulty thresholds b) for each polytomous item; item and test information functions plotted across the latent empathy continuum.
Measurement Invariance
Sequential testing of configural → metric → scalar invariance across gender groups using ΔCFI and ΔRMSEA criteria.
Differential Item Functioning (DIF)
Mantel-Haenszel statistic per item; ETS Δ effect size classification (A/B/C); Bonferroni-corrected p-values.
Network Psychometrics
EBICglasso partial correlation network; node strength, betweenness, and closeness centrality; bridge item identification.
Latent Profile Analysis
Gaussian Mixture Model (K=2–6); BIC/AIC/Silhouette model selection; profile characterisation and gender cross-tabulation.
UMAP + t-SNE Clustering
Dimensionality reduction of 16-item responses into 2D; K-Means and HDBSCAN clustering; silhouette and Calinski-Harabász validation.
Hypothesis Testing
8 formal hypotheses tested; Bonferroni correction for multiple comparisons; effect sizes reported for all significant findings.
Participants
Who responded? Demographic breakdown of the 277 participants.
Male participants outnumber female participants 57% to 41% in this sample. This matters for interpretation: the gender empathy difference found later in the analysis (females scoring ~4 points higher) is even more striking because it occurs despite males being the majority group. The small "Other / Prefer not to say" category (N=4) is kept separate throughout to avoid misclassification.
The distribution is right-skewed — most respondents are 20–25 (the typical university age range at DAU), with a small tail of older participants pulling the mean up to 24.6. This narrow age band largely explains why age turned out to be a non-significant predictor of empathy (β=−0.051, p=0.381): there simply isn't enough age variation in the sample to detect a meaningful relationship, even if one exists in the broader population.
Exploratory Data Analysis
A first look at the data — score distributions, how each question was answered, and how different questions relate to each other.
Before running complex statistics, we look at the raw patterns. How did people score overall? Which questions got the most "Always" or "Never" responses? And are questions about different aspects of empathy correlated — meaning do people who score high on one tend to score high on others? This section answers all three.
Scores cluster around 43.48 — slightly below the published TEQ norm of ~47 for university students. The distribution is approximately bell-shaped with a mild negative skew (−0.12), meaning slightly more participants score above the mean than below. The range of 21–63 (out of 0–64 possible) shows healthy variance, with no floor or ceiling effects that would limit the scale's ability to distinguish between participants.
Q5 ("I enjoy making others feel better," M=3.18) and Q7 ("I try to steer away from friends' problems," M=3.13) score highest — a mix of prosocial helping and empathy inhibition. Q9 ("I'm in tune with others' moods," M=2.12) is the lowest-scoring item, suggesting that passive emotional attunement is weaker than active helping behaviour in this sample. Items Q2 (M=2.24) and Q10 (M=2.26) also score low and are the two weakest contributors to reliability (item-total r = 0.22 and 0.17 respectively).
Darker blue cells indicate stronger positive correlations. A perfectly unidimensional empathy scale would show uniform, moderate correlations across all items. The actual pattern here is uneven — some pairs correlate strongly (particularly the prosocial cluster: Q3, Q5, Q13) while others barely correlate at all (Q2 and Q10 with most items). This uneven texture is the visual fingerprint of the two-factor structure confirmed in the EFA: two groups of items that are internally coherent but only moderately related to each other.
Mean TEQ = 43.48 — below published norms (M ≈ 47 for university samples). Q10 and Q2 show the lowest inter-item correlations. The distribution is slightly left-skewed but approximately normal overall. No items show correlations below 0.10 after reverse-scoring.
Reliability Analysis
Is the TEQ consistent? Reliability tells us whether the scale is measuring empathy stably — not randomly fluctuating.
Think of reliability like measuring your height five times with the same ruler — you want consistent readings. For a questionnaire, reliability means people who are truly empathetic score high repeatedly across all questions. We test this using Cronbach's alpha (α) — a score above 0.70 is considered acceptable, above 0.80 is good, and above 0.90 is excellent.
Bars exceeding the red line indicate items that, if removed, would improve the scale's internal consistency. Only Q10 (α-if-deleted = 0.777) clearly crosses this threshold. All other items contribute positively — removing them would lower reliability. Q3 (α-if-deleted = 0.751) is the single most important contributor to reliability: removing it causes the largest drop in α, confirming it is the most central item in the scale — consistent with its network hub status.
