The Vicious Circle

1. People Who Qualify for Disability Benefits Are Not Receiving Them

Israel’s National Insurance Institute pays disability benefits to hundreds of thousands of people. But in some localities, far fewer people claim than we would expect given their socio-economic conditions. We analyzed 278 localities with 15 indicators to find out where — and why.

What we did

We trained machine learning models to predict disability benefit rates from socio-economic data (income, education, demographics, geography). Localities where actual claiming is far below the prediction are potential under-utilization pockets — places where eligible people may face barriers to access.

278

Localities Analyzed

Across all regions and population groups

Socio-Economic Indicators

Income, education, demographics, geography

Research Questions

Under-utilization, intergenerational, conflict, distance

ML Models

Random Forest, XGBoost, TabPFN, Linear Regression

2. Research Question 1 Localities That Claim Far Less Than Their Conditions Predict

Two independent methods — a non-linear ensemble and a linear regression — predict disability rates from socio-economic data. Localities beyond ±1.5σ are flagged as anomalies.

2.1 Expected vs Observed Disability Rate (Ensemble: RF + XGBoost + TabPFN)

Green diamonds = under-utilization. Red crosses = hidden burden. Dashed bands = ±1.5σ (2.2 pp).

Sector breakdown of the 17 under-utilization localities (green diamonds) flagged in the graph above:

Arab under-utilization

1 of 17 · click to view

35%

Haredi under-utilization

6 of 17 · click to view

59%

Other under-utilization

10 of 17 · click to view

Note: ensemble R² < 0 for Haredi sector — Haredi flags may be unreliable.

2.2 Expected vs Observed Disability Rate (Linear Regression, R² = 0.699)

Sector breakdown of the 13 under-utilization localities (green diamonds) flagged in the graph above:

