Decision-tree analysis of characteristics associated with 6-month smoking abstinence in Taiwan

Abstract

Objective

To identify smoker profiles associated with higher 6-month smoking abstinence using decision-tree analysis in a nationwide cessation service.

Design

Retrospective nationwide service-based study. Exploratory classification and regression tree (CART) analysis described subgroups with different observed abstinence probabilities, while multivariable logistic regression using generalised estimating equations (GEE) estimated adjusted associations accounting for repeated annual records within service users.

Setting

Taiwan’s Second-Generation Smoking Cessation Service Programme, a nationwide, multi-institutional smoking cessation service programme across multiple levels of care (2012–2022).

Participants

906 827 person-year records from 646 851 service users with valid 6-month abstinence outcomes aggregated by service user and calendar year. Annual records summarised service use, baseline characteristics and the last valid 6-month abstinence outcome.

Exposures

Variables included age, sex, smoking-related characteristics, service use, prescribed pharmacotherapy duration, calendar year, region, urbanisation, institution level and health professional type.

Primary outcome measures

Self-reported 7-day point-prevalence abstinence assessed at 6 months.

Results

In CART, the highest observed abstinence probability was 0.59 among records with at least 6 weeks of prescribed pharmacotherapy and time to first cigarette >30 min after waking (16% of records). Among records with at least 6 weeks of prescribed pharmacotherapy, time to first cigarette of ≤30 min and <20 cigarettes/day defined a large subgroup with an observed probability of 0.53 (39%). In adjusted GEE analysis, each 1-point increase in the Fagerström Test for Nicotine Dependence total score was associated with lower odds of abstinence (OR 0.92, 95% CI 0.91 to 0.92).

Conclusions

Exploratory CART identified person-year subgroups with differing observed abstinence probabilities. Longer recorded prescribed pharmacotherapy duration and lower overall nicotine dependence were generally associated with higher odds of abstinence in adjusted GEE analyses, but these observational findings do not establish causal effects or clinically validated treatment thresholds.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• Large, nationwide smoking cessation service database (2012–2022) spanning multiple levels of care in Taiwan, enabling large-scale, real-world analysis with harmonised definitions.

• Interpretable classification and regression tree approach used for exploratory subgroup profiling rather than causal inference or validated individual prediction.

• Cumulative prescribed pharmacotherapy weeks were derived from routine service records, but medication dispensing, actual use, ingestion and adherence were not verified.

• Outcome ascertainment by self-reported 7-day point-prevalence abstinence assessed at 6 months without biochemical verification may introduce misclassification.

• Observational, retrospective design; the main complete-case analysis was defined with respect to 6-month outcome availability, which was related to service engagement, and selection bias cannot be eliminated.

Introduction

Smoking remains a critical public health issue worldwide, with significant implications for individual health outcomes and population well-being. In Taiwan, the prevalence of smoking among adults aged ≥18 years decreased from 21.9% in 2008 to 14.0% in 2022, indicating progress towards the WHO target of a 30% relative reduction in tobacco use by 2025.¹ Despite this progress, the smoking cessation rate has plateaued, indicating a need for further improvement. Targeted interventions must address the specific factors contributing to smoking cessation to ensure a comprehensive and effective approach.

Regression studies have identified demographic, smoking-related and psychosocial characteristics associated with smoking cessation outcomes.^2–4 These models estimate adjusted associations for individual variables, but combinations of characteristics may be less immediately visible. Classification and regression tree (CART) methods offer an interpretable, data-driven way to describe subgroups defined by sequential combinations of variables.^{5 6} Such trees are useful for exploratory subgroup profiling, but they should not be interpreted as causal pathways or as validated clinical prediction rules without appropriate validation. This study used CART to describe person-year subgroups with different observed probabilities of self-reported 7-day point-prevalence abstinence assessed at 6 months in Taiwan’s national smoking cessation service dataset.

Materials and methods

Study design and analytical unit

We conducted a retrospective study using data from Taiwan’s Second-Generation Smoking Cessation Service Programme (SGSCSP), a nationwide service programme overseen by the Taiwan Health Promotion Administration, Ministry of Health and Welfare. Launched in March 2012, the SGSCSP introduced an institution-based smoking cessation service system. Under this programme, each SGSCSP user is eligible for up to four health education and four pharmacotherapy sessions per year, with up to eight visits for health education and up to 8 weeks of pharmacotherapy in each session.⁷ This study focused on SGSCSP participants between 2012 and 2022 and comprised both health education and pharmacotherapy sessions. Pharmacotherapy sessions included prescription medications such as varenicline and bupropion, as well as nicotine replacement therapies (NRTs). Because participation reflects the use of smoking cessation services rather than probability sampling of the general smoking population, the study represents a nationwide service-based cohort rather than a nationally representative sample of all smokers in Taiwan.