Q2 (r = 0.215) and Q10 (r = 0.168) show item-total correlations below 0.30 — the accepted minimum. Both are reverse-scored items measuring empathy inhibition. Their removal would marginally increase α (to 0.772 and 0.777 respectively). McDonald's ω > 1 from semopy is a computation artefact; treat reliability as α = 0.772.
Normality Testing
Do the scores follow a normal "bell curve"? This determines which statistical tests are valid downstream.
Many statistical tests assume scores are distributed like a bell curve — most people in the middle, fewer at the extremes. If scores are very skewed (e.g., most people scoring very high), some tests become unreliable. We formally test this assumption for every item and the total score using three different normality tests.
A Q-Q plot checks normality by comparing your data's quantiles to a perfect normal distribution — points on the diagonal line mean normal, deviations away from it mean non-normal. The TEQ total follows the line closely in the centre but curves slightly at both tails. This S-shaped tail pattern is consistent with mild negative skewness and slightly light tails. The distribution is close enough to normal for the analysis to proceed, but all group comparisons use non-parametric tests out of caution.
The bars (observed frequencies) and orange curve (expected normal distribution) are remarkably close — visually, the TEQ total is approximately normally distributed. The slight excess of scores in the 44–50 range above what the normal curve predicts reflects the mild negative skew (−0.12): a few more people than expected score just above the mean. For a scale with 260 respondents across 16 5-point items, this level of near-normality is typical and expected.
Normality Test Results
| Variable | Shapiro-Wilk W | SW p-value | Skewness | Kurtosis | Normal? |
|---|---|---|---|---|---|
| TEQ_Total | 0.9891 | 0.048 | −0.123 | −0.584 | Marginal |
| All 16 items fail normality (SW p < .001). TEQ Total: SW W=0.989, p=0.048 — barely non-normal. All group comparisons use non-parametric methods (Mann-Whitney U, Kruskal-Wallis). | |||||
All 16 items and TEQ_Total fail the Shapiro-Wilk normality test. TEQ_Total is borderline (p = 0.048) with mild skewness (−0.12) and negative kurtosis (−0.58). All group comparisons therefore use non-parametric methods — Mann-Whitney U for gender and Kruskal-Wallis for age groups.
Exploratory Factor Analysis
How many underlying dimensions does empathy have — and are all 16 items measuring the same thing?
Imagine empathy isn't one single thing but potentially several related things — like "feeling what others feel" vs "wanting to help when someone is upset." Factor analysis discovers how many such underlying "factors" explain the patterns in the data. The TEQ is theoretically a single-factor scale, so we're checking whether the data agrees.
Each point on the blue line represents one potential factor, measured by how much variance it explains (its eigenvalue). The red dashed line is the parallel analysis benchmark — a simulation showing what eigenvalues would look like if the data were purely random. Only factors above the red line explain more than chance. Here, the first two factors cross this threshold, with a sharp elbow after Factor 1. The break is less dramatic than in a purely unidimensional scale — another signal that two factors are present in the data.
Factor loadings show how strongly each item belongs to the single "empathy" dimension. Q3 (0.646), Q5 (0.616), and Q13 (0.604) are the most reliable markers of empathy in this sample — and these are exactly the same items that emerge as network hubs, IRT high-discriminators, and items with the highest item-total correlations. Q10 (0.122) and Q2 (0.170) barely load onto Factor 1 — they load on Factor 2 in the 2-factor solution instead, measuring empathy inhibition rather than affective empathy.
Parallel analysis recommends 2 factors, not 1. Factor 1 (affective/prosocial empathy — Q3, Q5, Q6, Q8, Q9, Q13, Q16) explains 19.3% of variance; Factor 2 (empathy inhibition — Q4, Q7, Q10, Q11, Q12, Q14, Q15) adds another 14.0%, totalling 31.5%. Items Q10 (0.12) and Q2 (0.17) load poorly on Factor 1, and both load meaningfully on Factor 2 instead. This suggests the TEQ captures two related but distinct dimensions of empathy in this sample — consistent with the poor CFA fit for the 1-factor model.
Structural Equation Modelling (CFA)
Formally testing whether the single-factor structure proposed by the TEQ authors holds in this dataset.