46%

Arab under-utilization

6 of 13 · click to view

54%

Other under-utilization

7 of 13 · click to view

Flagged by both methods

בסמ"ה

▶ Explore All 278 Localities

Locality ▲▼	Actual ▲▼	Expected (Ensemble) ▲▼	Gap (Ensemble) ▲▼	Population ▲▼	Group ▲▼	Ensemble	OLS
ברוכין	0.5%	4.5%	+4.0	2,637	Other	Under-util
בית שמש	5.9%	9.1%	+3.3	175,767	Haredi	Under-util
ברכה	2.3%	5.5%	+3.2	3,199	Other	Under-util
אשדוד	6.8%	9.9%	+3.1	227,285	Other	Under-util
יד בנימין	3.0%	5.9%	+3.0	4,326	Other	Under-util
עלי זהב	1.7%	4.6%	+2.9	5,354	Other	Under-util	Hidden
טלמון	2.4%	5.2%	+2.9	5,886	Other	Under-util
גבעת זאב	4.7%	7.5%	+2.8	23,424	Haredi	Under-util
רבבה	2.1%	4.7%	+2.6	3,157	Other	Under-util
ירושלים	5.4%	8.0%	+2.6	1,014,874	Other	Under-util
קריית יערים	4.1%	6.6%	+2.5	7,036	Haredi	Under-util
אלעד	4.9%	7.4%	+2.4	49,582	Haredi	Under-util
מודיעין עילית	5.0%	7.4%	+2.4	86,326	Haredi	Under-util
כפר חב"ד	4.0%	6.4%	+2.3	6,557	Haredi	Under-util
בסמ"ה	5.1%	7.4%	+2.3	11,698	Arab	Under-util	Under-util
לוד	7.4%	9.6%	+2.2	89,230	Other	Under-util
צור יצחק	3.4%	5.6%	+2.2	7,047	Other	Under-util
עין נקובא	3.8%	6.0%	+2.2	3,799	Arab		Under-util
זמר	5.1%	7.2%	+2.1	7,518	Arab
גדרה	4.2%	6.3%	+2.0	29,884	Other
מעלה אדומים	5.0%	7.0%	+2.0	37,165	Other
גן יבנה	4.2%	6.2%	+2.0	24,777	Other
גבע בנימין	5.2%	7.1%	+1.9	5,997	Other
באר יעקב	4.1%	6.0%	+1.9	33,689	Other
אלעזר	2.9%	4.8%	+1.9	2,735	Other
ברקת	4.2%	6.1%	+1.9	2,097	Other		Under-util
כוכב השחר	3.2%	5.1%	+1.9	2,850	Other
ביתר עילית	5.8%	7.6%	+1.8	71,073	Haredi
מעלה עירון	6.0%	7.8%	+1.8	16,152	Arab
חריש	6.8%	8.6%	+1.8	37,909	Other
פוריידיס	6.1%	7.8%	+1.7	13,838	Arab
באר גנים	2.8%	4.5%	+1.7	2,292	Other
אור עקיבא	7.2%	8.9%	+1.7	24,165	Other
עלי	2.6%	4.2%	+1.7	4,585	Other
אבן שמואל	3.7%	5.4%	+1.7	2,065	Other
ראשון לציון	5.7%	7.4%	+1.6	254,174	Other
הושעיה	2.1%	3.7%	+1.6	2,052	Other
מודיעין-מכבים-רעות	3.2%	4.8%	+1.6	100,113	Other
יקיר	2.2%	3.8%	+1.6	2,793	Other
כוכב יעקב	4.9%	6.5%	+1.6	3,485	Haredi
מזרעה	5.9%	7.5%	+1.6	4,149	Arab
נוקדים	3.6%	5.2%	+1.6	3,197	Other
ע'ג'ר	5.0%	6.5%	+1.6	2,887	Arab
אריאל	5.4%	7.0%	+1.5	19,128	Other
בית אל	3.8%	5.3%	+1.5	6,722	Other		Under-util
ערערה	6.1%	7.6%	+1.5	21,203	Arab
רמלה	8.1%	9.5%	+1.5	82,371	Other
בית אריה-עופרים	4.3%	5.7%	+1.4	5,503	Other
בת ים	8.3%	9.7%	+1.4	124,600	Other
בני ברק	6.8%	8.1%	+1.4	223,585	Haredi
ניצן	3.1%	4.5%	+1.3	2,563	Other
אחיעזר	4.3%	5.6%	+1.3	2,059	Other
רעננה	4.7%	6.0%	+1.3	82,844	Other
פדואל	2.4%	3.7%	+1.3	2,223	Other
שוהם	2.9%	4.2%	+1.3	23,519	Other
מיתר	2.8%	4.1%	+1.3	11,790	Other
נעלה	4.1%	5.4%	+1.3	2,794	Other
מתתיהו	5.2%	6.5%	+1.3	3,437	Haredi
רחובות	6.4%	7.6%	+1.3	149,399	Other
יבנה	5.7%	7.0%	+1.3	56,921	Other
קריית גת	8.7%	10.0%	+1.3	70,214	Other
בני עי"ש	7.0%	8.3%	+1.2	6,515	Other		Under-util
שבלי - אום אל-גנם	6.8%	8.1%	+1.2	6,830	Arab
ג'ת	4.5%	5.7%	+1.2	9,384	Arab
עפרה	3.0%	4.2%	+1.2	3,210	Other		Under-util
פתח תקווה	6.3%	7.5%	+1.2	255,952	Other
מבוא חורון	3.7%	4.9%	+1.2	2,630	Other
באקה אל-גרביה	6.1%	7.3%	+1.2	35,170	Arab
רמת השרון	4.6%	5.7%	+1.1	45,758	Other
ניל"י	2.3%	3.4%	+1.1	2,150	Other
צור משה	2.8%	3.9%	+1.1	2,617	Other
קרני שומרון	5.4%	6.5%	+1.1	10,551	Other
קריית עקרון	7.7%	8.8%	+1.1	10,888	Other
אליכין	4.0%	5.1%	+1.1	3,485	Other		Under-util
כפר אדומים	2.4%	3.5%	+1.1	4,738	Other
ראש העין	5.1%	6.1%	+1.0	74,367	Other
אפרת	4.1%	5.1%	+1.0	12,746	Other
טייבה	6.5%	7.5%	+1.0	47,069	Arab
יפיע	6.7%	7.7%	+1.0	19,976	Arab
מג'ד אל-כרום	6.0%	7.0%	+1.0	21,794	Arab
כפר האורנים	1.8%	2.8%	+1.0	2,114	Other
קיסריה	3.3%	4.2%	+1.0	6,031	Other
אום אל-פחם	6.8%	7.8%	+1.0	60,641	Arab
חולון	7.3%	8.3%	+1.0	187,734	Other
להבים	3.0%	4.0%	+0.9	7,241	Other