The source data contained 1 013 131 pharmacotherapy treatment-course records and 793 152 health education treatment-course records, for a total of 1 806 283 eligible treatment-course records between 2012 and 2022. At the treatment-course level, 1 404 699 records had completed follow-up, and 401 584 had no completed follow-up or no valid abstinence outcome.

All eligible treatment-course records for the same service user and calendar year were aggregated using service user ID and calendar year, yielding 1 158 098 person-year records from 805 696 unique service users. The same service user could contribute more than one person-year in different calendar years. Calendar-year baseline individual and institutional characteristics were taken from the first treatment-course record in that year. Cumulative prescribed pharmacotherapy weeks, number of treatment-course records and service type were summarised across all records in that person-year. For each person-year, the abstinence outcome was assigned from the last treatment-course record in that calendar year with a valid 6-month abstinence assessment. A valid outcome was available for 906 827 person-year records from 646 851 service users, which formed the main complete-case analysis with respect to outcome availability; 251 271 person-year records had no valid outcome and were retained for the follow-up comparison and sensitivity analyses (figure 1).

Figure 1. Flow chart of treatment-course records and person-year records included in the complete-case analysis.

Variables potentially associated with 6-month smoking abstinence

Three categories of variables were considered in this study, namely, individual, ecological and structural factors. Calendar-year baseline individual and institutional variables were derived from the first health education or pharmacotherapy treatment-course record in that person-year. Service-engagement variables, including cumulative prescribed pharmacotherapy weeks, number of treatment-course records and service type, were summarised across all treatment-course records for the same service user and calendar year. The ascertainment periods for the study variables are shown in online supplemental efigure 1.

Individual factors

Individual variables included age, sex, smoking duration and nicotine dependence measured using the Fagerström Test for Nicotine Dependence (FTND).^{8 9} Race and ethnicity were not collected in the SGSCSP database and were therefore unavailable for analysis. The six FTND items were time to first cigarette after waking, difficulty refraining from smoking in forbidden places, cigarette most difficult to give up, number of cigarettes smoked per day, smoking more frequently during the first hours after waking and smoking when ill in bed. The item response categories are reported in table 1.

Table 1. Fagerström Test for Nicotine Dependence item distributions among person-year records included in the main complete-case analysis.

FTND item	N*	%
FTND score†, mean±SD	6.14±2.41
≤3	108 160	11.93
4–6	371 918	41.01
≥7	426 749	47.06
Number of cigarettes smoked per day, mean±SD	21.55±14.83
≤10	204 228	22.52
11–20	404 451	44.60
≥21	298 148	32.88
Time to first cigarette after waking
Within 5 min	469 088	51.73
6–30 min	271 999	29.99
31–60 min	165 740	18.28
Difficulty refraining from smoking in forbidden places
Yes	516 924	57.00
No	389 903	43.00
Cigarette most difficult to give up
First cigarette in the morning	669 056	73.78
Other cigarettes	237 771	26.22
Smoking more frequently during the first hours after waking
Yes	499 082	55.04
No	407 745	44.96
Smoking when ill in bed
Yes	398 413	43.93
No	508 414	56.07

^*Values are expressed as number (percentage) unless otherwise indicated. For variables with missing values, percentages were calculated among records with non-missing data.

^†FTND score: nicotine dependence scores were assessed according to the 1991 version of the Fagerström Test for Nicotine Dependence.

FTND, Fagerström Test for Nicotine Dependence.

Based on the SGSCSP database, the type of service use in a calendar year was classified as health education session only, pharmacotherapy session only or both health education and pharmacotherapy sessions. Pharmacotherapy type was also classified as prescription medications only, NRT only or both prescription medications and NRT. Prescription medications included varenicline and bupropion.

Previous studies have reported associations between pharmacotherapy duration and smoking abstinence.^{10 11} We therefore calculated cumulative prescribed pharmacotherapy weeks during the calendar year. Cumulative prescribed pharmacotherapy weeks represented recorded prescribed duration and did not establish medication dispensing, actual medication use, ingestion or adherence.

