Quantitative relationship of effect estimates from spouse and disease controls in case control study

Abstract

Background

Disease control (DC) and spouse control (SC) are commonly used in case control study. The exposure level of these two kinds of control may be different, especially in smoking status. However, few studies concentrate on the association of DC and SC. Based on this situation, this study aims to explore the quantitative relationship of Odds Ratio (OR) between DC () and SC ( for lung cancer.

Methods

In this study, the data of cases, DC and SC were collected from the National Retrospective Survey of Deaths, including one million deaths during 1986 to 1988. Referring to cross-walk model, four forms of model, stratified by urban/rural areas, genders and 5-year age categories, were established to explore the suitable one, for adjusting and .

Results

After matching, a total of 57,789 participants were collected. Accordingly, the smoking prevalence of case, SC and DC groups were 54.39%, 31.72%, 33.54%, respectively. In the urban areas, the total of males and females were higher than , while the quantitative relation was the opposite in the rural areas. The and value gradually increases with the age group, reaching its maximum at 60–64 in urban area and 50–59 in rural area. After establishing the cross-walk model, in urban area, and were used for males and females, respectively. In rural area, and were used for males and females, respectively.

Conclusions

were commonly larger than in urban areas, whereas the opposite quantitative relation was observed in rural areas. Any modeling of the relationship between and is presented solely as a descriptive exercise for this historical dataset and should not be interpreted as a validated adjustment tool or a generally applicable method for translating effect estimates between control types. As an exploratory, proof-of-concept analysis, this study illustrates how effect estimates may vary across control definitions in case-control studies, highlighting a fundamental methodological caution.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04855-y.

Introduction

In systematic reviews, the definition and methods of data synthesis may vary depending on the study design, data type, or analytical objectives. Typically, data synthesis refers to the process of integrating, summarizing, and analyzing data from multiple independent studies, which aims to derive broader and more robust conclusions by combining the results of different studies [1]. To ensure the reliability and comparability of the results, it is necessary to standardize the data and make comparisons on a unified basis. Some researchers set up adjustive methods to estimate the different data sources, and the cross-walk model is one of the explorations. Cross-walk is performed by establishing models to adjust the different categories and definitions or consumption frequencies of research factors, so that they are unitized and comparable [2]. Cross-walk can map the indexes or results from one system to another by establishing quantitative relationships. When a certain research design or data type is lacking in some regions or time periods, the existing data can be extrapolated to other conditions through a cross-walk model. In Ng et al. study [2], researchers established regression models (cross-walk models) and adjusted different definitions or non-standardized age-gender and multi-source data of smoking product and consumption, and calculated the smoking prevalence of particular age-sex-country-year. In the Global Burden of Disease (GBD) study, this model was used in definitions or consumption frequencies of smoking, ambient particulate matter pollution, family size of household air pollution, alcohol consumption and physical activity [3, 4]. The above studies illustrate that adjustment methods such as cross-walk models can be useful when differences arise from varying definitions or measurement scales. However, when discrepancies originate from fundamentally different study bases rather than measurement definitions, statistical adjustment alone may not resolve the underlying methodological inconsistency. This distinction is particularly important in epidemiological case-control studies, where comparability of source populations is essential for valid inference.

In epidemiology, case-control study is a type of observational study commonly used to look at factors associated with diseases or outcome, which may exist selection bias and recalling bias. The selection of cases and controls is a crucial aspect of this study design. The key principle is that both groups should originate from the same study base. Study base comparability is widely regarded as the sine qua non of valid case-control inference, because violations of this principle may lead to systematic differences unrelated to the exposure-disease relationship itself. When controls arise from populations governed by different selection mechanisms, estimated associations may reflect these mechanisms rather than true causal effects. In general, there are several sources for constructing controls, including: general population; relatives or friends; or hospital patients [5]. The latter two types are commonly used in some case-control studies of China [6, 7]. DCs are deceased individuals recruited from hospitals, representing a patient-based population, whereas SCs are surviving spouses of deceased cases, representing a population defined by marital linkage and survival. This intrinsic incomparability in study base represents a central methodological challenge in case-control studies. Rather than assuming such differences can be removed through statistical correction, it is important to empirically examine how large the resulting discrepancies in effect estimates may become when different control definitions are used. Such comparisons can serve as cautionary examples illustrating the consequences of study base violations. For example, in the study of estimating the risk of smoking on cancer deaths in China [8, 9], the deceased who died from causes other than those related to smoking and surviving spouses of cancer deceased were invited as disease controls (DC) and spouse controls (SC), respectively. It is worth noting that different controls may bring certain differences in the research. Jiang et al. [10] compared these two types of control by calculating the ratio of and , and did not detect quantified difference between them.

Therefore, based on the previous studies, the primary aim of the present study is to use this historical dataset as a methodological case study to examine how effect estimates differ when spouse and disease controls are used. We further explore the association of DC and SC, and estimate the quantitative relationship of and in the context of lung cancer associated with smoking, utilizing the cross-walk model of the GBD study as an adjustment approach, to provide a methodological proof-of-concept using a unique but historical dataset. However, this study serves as a preliminary methodological demonstration, and when we have updated or other data, the method should be explored specially in different situations. The findings may provide methodological insights for systematic reviews and meta-analyses in epidemiological studies.