אלקנה	2.5%	3.4%	+0.9	4,565	Other
אלון מורה	3.0%	3.9%	+0.9	2,085	Other
קצר א-סר	3.6%	4.5%	+0.9	2,704	Arab		Under-util
אילת	7.9%	8.8%	+0.9	57,781	Other
שייח' דנון	5.9%	6.8%	+0.9	2,946	Arab
כפר ורדים	4.5%	5.4%	+0.9	5,549	Other
הר אדר	2.2%	3.1%	+0.9	3,966	Other
אשקלון	8.2%	9.0%	+0.8	160,929	Other
מרכז שפירא	3.8%	4.6%	+0.8	3,058	Other
קריית מלאכי	9.1%	9.9%	+0.8	27,668	Other
מעלות-תרשיחא	6.5%	7.3%	+0.8	23,448	Other		Under-util
כפר קאסם	6.6%	7.4%	+0.8	26,353	Arab
שילה	3.5%	4.2%	+0.8	5,767	Other
אזור	6.2%	7.0%	+0.8	12,496	Other
ג'ש (גוש חלב)	4.7%	5.4%	+0.8	3,111	Arab
שדרות	7.4%	8.1%	+0.8	35,796	Other
שלומי	6.8%	7.6%	+0.8	8,093	Other
יצהר	3.2%	4.0%	+0.8	2,426	Other
גבעתיים	4.0%	4.8%	+0.8	57,327	Other
שמשית	2.7%	3.4%	+0.8	2,279	Other
גבעת שמואל	4.0%	4.8%	+0.7	29,004	Other
צור הדסה	2.9%	3.7%	+0.7	10,083	Other
בית דגן	5.2%	5.9%	+0.7	7,226	Other
תל אביב -יפו	5.3%	6.0%	+0.7	461,649	Other
קלנסווה	7.0%	7.7%	+0.7	24,627	Arab
סביון	2.4%	3.1%	+0.7	4,176	Other
מעלה מכמש	3.7%	4.4%	+0.7	2,116	Other
עתלית	4.6%	5.2%	+0.7	10,912	Other
נס ציונה	4.3%	5.0%	+0.7	45,315	Other
נופית	3.1%	3.8%	+0.6	2,402	Other
אור יהודה	8.1%	8.7%	+0.6	39,382	Other
כוכב יאיר	3.9%	4.5%	+0.6	9,796	Other
כעביה-טבאש-חג'אג'רה	6.9%	7.5%	+0.6	6,291	Arab
אכסאל	7.0%	7.6%	+0.6	15,791	Arab
ג'לג'וליה	7.0%	7.6%	+0.6	10,877	Arab
מתן	2.4%	2.9%	+0.6	3,197	Other
קריית אונו	4.2%	4.8%	+0.6	42,385	Other
גבעת ברנר	3.5%	4.1%	+0.5	2,234	Other
גני תקווה	3.9%	4.5%	+0.5	24,163	Other
יהוד	5.1%	5.6%	+0.5	31,324	Other
אלון שבות	4.2%	4.7%	+0.5	3,290	Other
מעיליא	4.2%	4.7%	+0.5	3,165	Arab		Under-util
פקיעין (בוקייעה)	5.4%	5.9%	+0.5	6,114	Arab		Under-util
זכרון יעקב	4.4%	4.8%	+0.5	23,877	Other
עמנואל	7.0%	7.4%	+0.5	5,569	Haredi
שגב-שלום	6.1%	6.6%	+0.5	13,396	Arab
פסגות	3.3%	3.7%	+0.5	2,365	Other
נווה דניאל	4.1%	4.5%	+0.4	2,568	Other
פרדסייה	3.3%	3.8%	+0.4	7,755	Other
תקוע	3.6%	4.0%	+0.4	4,465	Other
עכו	8.4%	8.7%	+0.4	53,986	Other
יקנעם עילית	5.6%	5.9%	+0.3	24,749	Other
נהרייה	6.9%	7.2%	+0.3	67,695	Other
מבשרת ציון	4.9%	5.2%	+0.3	26,097	Other
נאעורה	6.0%	6.4%	+0.3	3,169	Arab
כפר מנדא	7.7%	8.0%	+0.3	22,269	Arab
כאוכב אבו אל-היג'א	6.0%	6.3%	+0.3	3,658	Arab
אלפי מנשה	4.2%	4.5%	+0.3	7,940	Other
בית שאן	9.9%	10.2%	+0.3	21,317	Other
אחוזת ברק	3.5%	3.7%	+0.2	2,446	Other
בנימינה-גבעת עדה	4.2%	4.4%	+0.2	15,029	Other
פסוטה	4.8%	5.0%	+0.2	3,431	Arab
מזכרת בתיה	4.1%	4.3%	+0.2	16,185	Other
ג'ולס	6.4%	6.6%	+0.2	6,871	Arab
מצפה יריחו	4.6%	4.7%	+0.1	2,753	Other
חורפיש	4.8%	5.0%	+0.1	6,794	Arab
גן נר	4.8%	4.9%	+0.1	2,642	Other
הוד השרון	4.8%	4.9%	+0.1	63,403	Other
כפר קרע	6.2%	6.3%	+0.0	20,799	Arab
רמת גן	5.3%	5.3%	+0.0	162,214	Other
הרצלייה	5.1%	5.1%	+0.0	104,165	Other
אבן יהודה	3.5%	3.5%	+0.0	14,088	Other
כסיפה	6.2%	6.2%	-0.0	21,654	Arab
קדומים	3.6%	3.6%	-0.0	4,871	Other
רמת ישי	4.5%	4.4%	-0.0	7,841	Other
אורנית	3.5%	3.4%	-0.1	9,317	Other
תל מונד	3.8%	3.7%	-0.1	14,760	Other
ג'דיידה-מכר	7.1%	7.0%	-0.1	21,134	Arab
דאלית אל-כרמל	7.8%	7.7%	-0.1	20,395	Arab
נוף איילון	3.0%	2.9%	-0.1	2,134	Other
מייסר	5.7%	5.5%	-0.1	2,064	Arab
בית יצחק-שער חפר	3.2%	3.0%	-0.2	2,142	Other
דימונה	9.5%	9.3%	-0.2	41,136	Other
לקיה	6.4%	6.2%	-0.2	18,401	Arab
כפר תבור	5.0%	4.8%	-0.2	4,516	Other
בית חשמונאי	3.0%	2.8%	-0.2	2,093	Other
סעוה	4.9%	4.7%	-0.2	2,332	Arab
טורעאן	6.4%	6.2%	-0.2	15,136	Arab
לפיד	3.1%	2.9%	-0.2	2,108	Other
חדרה	7.5%	7.3%	-0.2	103,959	Other
קדימה-צורן	4.6%	4.3%	-0.2	22,017	Other
ביר הדאג'	4.7%	4.5%	-0.3	3,666	Arab		Under-util
כפר ברא	6.4%	6.0%	-0.3	4,148	Arab
כפר כמא	6.5%	6.2%	-0.3	3,583	Arab
נתניה	7.3%	6.9%	-0.4	226,189	Other
חורה	6.6%	6.3%	-0.4	23,033	Arab
גבעת אבני	5.4%	5.0%	-0.4	2,155	Other