Prior data suggested that smoking cessation service use may be associated with abstinence,¹² and other studies have reported associations between more frequent visits or counselling and cessation outcomes.^13–15 We therefore included smoking cessation service use rate as a service-engagement variable. The definition of high use differed between health education and pharmacotherapy sessions. For health education sessions, high use was defined as at least two visits. For pharmacotherapy sessions, high use was defined as at least 4 weeks of prescribed pharmacotherapy within a session, based on the service programme structure.¹⁶

Ecological factors

Ecological factors included the year of participating in the service programme, region and urbanisation of institutions. Muzi et al¹⁷ and Northridge et al¹⁸ found that smoking cessation was more successful in urban areas than in rural areas, with Muzi et al¹⁷ also noting higher cessation rates in more urbanised places.

This study used an urbanisation classification scheme proposed by Wu et al,¹⁹ who classified all 358 city districts and townships in Taiwan into seven levels of urbanisation based on population density, proportion of residents with a college or higher education, percentage of older people (age >65 years) and proportions of agricultural, industrial and professional or managerial workforces, as well as the proportion of people aged 15–65. In this study, we classified levels 1 and 2 as high, 3 and 4 as median, and 5–7 as low urbanisation levels.

Structure factors

Many studies have explored the impact of different structural measures on the success of smoking cessation interventions.^{20 21} The type of institution and specific smoking cessation services they offer can significantly impact success rates. Yu et al²⁰ showed that higher success rates were associated with clinics providing comprehensive support, such as medication compliance, multiple clinic visits and multi-professional interventions. For instance, clinics with higher medication compliance had lower primary failure risks, and those offering interventions from multiple health professionals saw substantial increases in quitting success.²¹ A meta-analysis highlighted that a team of physicians and non-physicians using multiple intervention modalities was most effective.²²

Two structural measures (ie, level of institution and type of healthcare professionals) were also used for categorised smokers with success in smoking cessation. Taiwan’s SGSCSP included various types of institutions to provide smoking cessation services, including medical centres, regional hospitals, community hospitals, clinics, public health centres, dental clinics and community pharmacies. The healthcare professionals include physicians, pharmacists, health educators and dentists.

Outcome measure: 6-month smoking abstinence

The outcome was the self-reported abstinence from smoking during the previous 7 days assessed 6 months after the first day of each session, based on the information from the most recent health education or pharmacotherapy session in a given year. The definition of outcome measure is shown in online supplemental efigure 1. To acquire information on the 6-month smoking status, institutional healthcare staff or qualified health educators conducted a telephone interview with the participants: “Did you smoke in the past 7 days?” Participants who answered “no” were classified as abstinent. The follow-up completion rate was 78% (1 404 699/1 806 283).

Statistical analysis

Continuous variables were summarised using means and SD, and categorical distributions were additionally reported for key continuous variables. An exploratory CART analysis was conducted to partition person-year records in the main complete-case analysis into subgroups with different observed probabilities of self-reported 7-day point-prevalence abstinence assessed at 6 months.^{23 24} At each node, the algorithm selected the variable and cut-off that resulted in the greatest reduction in Gini impurity. Tree growth and pruning were controlled using a complexity parameter (cp) of 0.01. A node required at least 20 observations to be considered for splitting (minsplit=20), each terminal node required at least seven observations (minbucket=7) and the maximum tree depth was 30 (maxdepth=30). Ten-fold cross-validation estimates were calculated as part of the tree-fitting procedure (xval=10). Each terminal node displays the observed abstinence probability and terminal-node size, expressed as the proportion of person-year records. The 6-week pharmacotherapy split was data-driven rather than prespecified or clinically validated. The tree was used for exploratory subgroup profiling, not for causal inference or development of a validated prediction model. Unlike the generalised estimating equation (GEE) analyses, the CART analysis treated person-year records as observations and did not explicitly model within-user correlation.

Multivariable logistic regression models were fitted using GEE, with service user ID specified as the clustering variable to account for repeated annual person-year records.²⁵ Separate GEE models examined nicotine dependence: model 1 included the FTND total score, whereas model 2 replaced the total score with all six FTND item scores entered simultaneously in the same model. Higher FTND total score and item scores indicated greater nicotine dependence. Both models adjusted for age, sex, smoking duration, cumulative prescribed pharmacotherapy weeks, number of treatment-course records in the year, service type, calendar year, health insurance branch region, urbanisation level, institution level and health professional type.

Covariate completeness was assessed before modelling. The main complete-case cohort was defined by the availability of a valid 6-month abstinence outcome, and each multivariable model included records with complete data for all covariates required in that model. Analyses of cumulative prescribed pharmacotherapy duration were restricted to person-year records with at least 1 week of prescribed pharmacotherapy; health-education-only records were therefore treated as not applicable rather than as having missing pharmacotherapy exposure.