Confirmatory Factor Analysis is like setting a hypothesis before looking at the data. We say: "We believe all 16 items measure one underlying thing called empathy." Then we test how well that model actually fits the data using four key fit indices — think of them as different ways to ask "does this model match reality?"
While most items show positive loadings onto the single EMPATHY factor, the overall model fit is poor (CFI=0.632, well below the 0.90 threshold; RMSEA=0.104, above the 0.10 ceiling for acceptable fit). This is consistent with the EFA evidence for 2 factors. Total TEQ scores can still be used as a rough empathy measure, but researchers should note that the unidimensional model does not hold in this sample.
Fit Index Benchmarks
| Index | Obtained | Excellent | Acceptable | Poor | Interpretation |
|---|---|---|---|---|---|
| CFI | 0.632 | >0.95 | >0.90 | <0.90 | Poor fit |
| TLI | 0.576 | >0.95 | >0.90 | <0.90 | Poor fit |
| RMSEA | 0.104 | <0.05 | <0.08 | >0.10 | Poor fit |
| GFI | 0.567 | >0.95 | >0.90 | <0.90 | Poor fit |
The 1-factor CFA shows poor fit (CFI=0.632, RMSEA=0.104) — both below their respective thresholds. This aligns with the EFA finding that 2 factors better explain the data. The TEQ, at least in this DAU sample, does not function as a fully unidimensional scale. Q2 and Q10 show near-zero standardised loadings (0.242 and 0.128), confirming they are the weakest items. This has implications for DD8: universal design claims based on a single empathy score should be made with caution. (pairs of items whose correlated residuals would improve fit — e.g., Q8 and Q9 both measure mood-detection). This does not invalidate the scale, but informs nuanced interpretation.
Item Response Theory (IRT)
Which questions are the best at distinguishing between people with different levels of empathy?
Not all questions in a survey are equally useful. Some questions are excellent at telling apart someone with very high empathy from someone with moderate empathy. Others are only useful for distinguishing very low-empathy people from the rest. IRT maps each question's discrimination (how sharp a separator it is) and difficulty (at what empathy level it becomes informative). We used the Graded Response Model, designed specifically for multi-category questions like this 5-point scale.
Each curve shows how much information (measurement precision) a question provides at each empathy level. Q3 (a=1.91) and Q5, Q13 (a≈1.82) peak sharply around moderate empathy levels — they are the scale's most precise instruments. Q10 (a=0.45) barely rises above zero, contributing almost no measurement precision across the entire empathy spectrum. The dotted line shows total test information (TIF), which peaks around θ=−1 to 0.
GRM Parameter Summary
| Item | Discrimination (a) | Threshold b1 | Threshold b2 | Threshold b3 | Threshold b4 | Interpretation |
|---|---|---|---|---|---|---|
| Q3 | 1.907 | −2.772 | −2.159 | −1.028 | 0.381 | Strong |
| Q5 | 1.820 | −3.096 | −2.429 | −0.979 | 0.079 | Strong |
| Q13 | 1.822 | −2.428 | −1.555 | −0.458 | 0.690 | Strong |
| Q16 | 1.335 | −3.071 | −1.867 | −0.471 | 0.827 | Good |
| Q6 | 1.219 | −2.503 | −1.517 | 0.049 | 1.443 | Good |
| Q14 | 1.107 | −3.393 | −2.429 | −1.226 | 0.332 | Good |
| Q8 | 1.075 | −3.043 | −1.851 | −0.231 | 1.704 | Good |
| Q9 | 0.931 | −3.241 | −1.462 | 0.850 | 2.920 | Moderate |
| Q1 | 0.883 | −3.716 | −2.285 | 0.265 | 2.717 | Moderate |
| Q7 | 0.877 | −5.933 | −3.323 | −1.370 | 0.185 | Moderate |
| Q11 | 0.813 | −4.988 | −3.009 | −0.989 | 0.776 | Moderate |
| Q12 | 0.822 | −4.160 | −2.543 | −1.029 | 0.862 | Moderate |
| Q15 | 0.594 | −5.881 | −3.788 | −1.539 | 0.140 | Weak |
| Q4 | 0.579 | −4.437 | −2.931 | −1.606 | 1.057 | Weak |
| Q2 | 0.474 | −5.901 | −2.759 | 0.901 | 4.395 | Weak |
| Q10 | 0.445 | −5.863 | −2.875 | 0.768 | 4.269 | Weak |
Three items exceed a=1.0 discrimination: Q3 (a=1.907), Q13 (a=1.822), Q5 (a=1.820). These are the most informative items for distinguishing empathy levels — ideal design triggers for empathy-responsive features. Q10 (a=0.445) and Q2 (a=0.474) are weak discriminators — consistent with their low item-total correlations and EFA loadings. — these are the most valuable questions). Items with low discrimination (a < 0.50) may contribute little to measurement precision. The test information function shows at which empathy level (θ) the TEQ is most precise overall.