עוזייר	8.2%	7.8%	-0.4	3,641	Arab
ערד	10.4%	10.0%	-0.4	31,958	Other
עומר	4.5%	4.1%	-0.4	8,464	Other
בית ג'ן	6.8%	6.4%	-0.4	12,896	Arab
יאנוח-ג'ת	6.7%	6.2%	-0.5	6,969	Arab
מסעדה	7.3%	6.9%	-0.5	4,198	Arab
נצרת	7.9%	7.4%	-0.5	78,999	Arab
ברקן	4.5%	4.0%	-0.5	2,068	Other
כרמיאל	7.8%	7.3%	-0.5	47,769	Other
אבו סנאן	7.8%	7.3%	-0.5	14,502	Arab
דייר אל-אסד	6.5%	6.0%	-0.5	10,052	Arab
אופקים	8.8%	8.3%	-0.5	40,184	Other
אבו גוש	8.0%	7.5%	-0.5	8,426	Arab
ירוחם	8.2%	7.6%	-0.6	12,061	Other
ריינה	8.1%	7.5%	-0.6	17,247	Arab
רהט	7.0%	6.4%	-0.7	79,100	Arab
ניין	7.2%	6.5%	-0.7	2,016	Arab
בת חפר	3.4%	2.7%	-0.7	5,117	Other
אבו תלול	5.9%	5.1%	-0.7	2,867	Arab
סלמה	8.6%	7.8%	-0.7	3,681	Arab		Hidden
טירת כרמל	9.2%	8.5%	-0.7	30,938	Other
אל סייד	5.6%	4.8%	-0.7	3,407	Arab
קציר	6.1%	5.3%	-0.8	2,678	Other
טירה	7.6%	6.8%	-0.8	26,833	Arab
סח'נין	7.5%	6.7%	-0.8	35,524	Arab
ראמה	7.7%	6.9%	-0.8	8,469	Arab
כסרא-סמיע	7.0%	6.1%	-0.8	9,582	Arab
כפר מצר	6.8%	5.9%	-0.8	2,814	Arab
אעבלין	7.7%	6.9%	-0.8	13,228	Arab
נוף הגליל	9.1%	8.3%	-0.9	47,542	Other
צופים	4.6%	3.7%	-0.9	2,628	Other
עילוט	8.1%	7.2%	-0.9	9,163	Arab
דבורייה	7.1%	6.2%	-0.9	11,265	Arab
תל שבע	7.6%	6.7%	-0.9	24,242	Arab
בועיינה-נוג'ידאת	7.8%	6.9%	-0.9	10,969	Arab
שפרעם	8.4%	7.4%	-0.9	43,185	Arab
ערערה-בנגב	7.5%	6.5%	-1.0	21,590	Arab
חשמונאים	4.8%	3.8%	-1.0	3,186	Other
בסמת טבעון	8.0%	7.0%	-1.0	8,245	Arab
נתיבות	7.4%	6.4%	-1.0	54,781	Other
כפר יונה	5.7%	4.7%	-1.0	28,151	Other
עין מאהל	7.5%	6.4%	-1.1	14,404	Arab
גני מודיעין	8.0%	6.9%	-1.1	3,084	Haredi		Hidden
נשר	7.6%	6.5%	-1.1	21,881	Other
דייר חנא	8.7%	7.6%	-1.1	11,279	Arab
רכסים	6.9%	5.8%	-1.1	14,689	Haredi
עראבה	7.9%	6.8%	-1.1	28,010	Arab
בוקעאתא	7.8%	6.7%	-1.1	7,024	Arab
מוקייבלה	7.5%	6.3%	-1.2	4,530	Arab
מג'דל שמס	7.6%	6.4%	-1.2	11,660	Arab
רומת הייב	7.3%	6.1%	-1.2	2,182	Arab
קריית ארבע	6.2%	5.0%	-1.2	7,910	Other
כפר יאסיף	7.0%	5.8%	-1.2	10,620	Arab
כפר סבא	6.6%	5.4%	-1.2	99,158	Other
ביר אל-מכסור	8.1%	6.8%	-1.3	11,104	Arab
קריית ים	10.2%	8.9%	-1.3	41,520	Other
עספיא	8.7%	7.4%	-1.3	11,058	Arab		Hidden
בענה	7.9%	6.6%	-1.3	6,611	Arab
טמרה	8.0%	6.7%	-1.3	36,657	Arab
פרדס חנה-כרכור	6.7%	5.3%	-1.4	45,270	Other
כפר כנא	7.9%	6.5%	-1.4	24,636	Arab
קריית אתא	9.3%	7.9%	-1.4	61,710	Other
זרזיר	8.8%	7.3%	-1.5	9,299	Arab
חצור הגלילית	11.1%	9.6%	-1.5	11,134	Other		Hidden
קריית שמונה	11.1%	9.4%	-1.6	24,400	Other
משהד	7.9%	6.2%	-1.7	8,900	Arab
אום בטין	7.7%	6.0%	-1.7	4,371	Arab
אבני חפץ	4.9%	3.2%	-1.7	2,537	Other
נחף	8.7%	6.9%	-1.8	14,252	Arab
מצפה רמון	8.6%	6.7%	-1.8	5,687	Other
חיפה	8.7%	6.8%	-1.8	279,968	Other
מגדל העמק	10.5%	8.6%	-1.9	28,708	Other
אבטין	9.6%	7.7%	-1.9	2,931	Arab
קריית מוצקין	8.1%	6.2%	-1.9	50,191	Other
סולם	8.8%	6.9%	-1.9	3,065	Arab
טובא-זנגרייה	9.1%	7.1%	-1.9	6,905	Arab		Hidden
מעגן מיכאל	6.5%	4.5%	-2.0	2,017	Other		Under-util
כאבול	8.7%	6.7%	-2.0	13,190	Arab
עיילבון	8.9%	6.9%	-2.0	5,706	Arab		Hidden
ירכא	9.3%	7.1%	-2.2	17,188	Arab		Hidden
תפרח	8.0%	5.8%	-2.2	2,277	Haredi
קריית ביאליק	8.8%	6.5%	-2.2	45,354	Other
עפולה	11.2%	9.0%	-2.2	65,478	Other	Hidden	Hidden
סאג'ור	8.2%	5.9%	-2.3	4,587	Arab	Hidden	Hidden
קצרין	9.6%	7.3%	-2.3	8,380	Other	Hidden
שעב	9.1%	6.6%	-2.4	7,869	Arab	Hidden
עין קנייא	8.6%	6.1%	-2.5	2,392	Arab	Hidden
באר שבע	10.3%	7.6%	-2.7	214,661	Other	Hidden	Hidden
ג'סר א-זרקא	10.1%	7.3%	-2.9	15,778	Arab	Hidden	Hidden
קריית טבעון	7.2%	4.3%	-3.0	17,981	Other	Hidden
מגאר	10.3%	7.1%	-3.1	24,667	Arab	Hidden	Hidden
צפת	12.7%	9.1%	-3.6	39,301	Haredi	Hidden	Hidden
ראש פינה	9.3%	4.8%	-4.6	3,175	Other	Hidden
טבריה	14.8%	10.0%	-4.9	52,388	Other	Hidden	Hidden
יבנאל	13.3%	7.8%	-5.5	4,502	Haredi	Hidden	Hidden
מגדל	13.4%	7.5%	-5.9	2,076	Other	Hidden