As a complementary sensitivity assessment of the CART-derived 6-week split, the existing week-specific analysis examined cumulative recorded prescribed pharmacotherapy duration using individual week-specific categories from 1 through ≥16 weeks in multivariable GEE models among person-year records with at least 1 week of prescribed pharmacotherapy. This analysis evaluated the association across the observed duration range without imposing a binary 6-week cut-off.

To assess potential selection associated with outcome availability, person-year records included in the complete-case analysis were compared with those excluded because valid 6-month outcomes were lacking. Absolute standardised mean differences (SMDs) were calculated; for multi-category variables, the maximum category-specific absolute SMD was used in online supplemental efigure 2. An absolute SMD of ≥0.10 was considered potentially meaningful.

For incomplete follow-up, the complete-case estimate was compared with two deterministic sensitivity analyses: all person-year records without a valid outcome were classified as not abstinent, or all were classified as abstinent.

To assess potential temporal changes during the COVID-19 period, the annual programme record volume and the observed abstinence proportion among person-year records with valid 6-month outcomes from 2012 to 2022 were additionally summarised. This assessment was descriptive and was not intended to estimate a causal effect of the COVID-19 pandemic. Calendar year was included as a covariate in the adjusted GEE models.

All statistical analyses and graphing were performed using the rpart package in R V.4.3.2 or SAS V.9.4 (SAS Institute, Cary, NC). Furthermore, all tests were two sided, and statistical significance was defined as p<0.05.

Patient and public involvement

Patients and the public were not involved in the design, conduct, reporting or dissemination plans of this research. This study analysed de-identified, routinely collected data from smoking-cessation services with no direct patient contact or recruitment. The findings will be disseminated via open-access publication, and a plain-language summary will be shared with service providers and relevant public health channels.

Results

Between 2012 and 2022, the main complete-case analysis with respect to outcome availability included 906 827 person-year records with valid 6-month abstinence outcomes. Table 2 shows the demographic characteristics of the person-year records. More than 88% of engaged person-year records were under the age of 65. The average age was 47.07 years (SD=13.92). Among records with known sex, 83.91% were male; sex was recorded as unknown for 95 143 (10.49%) person-year records, whereas other measured explanatory variables had <0.1% or no missingness. The mean smoking duration was 24.56 years (SD=13.29), the FTND score averaged 6.14 (SD=2.41) and daily cigarette consumption averaged 21.55 cigarettes (SD=14.83). The majority (51.73%) smoked their first cigarette within 5 min of awakening. For 62.28% of person-year records, use of smoking cessation services was low (ie, <2 visits of health education or <4 weeks of pharmacotherapy for the first session in a given year). Approximately 40% of institutions offering cessation assistance were in highly urbanised areas. Clinics provided most treatments (29.45%), with regional hospitals accounting for 19.66%. A sizeable proportion (37.74%) received pharmacotherapy sessions only, whereas 38.00% attended both health education and pharmacotherapy sessions. In terms of pharmacotherapy type, 54.35% used prescription medications only (varenicline and/or bupropion), 18.35% used NRT only and 25.76% received pharmacotherapy for at least 8 weeks in a given year.

Table 2. Characteristics of person-year records included in the main complete-case analysis (n=9 06 827).

Variable*	Person-year records, n†	%
Individual factors
Age (years), mean±SD	47.07±13.92
18–49	530 831	58.54
50–64	277 678	30.62
65–74	75 935	8.37
≥75	22 383	2.47
Sex
Male	681 094	83.91
Female	130 590	16.09
Smoking duration (years), mean±SD	24.56±13.29
≤10	165 319	18.23
11–20	267 034	29.45
21–30	243 543	26.86
≥31	230 931	25.47
Smoking cessation service use rate‡
Low	564 789	62.28
High	342 038	37.72
Types of service sessions in a year
Health education session only	219 924	24.25
Pharmacotherapy session only	342 276	37.74
Both health education and pharmacotherapy sessions	344 627	38.00
Types of pharmacotherapies used in a year
Health education session only	219 924	24.25
Prescription medications only§	492 889	54.35
NRTs only¶	166 444	18.35
Both prescription medications and NRTs	27 570	3.04
Total number of weeks on pharmacotherapy, mean±SD	5.93±4.77
<2	111 063	12.25
2–3	158 214	17.45
4–7	184 025	20.29
≥8	233 601	25.76
Ecological factors
Year of participating in the service programme
2012–2016	355 080	39.16
2017–2022	551 747	60.84
Urbanisation level of institutions
High	361 755	39.89
Medium	244 855	27.00
Low	300 217	33.11
Location of medical/health institutions
Taipei division	261 466	28.83
Northern division	111 726	12.32
Central division	176 573	19.47
Southern division	153 914	16.97
Kaohsiung-Pingtung region	182 572	20.13
Eastern region	20 576	2.27
Structure factors
Institutions level
Medical centre	99 736	11.00
Regional hospital	178 291	19.66
Community hospital	99 809	11.01
Clinic	267 076	29.45
Public health centre	132 927	14.66
Dental clinic	12 687	1.40
Community pharmacy	116 301	12.83
Health professional type
Physician	558 460	61.63
Pharmacist	118 165	13.04
Health educator	216 206	23.86
Dentist	13 310	1.47