Besides, our data remains the only national resource on this topic that includes DC and SC information. Although the data are historical and may not represent current population structures, they provide a unique opportunity to examine methodological issues arising from control selection, which are not restricted to a specific time period. Furthermore, extensive epidemiological and experimental evidence has confirmed the causal link between smoking and lung cancer. The previous meta-analysis [11] showed that the Relative Risk (RR) of smoking-attributed lung cancer has remained stable (ranged from 2 to 6) in the past four decades in China, and the RR, OR and Hazard ratio (HR) were equally collected in the study. Thus, our study used the death data from 1986 to 1988, which contained both DC and SC, to explore the quantitative relationship between and . The analysis is intended primarily to demonstrate the methodological implications of using non-comparable control groups rather than to develop a transportable calibration model.

Method

Data source

The data source was derived from the National Retrospective Survey of Deaths, including one million deaths from 1986 to 1988. To contain a wide geographical distribution, this survey covered 24 cities, which were chosen non-randomly, and 74 rural countries were chosen by stratified random sampling from over 2000 counties in China. The survey had interviewed surviving family members of one million people who had died, and collected the information of the deaths and surviving spouses, including smoking status [12]. Smoking was defined as smoking or not. Therefore, we could obtain data on the case, DC and SC. This study restricted the analyses to aged 35 and above, and the ages of participants were split into 5-year age categories, and finally eight age groups were divided. The main outcome of this study is lung cancer.

Definition and cross of cases and controls

Cases means the deaths with lung cancer as the underlying cause of death (UCOD), DC means the deaths with UCOD irrelative to smoking. SC means the surviving spouses of the cases, excluding the people with smoking-related diseases. The diseases which are irrelated with smoking were shown in Table 1S, coded according to the International Statistical Classification of Diseases and Related Health Problems (ICD).

Because of the pronounced gender difference in smoking behaviors, these analyses were conducted separately for males and females. This gender-specific stratification minimizes potential confounding caused by the uneven distribution of smoking exposure across sexes [13]. Thus, the cross matching of case, DC and SC in male and female was conducted. In the meanwhile, the education was also added to the matching, using Propensity Score Matching (PSM) (Fig. 1).

Fig. 1.

The frame of cross matching. As for male cases, we chose the husbands of the female cases as male SCs, and other deaths who were not died for smoking related diseases in the same hospital as male DCs, and so as the selection principle of female SCs and DCs

Statistical analysis

Data cleaning and statistical analyses were implemented using SAS 9.4 (SAS Institute Inc., Cary, North Carolina, USA) and R 4.3.2 (R Foundation for Statistical Computing; https://www.r-project.org). The data were cleaned by duplicate checking, processing extreme values, abnormal values and missing values, logical verification and so on. Descriptive analyses were shown as numbers and percentages. Smoking and smoking rate stratified by urban/rural area and gender were calculated. The multiple comparisons of chi-squared test were used to analyze the categorical variables. Cochran-Mantel-Haenszel test was used to calculate the ORs with 95% Confidence Interval (CI) among gender and age (5-year age categories).

Referring to the cross-walk method [2, 14] on smoking prevalence proposed by GBD, four forms of models among different urban/rural, genders and age groups were established for exploring the fit one to adjust and . The models were established as follows:

Model 1: normal scale with an intercept: ;

Model 2: normal scale without an intercept: ;

Model 3: logit scale with an intercept: ;

Model 4: logit scale without an intercept: ;

where is the intercept of the model, is the coefficient of or , and is the residual of the model. Several indices were used to evaluate model fit, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Akaike’s Information Criteria (AIC) and Bayesian information criteria (BIC). The lower the above four indexes, the better the model fit. All tests were two-sided and P-value < 0.05 was considered to be statistically significant.

Results

Demographic characteristics

After matching, a total of 57,789 participants were collected, including 7,438 cases, SCs, and DCs in males and 11,825 cases, SCs, and DCs in females. The educational degrees of most participants were under senior high school, accounting for 81.97%, 81.97%, 84.67% among case, SC and DC in males, and the same proportion (81.87%) among the above corresponding three groups in females. Among the three groups of males, the largest proportion of age groups was 70 + years old, as the case (28.98%) and DC (46.55%) groups of females. The proportions of urban participants for case, SC and DC groups were similar among males and females, accounting for 80.86%, 82.60 and 91.58% in males and 87.10%, 82.97% and 95.68% in females (Table 1).

Table 1.