2.3 What the Model Uses & What It Found

To explain the general disability benefit rate across localities, we used eight explanatory variables: log salary, peripherality index, income support rate, share of population aged 65+, labour participation rate, dependency ratio, and two dummy variables — Arab-majority locality and Haredi-majority locality (OLS with HC3 robust standard errors; N = 263, R² = 0.698).

Five of these variables are statistically significant (p < 0.05): log salary (β = −5.964, p < 0.001), Arab-majority locality (β = −1.414, p = 0.004), share 65+ (β = +0.175, p < 0.001), dependency ratio (β = −0.029, p < 0.001), and peripherality index (β = −0.236, p = 0.005). In plain terms: a locality with an Arab majority is associated with a 1.41 percentage-point lower disability rate, all else equal; a one-percentage-point increase in the share of residents aged 65 and above is associated with a 0.175 pp higher disability-benefit rate on average; and a 1% increase in average salary is associated with a 0.060 pp lower disability rate (Level-Log model: β / 100 = −5.964 / 100).

Variable	β (coef.)	Std. Error	p-value	Significance	Interpretation (1 unit / dummy)
Log Salary	−5.964	0.482	< 0.001	★★★	+1% in avg. salary → −0.060 pp disability rate (Level-Log: β/100)
Arab-majority (dummy)	−1.414	0.492	0.004	★★	Arab-majority locality → −1.41 pp disability rate, all else equal
Population 65+ %	+0.175	0.031	< 0.001	★★★	+1 pp in share of 65+ → +0.175 pp disability rate
Dependency Ratio	−0.029	0.006	< 0.001	★★★	+1 unit in dependency ratio → −0.029 pp disability rate
Peripherality Index	−0.236	0.084	0.005	★★	+1 unit in peripherality index → −0.236 pp disability rate
Labour Participation	−0.071	0.041	0.080	n.s.	Not statistically significant (p = 0.080)
Income Support Rate	+0.139	0.163	0.394	n.s.	Not statistically significant (p = 0.394)
Haredi-majority (dummy)	+0.342	0.789	0.665	n.s.	Not statistically significant (p = 0.665)

Variables are in their original (unstandardised) units. Coefficients reflect the change in the disability rate (percentage points) per one-unit increase in the explanatory variable. Dummy variables reflect the level difference relative to the reference group (non-Arab, non-Haredi localities). HC3 heteroscedasticity-robust standard errors. ★★★ p < 0.001 | ★★ p < 0.01 | ★ p < 0.05 | n.s. not significant.

3. Research Question 2 The Further from the Center, the Tighter the Intergenerational Disability Trap

Child disability rates correlate with adult disability rates across all socio-economic and geographic contexts.

3.2 Robustness Check: Does the Child–Adult Disability Link Survive Controls?

Five nested OLS models — HC3 robust standard errors, N = 263. Models 1–4: original (unstandardised) units. Model 5 (interaction): all continuous variables standardised to Z-scores (mean = 0, SD = 1) before constructing the interaction term.

Model 1 — Base

Tests the raw, direct relationship between the child disability rate and the adult disability rate — no controls. Establishes the baseline signal.

R² = 0.242

Model 2 — Structural

Adds economic and geographic controls: log salary, peripherality index, income support rate, share 65+, labour participation, dependency ratio. Tests whether the relationship survives controlling for poverty and distance from the centre.

R² = 0.767

Model 3 — District Fixed Effects

Adds district fixed effects (7 districts) on top of Model 2. Absorbs all unmeasured district-level differences in policy, infrastructure, and resources — identification comes from within-district variation only.

R² = 0.782

Model 4 — Sector Fixed Effects

Adds sector fixed effects (Arab / Haredi / Secular) on top of Model 2. Controls for all unmeasured cultural, institutional, and rights-access differences across population groups.

R² = 0.769

Model 5 — + Child×Peripherality

Extends Model 2 with an interaction term disabled_child_benefit_rate × peripherality_index_score. Tests whether the child–adult disability link is moderated by a locality's geographic peripherality.

R² = 0.773

✓

The child disability rate remains significant across all four models (p < 0.001)

Even after controlling for salary, peripherality, labour market participation, demographics, district identity, and sector identity, disabled_child_benefit_rate continues to significantly predict the adult disability rate. Its coefficient drops from β = 0.744 (Model 1) to β = 0.581 (Model 2), β = 0.584 (Model 3, district FE), and β = 0.565 (Model 4, sector FE) — highly significant in every specification. This is the core evidence for an intergenerational disability cycle that operates independently of geography, socio-economic conditions, and cultural context.