^*Baseline data are derived from the first health education or medication treatment session in a given year.

^†Values are expressed as number (percentage) unless otherwise indicated. For variables with missing values, percentages were calculated among records with non-missing data.

^‡High use indicated a minimum of 2 out of 8 weeks of health educational sessions or 4 out of 8 weeks of pharmacotherapy for the first session in a given year.

^§Prescription medications were to be either varenicline and/or bupropion.

^¶NRT comes in the form of gum, nasal sprays, transdermal patches and sublingual tablets.

NRT, nicotine replacement therapy.

Outcome availability differed most strongly according to service engagement, particularly cumulative prescribed pharmacotherapy weeks and number of treatment-course records. Compared with records without a valid outcome, included records had more cumulative prescribed pharmacotherapy weeks (4.23 vs 2.65; absolute SMD 0.397) and more treatment-course records (1.62 vs 1.35; SMD 0.368). Other potentially meaningful differences were observed for service type, health professional type, institution level and age. Differences in smoking duration, FTND total score, sex and most individual FTND items were smaller (online supplemental etable 1 and efigure 2).

In the exploratory CART analysis, the first split was at 6 weeks of cumulative prescribed pharmacotherapy. The terminal subgroup with the highest observed abstinence probability (0.59) comprised person-year records with at least 6 weeks of prescribed pharmacotherapy and a time to first cigarette of >30 min after waking; this subgroup represented 16% of records. Among records with at least 6 weeks of prescribed pharmacotherapy, a time to first cigarette of ≤30 min and <20 cigarettes per day, the observed abstinence probability was 0.53 (39% of records). The subgroup with <6 weeks of prescribed pharmacotherapy and >10 cigarettes per day had an observed probability of 0.42 (29% of records). These terminal nodes describe observed subgroup profiles rather than causal effects, clinically validated thresholds or individual prediction rules (figure 2).

Figure 2. Exploratory classification and regression tree for self-reported 7-day point-prevalence abstinence asssessed at 6 months. Terminal nodes show the observed abstinence probability and the proportion of person-year records in each subgroup. The 6-week pharmacotherapy split was data-driven rather than prespecified. The tree was used for exploratory subgroup profiling and should not be interpreted as a causal pathway, validated prediction model, clinical decision rule or optimal pharmacotherapy-duration threshold. Percentages may not sum to 100% because of rounding.

The distributions of the FTND total score and six FTND item responses are shown in table 1. The mean FTND total score was 6.14 (SD 2.41), and 47.06% of person-year records had a score of ≥7. 51.73% indicated smoking the first cigarette within 5 min after waking, 32.88% reported smoking ≥21 cigarettes per day and 57.00% reported difficulty refraining from smoking in forbidden places.

In adjusted GEE analyses, each 1-point increase in FTND total score was associated with lower odds of self-reported 7-day point-prevalence abstinence assessed at 6 months (OR 0.92, 95% CI 0.91 to 0.92). In the mutually adjusted item-level model, higher scores for time to first cigarette after waking (OR 0.85, 95% CI 0.85 to 0.86), number of cigarettes smoked per day (OR 0.86, 95% CI 0.86 to 0.87) and smoking when ill in bed (OR 0.87, 95% CI 0.86 to 0.88) were associated with lower odds of abstinence. Some item-level associations were small and in the positive direction. These estimates should be interpreted with caution because the six correlated FTND item scores were entered simultaneously in the same model, and they should not be interpreted as protective effects (table 3).