Measurement Invariance
Can we fairly compare empathy scores across gender groups? Only if the questionnaire measures the same construct in the same way for everyone.
Imagine a weighing scale that displays kilograms for some people and pounds for others — you can't compare the numbers directly! Measurement invariance testing checks whether the TEQ "works the same" for male and female respondents. If it does, we can trust that a higher score genuinely means more empathy regardless of who answered. We test three levels of this equivalence, each stricter than the last.
| Model | What's Constrained | CFI | RMSEA | ΔCFI | ΔRMSEA | Decision |
|---|---|---|---|---|---|---|
| Configural — Male | 1-factor, male group only | 0.519 | 0.113 | — | — | Poor base fit |
| Configural — Female | 1-factor, female group only | 0.726 | 0.099 | — | — | Poor base fit |
| Metric / Scalar | Full multi-group MI requires lavaan in R (code provided in Section 10). Base model fit is poor for both groups, which limits the validity of formal invariance testing. | |||||
Metric holds if ΔCFI ≥ −0.010 and ΔRMSEA ≤ +0.015 (Cheung & Rensvold, 2002). Scalar holds under same criteria. If scalar fails, partial invariance (freeing 1–2 intercepts) should be tested before making cross-group score comparisons.
Configural fit is poor for both gender groups (Male CFI=0.519, RMSEA=0.113; Female CFI=0.726, RMSEA=0.099), meaning the 1-factor structure does not hold well in either group separately. A loading comparison across gender groups reveals that 12 of 16 items show large standardised loading differences (Δ>0.15) — Q3, Q4, Q6, Q7, Q8, Q9, Q11, Q12, Q13, Q14, Q15, Q16. Only Q1, Q5, and Q10 show small differences (Δ<0.08). This extensive loading non-invariance confirms the TEQ measures partially different constructs in male and female respondents. Design implication (DD3 + DD9): H2 gender difference (p=0.0004) is valid as a raw score comparison but universal design claims should carry the caveat that empathy items may tap different psychological constructs across gender groups. Localise emotional design per demographic segment.
Differential Item Functioning
Are any individual questions systematically easier or harder to agree with depending on gender?
DIF detects bias at the individual question level. Even if two people have the same overall empathy level, a biased question might consistently get a higher response from women than men (or vice versa) — not because of real empathy differences, but because of how the question is worded or interpreted. We use the Mantel-Haenszel method, the gold standard for detecting DIF in survey research.
Bars beyond the orange dashed lines indicate gender-biased items. 12 of 16 items show Class B (|Δ|>1.0) or Class C (|Δ|>1.5) DIF — an unusually high rate. Q7 (Δ=2.49) and Q4 (Δ=2.19) show the largest gender-differential. Direction verification confirms all positive ETS Δ values indicate female-favouring DIF — females score above the median more consistently at equivalent total empathy levels (15/16 items; only Q5 marginally favours males by −0.13). This challenges the assumption that the TEQ is gender-neutral in this population.
12 of 16 items show at least moderate DIF by gender — a remarkably high proportion. Class C (large) DIF items: Q4 (Δ=2.19), Q7 (Δ=2.49), Q15 (Δ=1.92), Q14 (Δ=1.70), Q12 (Δ=1.69), Q3 (Δ=1.50). All DIF favours the focal group (Female) — confirmed: 15/16 items show higher female means after reverse-scoring (Q5 is the sole exception, Δ=0.39 marginal). This means females respond more consistently with high empathy on these items after controlling for overall empathy level. Items Q7 ('steer away from friends' problems', Δ=2.49) and Q11 ('irritated when someone cries', Δ=1.45) show the largest differential — likely reflecting culturally-gendered norms around emotional support rather than pure empathy capacity. Gender-neutral design assumptions for empathy-activating content are empirically unsupported in this sample.