Notable: In Model 3, the peripherality index attenuates to insignificance (β = −0.103, n.s.) — suggesting that peripherality's effect on adult disability operates partly through district-level differences in resources, not locality-level geography alone.

Variable	Model 1 Base	Model 2 Structural	Model 3 + District FE	Model 4 + Sector FE	Model 5 + Child×Periph (Z-scored)
disabled_child_benefit_rate ← variable of interest	0.744*** (0.091)	0.581*** (0.080)	0.584*** (0.079)	0.565*** (0.091)	0.812*** (0.089)
Log Salary	—	−4.291*** (0.343)	−4.037*** (0.344)	−4.190*** (0.450)	−1.429*** (0.115)
Peripherality Index	—	−0.334*** (0.081)	−0.141 (n.s.) (0.113)	−0.338*** (0.083)	−0.275*** (0.089)
Income Support Rate	—	−0.005 (n.s.) (0.103)	+0.030 (n.s.) (0.105)	+0.047 (n.s.) (0.111)	+0.020 (n.s.) (0.102)
Population 65+ %	—	+0.114*** (0.021)	+0.100*** (0.020)	+0.115*** (0.023)	+0.694*** (0.113)
Labour Participation	—	−0.062*** (0.016)	−0.046** (0.015)	−0.062** (0.021)	−0.445*** (0.105)
Dependency Ratio	—	−0.018*** (0.004)	−0.010 (n.s.) (0.005)	−0.021*** (0.005)	−0.406*** (0.075)
Child Disability Rate × Peripherality (interaction term, Z-scored)	—	—	—	—	−0.203** (0.075)
District fixed effects (7)	—	—	✓ included	—	—
Sector fixed effects (Arab / Haredi / Secular)	—	—	—	✓ included	—
N	263	263	263	263	263
R²	0.242	0.767	0.782	0.769	0.773
Adj. R²	0.239	0.761	0.770	0.761	0.766

Standard errors in parentheses. *** p < 0.001 | ** p < 0.01 | * p < 0.05 | n.s. not significant. Models 1–4: variables in original (unstandardised) units. HC3 heteroscedasticity-robust standard errors throughout. Model 5: all continuous predictors standardised to Z-scores before estimation; coefficients represent effects in units of the outcome per 1-SD change in each predictor. Model 3 includes 7 district fixed effects (one per CBS district); individual district coefficients are not reported — fixed effects serve as controls, not results of interest. Model 4 includes sector fixed effects (Arab / Haredi / Secular); neither dummy reaches significance after controlling for salary, peripherality, and demographics.

▶

In peripheral localities — where economic and social disadvantage are already most acute — the child–adult disability link is significantly stronger (interaction β = −0.203, p < 0.01, Model 5). Where support resources are thinnest, the intergenerational cycle transmits most fully. In a well-funded welfare state, targeted services should moderate this transmission. The data show they do not: geography shapes the intensity of the cycle, but cannot break it.

Overall Picture

Double-Burden localities combine: lower socio-economic standing (−0.42 vs −0.06), lower wages (₪11,480 vs ₪13,429, −14.5%), higher income support dependence (1.64% vs 1.03%), weaker labor participation (72.41% vs 75.72%), and lower academic attainment (23.83% vs 31.28%) — creating a self-reinforcing environment where both adults and children depend on disability benefits. 6 of 7 indicators show statistically significant differences (Welch’s T-test p < 0.05). Targeted holistic family rehabilitation programs are needed.

What This Means for Policy

The 32 high-risk localities need holistic family-level rehabilitation programs, not individual disability support alone. Breaking the intergenerational cycle requires simultaneous intervention in education, employment, and healthcare access.

4. Every Model We Tried Hit the Same Wall

We tested Random Forest, XGBoost, TabPFN, and Linear Regression. All four models explain disability patterns well for most localities — and fail for Arab communities.

“It’s Disgraceful Going through All this for Being an Arab and Disabled”

Alhuzeel et al. (2023) — Scandinavian Journal of Disability Research. Interviews with 15 Arab Israelis revealed multi-level barriers: language (Hebrew-only forms and committees), excessive bureaucracy (3 medical committees vs 1), lack of information about rights in Arabic, and cultural stigma within the community itself.

Arab self-reported disability prevalence: 21% — Jewish: 19%. Yet benefit claiming is lower. 35% of approved beneficiaries never exercised eligibility.

Brookdale Institute (2024) — People with Disabilities in the Arab Population

4.2 We Removed the Arab Variable — Nothing Changed

We re-ran the ensemble without arab_population_percentage. The under-utilization list stayed identical. The model doesn’t need an “Arab” label — it reconstructs the group from the remaining 14 indicators: low SES, low salary, zero Haredi %, specific peripherality and education patterns. No single variable carries the signal; it is distributed across the entire socio-economic profile.

This means the model normalizes the under-claiming pattern. It learns “localities that look like this tend to claim less” and adjusts its prediction downward — turning systemic access barriers into a statistical baseline. The discrimination is not in one column; it is woven into the structure of the data.

4.3 This Has Happened Before

Obermeyer et al. (2019, Science): A US healthcare algorithm trained on 200 million patients used cost as proxy for need. Because the system spent less on Black patients with equivalent illness, the algorithm concluded they were healthier.

Our case is analogous: NII data records benefit receipt as proxy for disability prevalence. In communities facing access barriers, fewer people claim — creating the illusion of lower need.