Table 3. Associations of the FTND total score and item-level FTND scores with self-reported 7-day point-prevalence abstinence assessed at 6 months using multivariable logistic regression with generalised estimating equations.

Model	Variable	Adjusted OR (95% CI)*	P value
Model 1: FTND total score†
	FTND total score	0.92 (0.91 to 0.92)	<0.0001
Model 2: FTND item-level scores†
	Number of cigarettes smoked per day	0.86 (0.86 to 0.87)	<0.0001
	Time to first cigarette after waking	0.85 (0.85 to 0.86)	<0.0001
	Difficulty refraining from smoking in forbidden places	1.02 (1.01 to 1.03)	<0.0001
	Cigarette most difficult to give up	0.99 (0.97 to 1.00)	0.0136
	Smoking more frequently during the first hours after waking	1.04 (1.03 to 1.05)	<0.0001
	Smoking when ill in bed	0.87 (0.86 to 0.88)	<0.0001

^*Adjusted ORs for the FTND total score and item scores represent the change in odds of self-reported 7-day point-prevalence abstinence assessed at 6 months per 1-point increase. Both models adjusted for age, sex, smoking duration, cumulative prescribed pharmacotherapy weeks, number of treatment-course records, service type, calendar year, region, urbanisation, institution level and health professional type, with repeated records clustered by service user ID. Multivariable models were fitted using records with available data for all variables required in the corresponding model.

^†Model 1 included the FTND total score. Model 2 replaced the FTND total score with the six FTND item scores entered simultaneously in the same model.

FTND, Fagerström Test for Nicotine Dependence.

Table 4 shows the association between cumulative recorded prescribed pharmacotherapy weeks and self-reported 7-day point-prevalence abstinence assessed at 6 months among person-year records with at least 1 week of prescribed pharmacotherapy. Records with 8 weeks of prescribed pharmacotherapy had an observed abstinence proportion of 49.04% and higher adjusted odds of abstinence than records with 1 week of prescribed pharmacotherapy (adjusted OR 1.70, 95% CI 1.65 to 1.75; p<0.0001) (online supplemental efigure 3). In the complementary week-specific sensitivity assessment, adjusted odds of abstinence were generally higher at longer recorded prescribed durations than at 1 week of prescribed pharmacotherapy, although the pattern was not strictly monotonic. Adjusted ORs were 1.39 (95% CI 1.35 to 1.44) at 5 weeks, 1.45 (95% CI 1.41 to 1.50) at 6 weeks, 1.59 (95% CI 1.53 to 1.66) at 7 weeks and 1.70 (95% CI 1.65 to 1.75) at 8 weeks. Thus, the broader duration–abstinence association was observed without imposing the CART-derived 6-week dichotomy.

Table 4. Associations of cumulative recorded prescribed pharmacotherapy weeks with self-reported 7-day point-prevalence abstinence assessed at 6 months.

Cumulative recorded prescribed pharmacotherapy weeks	Person-year records, n	Abstinent at 6 months, n (%)	Adjusted OR (95% CI) *	P value
1	111 063	31 162 (28.06)	1.00 (Ref.)	–
2	123 226	40 689 (33.02)	1.24 (1.22 to 1.26)	<0.0001
3	34 988	13 196 (37.72)	1.58 (1.54 to 1.62)	<0.0001
4	99 758	37 660 (37.75)	1.53 (1.50 to 1.56)	<0.0001
5	28 856	12 424 (43.06)	1.39 (1.35 to 1.44)	<0.0001
6	41 422	18 251 (44.06)	1.45 (1.41 to 1.50)	<0.0001
7	13 989	6370 (45.54)	1.59 (1.53 to 1.66)	<0.0001
8	85 121	41 744 (49.04)	1.70 (1.65 to 1.75)	<0.0001
9	7558	3132 (41.44)	1.47 (1.39 to 1.54)	<0.0001
10	16 258	7055 (43.39)	1.48 (1.42 to 1.54)	<0.0001
11	4984	2109 (42.32)	1.45 (1.36 to 1.54)	<0.0001
12	30 810	13 711 (44.50)	1.45 (1.40 to 1.51)	<0.0001
13	4160	1817 (43.68)	1.39 (1.30 to 1.49)	<0.0001
14	12 055	5505 (45.67)	1.43 (1.36 to 1.50)	<0.0001
15	4075	1855 (45.52)	1.47 (1.38 to 1.58)	<0.0001
≥16	68 580	34 464 (50.25)	1.55 (1.50 to 1.61)	<0.0001

^*Adjusted OR and 95% CI were estimated from multivariable logistic regression with generalised estimating equations. Adjusting for individual factors (including age, sex, smoking duration, FTND score, daily cigarette consumption, first cigarette after waking, smoking cessation service use rate, types of service sessions in a year and types of medications used in a year), ecological factors (including year of participating in the service programme, urbanisation level of medical institutions and location of medical/health institutions) and structure factors (including institutions level and health professional type).