Network Psychometrics
Instead of treating empathy as a single score, network analysis reveals which aspects of empathy are most tightly connected — and which act as bridges between others.
Imagine empathy as a map of cities (items) connected by roads (correlations). Some cities are major hubs — everything passes through them. Others are quieter, connected to only a few routes. Network psychometrics builds this map using partial correlations (controlling for all other items) to find the genuine relationships, not just surface-level ones. The most central items are the ones that, if you changed them, would most change the entire empathy construct.
Q3, Q5, and Q13 form the network's core hub — they have the highest Strength scores (0.89, 0.89, 0.87) and connect most densely to other items. Q2 and Q10 are peripheral nodes with limited connections. The network visually splits into two clusters: a prosocial cluster (top-left, blue-dominated) and an inhibition cluster (bottom-right, orange-dominated), mirroring the 2-factor EFA structure.
Centrality Indices
| Item | Strength | Betweenness | Closeness | Expected Influence | Role |
|---|---|---|---|---|---|
| Q3 | 0.894 | 0.067 | 0.625 | 0.894 | Hub — most central item |
| Q5 | 0.888 | 0.124 | 0.536 | 0.888 | Hub |
| Q13 | 0.871 | 0.000 | 0.600 | 0.871 | Hub |
| Q12 | 0.816 | 0.000 | 0.536 | 0.816 | Strong connector |
| Q14 | 0.807 | 0.138 | 0.682 | 0.807 | Bridge item |
| Q10 | 0.518 | 0.338 | 0.577 | 0.402 | Bridge — high betweenness |
| Q2 | 0.349 | 0.057 | 0.500 | 0.349 | Peripheral |
Q3 ('It upsets me to see someone treated disrespectfully') has the highest Strength (0.894) and is the network hub. Q5 (0.888) and Q13 (0.871) follow. These three items are tightly interconnected and form the core of the empathy construct in this network. Q10 has the highest Betweenness (0.338) despite low overall strength — making it a structural bridge between network clusters. Q2 has the lowest Strength (0.349), confirming it is a peripheral item. These are the "core" of empathy in this network — the aspects that connect most tightly to all other aspects. They have the highest design leverage: features that activate these items will ripple through the whole empathy construct.
Latent Profile Analysis
Discovering natural empathy "types" in the population — empirically-grounded user archetypes.
Rather than treating everyone as a point on a single scale, Latent Profile Analysis (LPA) asks: are there distinct groups of people who show similar patterns of empathy across all 16 questions? It's like discovering that people don't just vary continuously but actually cluster into meaningful types — "The Resonator," "The Pragmatist," "The Detached" — each with a different empathy signature across dimensions.
The BIC (blue line) reaches its minimum at K=2, indicating two profiles provide the best model fit without overfitting. The AIC continues declining to K=5, but AIC is known to overselect factors in smaller samples. The silhouette score (orange) also peaks at K=2 (S=0.13), though the overall silhouette values are modest — indicating the two profiles overlap somewhat, as expected for a continuous trait like empathy.
Profile 0 — The Resonator (blue, N=146) consistently scores higher than Profile 1 — The Pragmatist (orange, N=114) across every single item. The largest gaps appear on Q14 (+1.09), Q15 (+1.08), Q6 (+0.75), and Q5 (+0.80). Crucially, both profiles have the same shape — the same items rank highest in both — confirming these profiles differ in empathy intensity, not in which type of empathy they show.
Profile 1 — The Resonator (N=146, 56%, M=48.56): High scores across all items, especially Q5 (3.53), Q14 (3.52), Q15 (3.51). Slightly more male (51.7%) than female (46.9%). Profile 2 — The Pragmatist (N=114, 44%, M=36.97): Consistently lower scores across all 16 items. Substantially more male (64.0%) vs female (34.2%). The 11.6-point mean difference between profiles is large and practically significant for design segmentation. The largest item gaps are on Q14 (+1.09), Q15 (+1.08), Q5 (+0.80), and Q6 (+0.75). Profile × Gender (confirmed): χ²(1)=4.180, p=0.041, Cramér's V=0.128 (small but significant effect) — profile membership is significantly associated with gender. Males are overrepresented in The Pragmatist (64% male vs 57% sample base rate). This directly informs DD9: default empathy mode calibration should account for the gender distribution of the target user base.