5. Research Question 3 Did the Conflict Create a Disability Crisis?

Did the Swords of Iron conflict (October 2023 onward) cause a distinct spike in disability benefit rates in frontline localities?

+0.39 pp

Gaza Envelope — N=5 localities (2023→2025)

4.08% → 4.47%

+0.40 pp

Northern Frontline — N=21 localities (2023→2025)

4.11% → 4.51%

+0.32 pp

Other Population — N=170 localities (2023→2025)

3.41% → 3.73%

+0.09 pp

DiD — Northern Frontline

Unique effect vs national trend

+0.08 pp

DiD — Gaza Envelope

Unique effect vs national trend

Gaza Envelope — 5 localities

Ashkelon, Netivot, Ofakim, Sderot, Tifrah

Northern Frontline — 21 localities

Beit Jann, Buq‘ata, Ein Qiniyye, Fasuta, Ghajar, Hurfeish, Jish (Gush Halav), Katzrin, Kfar Vradim, Kiryat Shmona, Kisra-Sume‘a, Ma‘alot-Tarshiha, Ma‘alya, Majdal Shams, Mas‘ada, Mazra‘a, Nahariya, Peqi‘in (Buq‘ei‘a), Sheikh Dannun, Shlomi, Yanuh-Jat

Other Population — 168 localities

All remaining localities in the balanced panel

5.1 Disability Recipient Rate: Gaza Envelope vs Northern Frontline vs Other Population

Use the buttons to switch between the combined view and the Difference-in-Differences (DiD) view for each group. In the DiD views, the dashed line is the counterfactual — what the frontline rate would look like had it changed at the same pace as the rest of the country. The green segment at 2025 shows the DiD gap.

Note: The disability rate here is the general disability pension rate calculated relative to the total population (number of pension recipients in a locality divided by total population). This differs from the adult disability rate used in previous sections of this presentation, where the denominator was the working-age population (ages 18–64).

Conclusion (DiD: Northern Frontline = +0.09 pp, Gaza Envelope = +0.08 pp): Disability rates increased at a similar pace across all regions. The Difference-in-Differences values are near zero, meaning neither frontline group experienced a measurable spike beyond the national trend. The gaps between regions remained largely stable, pointing to structural rather than conflict-driven causes.

6. Research Question 4 Is Physical Distance the Barrier?

Is there a negative relationship between geographical distance from medical committee centers and disability benefit rates, after neutralising poverty and socio-economic variables?

6.1 Disability Rate vs Distance to Nearest National Insurance Institute (NII) Branch (km)

Each dot is a locality.

The graph shows a negative correlation between distance to National Insurance branches and disability benefit rates when considering distance to any branch (Pearson r≈−0.34), but almost no relationship when considering distance to a central branch only (r≈0.01). These findings indicate a moderate relationship that depends on how accessibility is defined, and likely reflects underlying socio-economic differences rather than a direct effect of distance.

6.2 OLS Regression: Distance to Medical Board Controlling for Socio-Economic Variables

Dependent variable: general disability rate. The key variable of interest is dist_central_branch_km (distance to nearest medical committee / central NII branch). Controls: labor participation, log salary, income support rate, dependency ratio, Arab-majority, Haredi-majority localities.

Variable	Coef.	95% CI
Distance to central branch (km)	−0.0054 (n.s.)	[−0.030, +0.019]
Labor participation rate	−0.1376**	—
Log average salary	−5.6499***	—
Income support rate	0.0930 (n.s.)	—
Dependency ratio	−0.0484***	—
Arab-majority locality	−2.8397***	—
Haredi-majority locality	−0.4542 (n.s.)	—

*** p<0.001 ** p<0.01 * p<0.05 (n.s.) not significant

Answer to Research Question 4: The hypothesis predicts that dist_central_branch_km will carry a negative and statistically significant coefficient. The coefficient is indeed negative (β = −0.005), consistent with the hypothesized direction, but it is not statistically significant (p = 0.666; 95% CI: −0.030 to +0.019). The confidence interval straddles zero, meaning we cannot distinguish the effect from chance. Once salary, dependency ratio, labour participation, and population composition are held constant, geographical distance to medical committee centres explains no independent variance in the disability benefit rate.

Conclusion: The research hypothesis is not supported. Distance is not a meaningful barrier to claiming disability benefits once structural socio-economic conditions are controlled. The bottleneck lies elsewhere — in information, language, bureaucratic capacity, and institutional trust — not in kilometres to the nearest branch.

7. Salary Is the Strongest Reducer of Disability Rates; Elderly Population the Strongest Booster

Knowing what drives disability tells us what to target. Two complementary views: Linear Regression shows direction (increases vs decreases), tree models show magnitude.

7.1 Linear Regression: Which Factors Increase or Decrease Disability?

Red bars = factor increases disability rate. Blue bars = factor decreases it. Salary is the strongest negative driver; population 65+ is the strongest positive driver. Income support rate, labour participation, and Haredi-majority are not statistically significant. R² = 0.698.

7.2 Tree Models: Which Factors Matter Most?

Importance regardless of direction. Averaged across Random Forest + XGBoost. Best tree-model R² = 0.640.

Both methods agree: salary and Arab population composition are the dominant factors. OLS identifies salary as the strongest single reducer and elderly population share as the strongest booster. Tree models additionally surface academic attainment, CBS socio-economic score, and child disability rate as major contributors — reinforcing the same socio-economic inequality axis and highlighting the intergenerational dimension of disability. Linear Regression tells us the direction (what to increase, what to decrease). Tree models tell us the importance (where to focus resources first). Together, they point directly to the actions in the last section.