Ref. Reference group.

The annual programme record volume declined during the later study years, including the COVID-19 period. The number of reconstructed person-year records was 143 953 in 2019, 118 307 in 2020 and 86 634 in 2021. Among person-year records with valid 6-month outcomes, the observed abstinence proportions were 39.90%, 41.08% and 39.88%, respectively; the corresponding proportion in 2022 was 39.54% (online supplemental etable 2).

Under the conservative assumption that all 251 271 person-year records without a valid outcome were not abstinent, the estimated abstinence proportion was 30.36% (351 566/1 158 098). Under the upper-bound assumption that all were abstinent, the proportion was 52.05% (602 837/1 158 098).

Discussion

In this study, we used a nationwide smoking cessation service dataset to describe person-year subgroups with differing observed probabilities of self-reported 7-day point-prevalence abstinence assessed at 6 months. Cumulative recorded prescribed pharmacotherapy duration and nicotine-dependence characteristics were prominent profiling variables.

A range of factors may influence smoking cessation outcomes. The success of smoking cessation is influenced by various factors, including demographic aspects such as older age and marital status, which provide stability and support. Socio-economic elements like higher income and education enhance access to resources.²⁶ Behavioural factors, including lower nicotine dependence and good mental health, significantly improve quitting rates.¹⁴ Additionally, a review article showed that environmental and social support systems from family, friends, and healthcare providers are also crucial.²⁷

The added value of the decision tree is descriptive, visual and exploratory. GEE models estimate adjusted population-averaged associations for prespecified variables, whereas CART provides an interpretable display of how treatment-course and smoking-related characteristics are sequentially combined to define subgroups with different observed abstinence probabilities. For example, after the 6-week pharmacotherapy split, time to first cigarette and daily cigarette consumption further differentiated observed abstinence probabilities. These profiles may help describe heterogeneity among service users; however, the tree was not developed or validated as a clinical prediction model and should not be used as an individual decision rule.

Our findings are consistent with the results of previous studies. Prolonged varenicline treatment, particularly beyond 5 weeks, appears beneficial for improving smoking cessation rates, aligning with broader efficacy observed in diverse populations.^{11 28} Guidelines also suggest the importance of tailored varenicline prescriptions to enhance smoking cessation success rates.²⁹ One study suggests that a minimum duration of 5 weeks may be necessary for NRTs to achieve optimal results in hospital-based settings.³⁰ However, recorded prescribed duration is not equivalent to medication exposure or adherence. Longer duration may be a marker of motivation, treatment persistence, provider support, tolerance of medication, service engagement or early treatment response. Conversely, individuals who relapse early, discontinue visits or experience adverse effects may have shorter recorded treatment duration. Reverse causation and residual confounding therefore remain plausible, and this study cannot establish that extending pharmacotherapy causes abstinence. The CART-derived 6-week split should be interpreted as a data-driven partition within the exploratory tree rather than an optimal treatment duration or clinically validated threshold. The complementary week-specific GEE analysis showed that the broader association with prescribed duration was present without imposing the binary 6-week split, although the pattern across individual weeks was not strictly monotonic. This analysis does not independently validate the stability of the exact 6-week CART cut-point; rather, it supports cautious interpretation of 6 weeks as an exploratory partition rather than a specific clinical treatment-duration threshold.

Multiple studies have consistently found that longer time to first cigarette in the morning was a significant predictor of successful smoking cessation.^31–34 A previous study of adolescent smokers found that smoking the first cigarette within 30 min after waking was associated with a greater likelihood of cessation failure.³² Shiffman et al even suggested that assigning the dose of nicotine gum based on the time to first cigarette could be an effective strategy.³⁵ The time to first cigarette after waking is an established indicator of nicotine dependence and has been associated with cessation outcomes.^31–36 It should not be interpreted as a causal behavioural target: asking a person to delay the first cigarette does not necessarily reduce underlying dependence or improve abstinence. In this study, longer time to first cigarette functioned as a marker that helped distinguish subgroups with different observed outcomes. The item-level GEE analysis likewise indicated that several FTND components contained outcome-related information; however, the small positive associations observed for some mutually adjusted items should not be interpreted as protective because the correlated item scores were entered simultaneously in the same model.