UMAP + t-SNE Dimensionality Reduction
Visualising all 277 people in 2D space to see whether natural clusters emerge from their full response patterns.
Each person answered 16 questions — so they exist in a 16-dimensional space (impossible to visualise). UMAP and t-SNE are mathematical techniques that "fold" this 16-dimensional space into 2 dimensions while trying to keep similar people near each other. If natural clusters appear in the resulting map, it confirms that the empathy profiles found by LPA are real and not just statistical artefacts.
UMAP compresses all 16 dimensions of each respondent's answers into a 2D map, preserving both local and global structure. This is real data from Colab, not simulation. The two LPA profiles separate clearly — confirming they represent geometrically distinct groupings, not just statistical model outputs. Resonators (Profile 0, blue) cluster at lower UMAP values; Pragmatists (Profile 1, orange) cluster at higher values.
The smooth red-to-green gradient confirms that empathy increases systematically across the UMAP space — this is real data, not simulation. High-empathy respondents (green) cluster in the lower UMAP region; low-empathy (red) in the upper region. This validates both the total scoring system and the LPA cluster structure.
t-SNE K-Means confirms K=2 as optimal (matching BIC). HDBSCAN on UMAP confirmed 3 valid clusters + 55 noise points (21.2%). Cross-tabulation with LPA reveals: Cluster 0 (N=174, 66.9%, Mean TEQ=45.87) — contains 119 Resonators + 55 Pragmatists, F1≈F2 balanced; Cluster 1 (N=13, 5.0%, Mean TEQ=31.92) — Factor 2 dominant (F2=3.044 vs F1=0.978), 69% male — confirmed Detached profile (empathy inhibition); Cluster 2 (N=18, 6.9%, Mean TEQ=31.89) — near-equal factors (F1=2.048, F2=1.952), 67% male — low-engagement subtype. Combined low-reactive subgroup (Clusters 1+2): N=31 (11.9%), mean TEQ≈32, ~68% male. 4-mode MedAssist decision: empirical signal for a Minimal mode is confirmed but N=31 is below threshold for full design specification. Recommendation: 3-mode primary architecture with a dormant Minimal state activated by passive signal thresholds.
Hypothesis Testing
Eight formal hypotheses tested, with Bonferroni correction for multiple comparisons.
A hypothesis is a testable prediction. We made 8 predictions before analysing the data, then checked whether the evidence supported them. Because we're running 8 tests at once, we use the Bonferroni correction — a stricter significance threshold — to prevent accidentally calling a result significant just by chance. Think of it as raising the bar for every additional bet you place.
| Hypothesis | Test | Statistic | p-value | Bonferroni p | Result |
|---|---|---|---|---|---|
| H1: TEQ scores are normally distributed | Shapiro-Wilk | W = 0.9891 | 0.0479 | 0.2393 | ❌ Not sig. after Bonferroni |
| H2: Significant gender differences in TEQ scores | Mann-Whitney U | U = 5642.0, r = 0.287 | 0.0001 | 0.0004 | ✅ Female > Male (+4.07 pts) |
| H3: Age predicts TEQ total score | Spearman ρ | ρ = −0.002 | 0.9738 | 1.000 | ❌ No age effect |
| H4: TEQ differs across age groups (3 cohorts) | Kruskal-Wallis | H = 1.457 | 0.4826 | 1.000 | ❌ No age-group effect |
| H5: TEQ α ≥ 0.70 (acceptable reliability) | Cronbach α + CI | α = 0.772 [0.731–0.813] | N/A | N/A | ✅ Acceptable (α ≥ .70) |
| H6: Forward & reverse item means differ significantly | Wilcoxon signed-rank | W = 11166.5 | 0.0096 | 0.0479 | ✅ Significant (marginal after Bonf.) |
| H7: ≥ 2 distinct latent empathy profiles exist | LPA / GMM (BIC) | Best K = 2 (BIC = 12627.8) | N/A | N/A | ✅ 2 profiles confirmed (BIC) |
| H8: Respondent clusters are structurally distinct | Silhouette coeff. | S = 0.454 (K = 2) | N/A | N/A | ✅ Distinct clusters |
With 6 testable p-values, Bonferroni-corrected threshold = 0.05 / 6 = 0.0083. H2 (gender, p=0.0004) is the only hypothesis that clearly survives correction. H6 (item direction, p=0.0479) survives at the uncorrected level but is marginal after correction. H1 (normality) barely fails the raw threshold (p=0.048) but does not survive correction.