8. Interaction Effects: When Context Changes the Rules

In this section, the dependent variable is the income support rate — the share of residents receiving income support benefits per locality. Multivariate regression with interaction terms (inspired by O'Brien, 2019) shows that sector and geography moderate the relationship between labor market conditions and this rate — explaining 47% of variance, up from 34%.

Methodology: Z-Score Normalization

All continuous variables were standardized to Z-scores (mean = 0, SD = 1) before constructing interaction terms. This serves two purposes: (1) comparable coefficients — each coefficient represents the effect of a one-standard-deviation change, enabling direct comparison across variables with different units; (2) reduced multicollinearity — centering variables before multiplying them prevents the interaction term from being strongly correlated with its components.

0.336

Baseline Adj. R²

Unemployment, dependency ratio, peripherality, salary

0.357

+Haredi × Unemp

Interaction not significant — negligible gain

0.427

+Periph Interactions

dep×periph and salary×periph both p<0.001

0.470

+Arab × Unemp

Best model — +40% relative gain over baseline

Interaction Term	Coefficient (β, standardized)	p-value	Significance	Interpretation
Peripherality × Dependency Ratio	−0.327	<0.001	*** highly significant	Peripherality amplifies the dependency effect — need intensifies uptake at the margins
Log-Salary × Peripherality	+0.267	<0.001	*** highly significant	Higher wages partially buffer the peripherality disadvantage
Arab Sector × Unemployment	−0.297	<0.001	*** highly significant	Unemployment–benefit link is significantly weaker in Arab-majority localities
Haredi Sector × Unemployment	+0.019	0.773	— not significant	No differential unemployment effect detected in Haredi localities

Insight 1: The Periphery Trap

The interaction between peripherality and dependency ratio (−0.327, p < 0.001) reveals that peripherality acts as a vulnerability magnifier. It doesn't simply add to the disadvantage; it compounds it, driving income support rates significantly higher in the periphery than the baseline model predicts.

Insight 2: The Arab Sector Paradox

In the general population, higher unemployment predicts higher income support uptake. In Arab-majority localities, this relationship is significantly weaker: the Arab × Unemployment interaction is −0.297 (p < 0.001). This is consistent with the institutional barriers documented elsewhere in this research — language barriers, limited digital access, and reduced NII outreach — which decouple economic need from actual benefit receipt. Unemployment rises, but claims do not follow.

Insight 3: The Haredi Non-Finding

Unlike the Arab sector, the Haredi × Unemployment interaction is not statistically significant (β = 0.019, p = 0.773). Within Haredi communities the standard unemployment–income support relationship operates similarly to the general population. Distinct patterns observed for Haredi localities in other analyses are therefore not driven by a differential sensitivity to unemployment, pointing toward structural or cultural factors operating through different channels.

Insight 4: Wages in the Periphery Hit Harder

The interaction between salary and peripherality (0.267, p < 0.001) demonstrates that the socio-economic impact of wages is highly location-dependent. In peripheral areas, the poverty-reducing effect of salary is significantly stronger. Consequently, low wages in the periphery drive up income support rates much more aggressively than identical low wages in central regions.

Bottom line: While peripherality significantly amplifies the impact of dependency and low wages on income support, a unique institutional 'decoupling' in the Arab sector—contrasting with the standard trends in Haredi localities—prevents economic distress from translating into actual benefit receipt.

9. What We Found

Four research questions, four answers — each revealing a different dimension of disability benefit under-utilization in Israel.

9.1 Where Is Under-Utilization Concentrated?

Ensemble anomaly detection flagged 17 localities claiming significantly less than their socio-economic profile predicts. Our models explain ~63% of overall variation but break down for Arab communities (R² = 0.18), pointing to barriers that are linguistic, institutional, and cultural — not captured in the data.

9.2 Is Under-Utilization Intergenerational?

Child and adult disability rates correlate strongly across all localities (ρ = 0.461, survives all 5 regression specifications). The link is stronger in more peripheral localities (interaction β = −0.203, p < 0.01): where support resources are thinnest, the disability cycle transmits most fully across generations. 32 localities (11.5%) were identified as “Double-Burden” — high child and adult disability rates simultaneously.

9.3 Did the Swords of Iron Conflict Increase Disability Claims?

Contrary to initial expectations, no significant impact of the conflict was detected. Gaza Envelope (5 localities, +0.39 pp) and Northern Frontline (21 localities, +0.40 pp) rose at nearly the same pace as the non-frontline population (170 localities, +0.32 pp). The gaps between groups remained stable, suggesting the differences reflect underlying structural factors rather than a short-term shock.

9.4 Is Physical Distance to NII Branches a Barrier?

Distance to the nearest National Insurance Institute branch showed no significant correlation with disability benefit uptake rates. This rules out geographic accessibility as a primary driver of under-utilization and shifts the focus to other barriers — awareness, language, and institutional trust — that do not appear in spatial data.

Project Summary: This research examines gaps in disability benefit utilization in Israel through a multi-dimensional lens. Beyond identifying under-utilization using statistical and machine learning models, it addresses intergenerational dynamics, conflict effects, and service accessibility — creating a comprehensive picture of not only who is not receiving benefits, but why. Key drivers such as income, age, and population structure were identified and shown to interact. At the same time, two intuitive hypotheses — distance to NII branches and war impact on frontline communities — were tested and not supported by the data. By combining anomaly detection with explanatory modeling, the analysis distinguishes broad systemic patterns from cases that remain unexplained by the model.