These results support cautious service-planning implications. Records characterised by shorter treatment engagement or stronger dependence indicators may identify groups that warrant additional follow-up and support, but intervention effects should be tested prospectively. Strategies such as adherence support, incentives or digital follow-up have been evaluated in other settings,^{36 37} yet the current observational analysis does not establish which intervention would improve outcomes in this programme.

The study period encompassed the COVID-19 pandemic, which may have affected access to and use of smoking cessation services. Annual programme record volume was lower during 2020–2021 than in 2019, whereas observed abstinence among records with valid follow-up remained broadly stable at approximately 40% during 2019–2022. In Taiwan, some participating settings, including community pharmacies and public health centres, also assumed additional epidemic-response responsibilities during the pandemic, including participation in the national mask-rationing system.³⁸ These operational demands may have affected routine service capacity, although such changes were not measured in our data. Calendar year was included in the adjusted GEE models; nevertheless, changes in service delivery, healthcare use or participant selection during the pandemic period cannot be excluded.

This study has several strengths. It used a large, nationwide smoking cessation service database spanning multiple care levels in Taiwan between 2012 and 2022. The person-year structure allowed repeated annual service use to be represented while preserving calendar-year treatment-course information. The CART approach provided an interpretable description of subgroup heterogeneity in observed abstinence probabilities, consistent with its use for describing sequential combinations of predictors in smoking cessation research.^{5 6} However, the findings should be interpreted in light of several limitations.

First, the main complete-case analysis was defined with respect to availability of a valid 6-month outcome. The comparison with excluded records suggested that missing outcome status was related to service engagement and institutional characteristics. Included records showed substantially greater treatment engagement than records without a valid outcome, particularly in cumulative prescribed weeks and number of treatment-course records. People who remained engaged may have been more likely both to complete follow-up and report abstinence, so the complete-case estimate may overstate programme-wide abstinence. Deterministic sensitivity analyses provide boundary estimates rather than a model-based correction and cannot eliminate selection bias. In addition, sex was unknown for 10.49% of person-year records, and multivariable analyses used records with available data for all variables required in the corresponding model.

Second, the outcome was self-reported 7-day point-prevalence abstinence assessed at 6 months and was not biochemically verified. Although self-reported smoking status has been used and evaluated in previous cessation studies,^{39 40} misclassification remains possible, particularly when compared with biochemical measures such as cotinine.⁴¹ This may have affected the estimated abstinence proportion.

Third, cumulative pharmacotherapy weeks represented recorded prescribed duration, not confirmed dispensing, actual medication use, ingestion or adherence. Prior studies have reported associations between pharmacotherapy duration and smoking abstinence,^{10 11 28–30} but recorded prescribed duration in routine service data should not be interpreted as verified medication exposure or adherence. Motivation to quit, prior quit attempts, socio-economic status, mental health, comorbidities, social support, provider quality, adverse effects and treatment adherence were unavailable. These unmeasured factors may have influenced both treatment engagement and self-reported 7-day point-prevalence abstinence assessed at 6 months, leaving residual confounding and limiting causal interpretation.

Fourth, the same service user could contribute records in multiple calendar years. GEE analyses accounted for within-user correlation, but the exploratory CART analysis did not explicitly model clustering. The tree was intended for subgroup profiling and was not externally validated for prediction; terminal-node probabilities should not be interpreted as stable individual risks, causal pathways or clinical thresholds. Race and ethnicity were not collected in the source database and therefore could not be evaluated. Finally, although the SGSCSP provides broad nationwide coverage across geographical regions and multiple levels of care, participation was service-based and did not constitute probability sampling of all smokers in Taiwan. The findings are therefore most directly generalisable to smokers who accessed the national smoking cessation service, particularly those with an observed 6-month outcome, rather than to the entire smoking population.

Conclusion

In this nationwide smoking cessation service dataset, an exploratory decision tree described person-year subgroups with different observed probabilities of self-reported 7-day point-prevalence abstinence assessed at 6 months, while GEE models quantified adjusted population-averaged associations. Cumulative prescribed pharmacotherapy weeks and nicotine-dependence indicators were prominent profiling variables, but their associations do not establish causal effects or clinically validated thresholds. Non-random availability of 6-month outcomes materially affected the estimated abstinence proportion. Prospective studies are needed to determine whether interventions that improve treatment persistence or address nicotine dependence increase abstinence.

Data availability statement

No data are available.