Group Comparisons & Regression
How do empathy scores vary across gender groups? Does age predict empathy? Item-level gender comparisons reveal finer-grained patterns.
We compared average empathy scores between male and female participants using the Mann-Whitney U test (a comparison method that works even without a normal distribution). We also checked whether older participants tend to be more or less empathetic using regression analysis — a method for understanding how one variable predicts another.
Female participants scored an average of 4.07 points higher than male participants — a statistically significant and practically meaningful difference (r=0.287 = moderate effect size). The strip of dots overlaid on each box shows individual participant scores, revealing substantial overlap between groups: there are high-empathy males and low-empathy females, but the central tendency differs meaningfully. Only Q11 and Q15 remained significant after Bonferroni correction at the item level.
The near-flat orange regression line tells the full story: age explains essentially none of the variance in empathy scores (R²=0.003). The wide scatter of dots at every age confirms that empathy is highly variable within any age group — a 20-year-old and a 30-year-old are equally likely to score anywhere across the range. This finding is typical for university-age samples with restricted age variance.
Female participants score 4.07 points higher than males (M♀=45.84, SD=7.91 vs M♂=41.77, SD=7.86). Effect size r=0.287 is a medium effect. This is the strongest and most robust finding in the study — significant both raw (p=0.0001) and after Bonferroni correction (p=0.0004). At item level, only Q11 ('I become irritated when someone cries') and Q15 ('I find it silly for people to cry') survive Bonferroni correction for item-level gender differences, both favouring females. Age is not predictive of empathy (ρ=−0.002, p=0.974, R²=0.003). (e.g., female > male by X points on average). Note effect size r and whether it survived Bonferroni correction. For the age regression, state whether age is a significant predictor (expected: weak or non-significant for a university-age sample with limited age range).
Design Directions
Eight evidence-based directions for designing products and interfaces that are responsive to human empathy — derived from the statistical findings.
This is where the numbers become design thinking. Each direction below is grounded in a specific statistical finding — not general intuition, but patterns measured across 277 people. They answer: given what we know about how people vary in empathy, how should we design?
Empathy-Adaptive Interface Layering
Three statistically distinct user types exist in this population. Design systems should support adaptive empathy registers — different tones, information densities, and interaction patterns for each profile.
- High Empathizers: Emotional narrative, community features, impact visibility, rich social proof
- Moderates: Clear hierarchies, rational framing, opt-in emotional content
- Low-Reactive: Utility-first design, factual CTAs, minimal emotional friction
Design for Core Empathy Drivers First
Network centrality reveals which aspects of empathy are most influential across the construct. Design features that activate high-centrality items propagate through the entire empathy network.
- High-strength items → use as primary UX triggers (help prompts, donate buttons, share features)
- Bridge items (high betweenness) → design as empathy "escalators" in the user journey
- Low-strength items → treat as context-specific; don't assume generalisation
Gender-Aware Emotional Design
Gender groups differ in mean empathy and in how they respond to specific questions. DIF-flagged items represent gender-sensitive constructs that should not be assumed universal.
- Test empathy-activating copy separately across gender groups
- Design emotionally charged prompts as gender-agnostic by default, with calibratable tone
- Avoid assuming that emotionally resonant design for one group works equally for another
IRT-Informed Micro-Moment Design
Items with high discrimination (a > 1.0) separate empathy levels precisely — they are the most powerful triggers for empathy-responsive design.
- High-a items → ideal triggers for emotionally activated moments (Q13: urge to help, Q5: making better)
- High-b items → reserved for experienced users; place later in the user journey after empathy "earned"
- Low-a items → broad-spectrum, safe for universal audiences early in onboarding
Reducing Friction for Low-Reactive Profiles
A statistically meaningful portion of users show low reactive empathy. These users need to engage with socially-oriented products without encountering design that presumes high emotional responsiveness.
- Avoid shame-based or guilt-inducing design patterns for low-reactive users
- Use rational motivation hooks first; introduce emotional content progressively
- Provide explicit control over the emotional "temperature" of the interface
Active Cues + Passive Inhibitor Reduction
Empathy has two dimensions: active prosocial tendencies (helping, connecting) and passive inhibitors (irritation, disinterest). Good design addresses both, not just one.
- Activate forward empathy with prompts that directly invite helping or connecting
- Reduce inhibitors: minimise cognitive friction, waiting, and emotional labour that triggers irritation
- Feedback mechanisms and error states should be tested with high-reverse-item respondents
Universal Design Claim Validity
Whether the same empathy design works universally depends on scalar invariance. If it holds, universal emotional design is empirically supported. If it fails, localisation is mandatory.
- Scalar invariance holds → empathy-based features can be designed with cross-group confidence
- Scalar invariance fails → localise emotional design per demographic segment
- Use MI evidence in ethics reviews and design documentation for emotionally sensitive products
Gender-Calibrated Default Mode
Profile membership is significantly associated with gender. The Pragmatist is male-skewed (64% male vs 57% base rate). Gender-neutral default empathy modes are empirically unjustified for this population.
- Male-majority products: default empathy mode → Balanced or Factual
- Mixed / female-majority: default → Balanced or Empathetic
- A/B test default states stratified by gender cohort before releasing to full audience
- Do not assume The Resonator is the universal default — it carries a measurable gender skew
- Gender-aware onboarding calibration where ethically appropriate
The Fourth Design Target — Minimal Mode
HDBSCAN confirms a geometrically distinct low-reactive subgroup. Cluster 1 (N=13) is Factor 2 (empathy inhibition) dominant — qualitatively different from The Pragmatist, not merely lower-intensity. Implement as a dormant Minimal mode activated by passive signal thresholds, not a default classification.
- Minimal mode requires: maximum control and privacy framing; zero emotional coercion; purely rational, outcome-oriented delivery; no forced emotional engagement; passive opt-in for emotional content
- The Detached ≠ character flaw — empathy inhibition may reflect cultural norms or protective coping mechanisms
- N=31 is below the threshold for full 4-mode design specification; confirmed as a real signal, not an artefact
- ~68% male: test Minimal mode interaction patterns with male user cohort first
- Two confirmed sub-types (F2-dominant inhibitors, Cluster 1; low-engagement across both factors, Cluster 2) — monitor in deployment for further differentiation
Instrument-Informed Persona Confidence
The TEQ shows acceptable-to-good reliability and model fit. This gives statistical backing to the confidence level we can place in the derived personas.
- Personas derived from profiles with α > 0.70 and CFI > 0.90 carry empirical weight in stakeholder presentations
- Low-loading items in CFA → aspects of empathy that may need separate measurement in future studies
- Use reliability coefficients as a "confidence interval" when communicating persona validity to non-technical stakeholders
User Archetypes
Three empirically-grounded personas, derived from the Latent Profile Analysis and UMAP cluster structure.
These are not made-up personas — they emerge directly from the data. Each archetype represents a statistically distinct group of real respondents who share a similar pattern of empathic tendencies. They are named for clarity, not stereotype, and the UX implications are derived from their specific item-level profiles.
- Social features, community & contribution affordances
- Impact visibility ("your action reached X people")
- Emotional storytelling & narrative-first UX
- Rich feedback loops; empathy-affirming micro-copy
- High-centrality triggers: Q3, Q5, Q13 activations
- Utility-first design; clear purpose & rational framing
- No guilt-based or emotionally coercive patterns
- Opt-in emotional content — never assumed
- Progressive emotional ramp-up; control stays with user
- Male-skewed (64%) — test empathy-prompts across genders
- Control, privacy & autonomy as core values
- Rational motivation — data, outcomes, not emotion
- Zero guilt-based or coercive patterns
- Progressive emotional disclosure on their terms
- High-inhibition items signal friction risk in flows
All three archetypes emerge from independent analytical methods. The Resonator and The Pragmatist come from BIC-optimal LPA (K=2, silhouette=0.454), with an 11.6-point mean TEQ gap and a significant gender skew (Pragmatist: 64% male; χ²(1)=4.180, p=0.041). The Detached derives from the confirmed HDBSCAN 3-cluster solution on UMAP — N=31 (11.9%) across two sub-clusters: Cluster 1 (N=13, Factor 2 dominant: F2=3.044 vs F1=0.978, 69% male — empathy inhibition profile) and Cluster 2 (N=18, near-equal factors, 67% male — low-engagement profile). The three personas form a continuous empathy spectrum, not discrete types. Note: 55 noise points (21.2%) from HDBSCAN are unclassified and may include additional low-empathy users.