Demographic characteristics of the participants

Variate	Male			Female
	Case (%)	SC (%)	DC (%)	Case (%)	SC (%)	DC (%)
Total	7438	7438	7438	11,825	11,825	11,825
Educational degree
High	1341 (18.03)	1341 (18.03)	1140 (15.33)	2144 (18.13)	2144 (18.13)	2144 (18.13)
Low	6097 (81.97)	6097 (81.97)	6298 (84.67)	9681 (81.87)	9681 (81.87)	9681 (81.87)
Age
35–39	70 (0.94)	111 (1.49)	262 (3.52)	253 (2.14)	119 (1.01)	537 (4.54)
40–44	98 (1.32)	216 (2.90)	252 (3.39)	367 (3.10)	280 (2.37)	469 (3.97)
45–49	215 (2.89)	401 (5.39)	285 (3.83)	697 (5.89)	703 (5.95)	552 (4.67)
50–54	561 (7.54)	802 (10.78)	577 (7.76)	1310 (11.08)	1664 (14.07)	1052 (8.90)
55–59	960 (12.91)	1214 (16.32)	687 (9.24)	1668 (14.11)	2339 (19.78)	1216 (10.28)
60–64	1405 (18.89)	1449 (19.48)	799 (10.74)	2030 (17.17)	2704 (22.87)	1260 (10.66)
65–69	1473 (19.80)	1292 (17.37)	882 (11.86)	2073 (17.53)	2090 (17.67)	1235 (10.44)
70+	2656 (35.71)	1953 (26.26)	3694 (49.66)	3427 (28.98)	1926 (16.29)	5504 (46.55)
Urban/Rural
Urban	6014 (80.86)	6144 (82.60)	6812 (91.58)	10,300 (87.10)	9811 (82.97)	11,314 (95.68)
Rural	1424 (19.14)	1294 (17.40)	626 (8.42)	1525 (12.90)	2014 (17.03)	511 (4.32)

SC: Spouse Control, DC: Disease Control, High: Senior High School and above, Low: Under Senior High School

Smoking among urban/rural area and genders

The smoking rate of the total case, SC and DC groups were 54.39%, 31.72%, 33.54%, with 79.87%, 54.41% and 52.31% in males, and 38.36%, 17.45%, 21.73% in females, respectively. The smoking rates of males were consistently higher than those of females in three groups. Cases exhibited the highest smoking rate, followed by SC and DC in males, and DC higher than SC in females of urban and rural areas. As for males, the smoking rates of case were similar in urban and rural areas, and the SC and DC of rural areas were higher than those of urban areas. The smoking rates of case, SC and DC of females in urban areas were higher than those of rural areas (Table 2).

Table 2.

Smoking rate stratified by urban/rural area and gender

Smoking	Male			Female
	Case (%)	SC (%)	DC (%)	Case (%)	SC (%)	DC (%)
Total	5941 (79.87)	4047 (54.41)	3891 (52.31)	4536 (38.36)	2063 (17.45)	2569 (21.73)
Urban/Rural
Urban	4805 (79.90)	3232 (52.60)	3531 (51.84)	4229 (41.06)	1866 (19.02)	2542 (22.47)
Rural	1136 (79.78)	815 (62.98)	360 (57.51)	307 (20.13)	197 (9.78)	27 (5.28)

Odds ratio of two types of control

Two types of OR ( and ) were calculated, stratified by urban/rural, gender and 5-year age categories (Table 3). In the urban areas, the total of males and females was higher than , while the quantitative relation was the opposite in the rural areas. The total and of males was higher than that of females, except for the of males and females in the rural areas. In the urban area, the highest and of males and females occurred in 60–64 age group, except the highest of females occurred in 70 + age group. In rural area, the highest and occurred in 50–59 age group, that is, the and value gradually increases with the age group, reaching its maximum at 60–64 in urban area and 50–59 in rural area. Especially, in the rural area, the group and of males in the age group 40–44 had a relatively high OR, for the small sample of the related groups.

Table 3.

Estimated Odds Ratio (OR) and 95% Confidence Interval (CI) by region, gender and age group

Age	Urban				Rural
	(95%CI)		(95%CI)		(95%CI)		(95%CI)
	Male	Female	Male	Female	Male	Female	Male	Female
Total	3.69(3.40, 4.00)	3.03(2.84, 3.23)	3.44(3.18, 3.73)	2.34(2.21, 2.49)	2.43(2.04, 2.89)	2.42(1.99, 2.95)	2.70(2.18, 3.33)	4.04(2.69, 6.09)
35–39	3.27(1.43, 7.45)	0.82(0.31, 2.17)	5.02(2.38, 10.55)	1.02(0.50, 2.12)	2.60(0.83, 8.16)	1.76(0.20, 15.83)	2.14(0.68, 6.75)	0.45(0.36, 0.55)
40–44	1.78(0.99, 3.18)	1.53(0.86, 2.70)	2.66(1.51, 4.48)	1.78(1.10, 2.88)	2.92(0.86, 9.87)	2.20(0.55, 8.87)	5.76(1.70, 19.55)	4.43(0.53, 37.27)
45–49	3.66(2.36, 5.68)	1.86(1.37, 2.52)	2.83(1.80, 4.46)	1.45(1.08, 1.95)	1.48(0.73, 3.00)	2.09(0.93, 4.67)	1.00(0.41, 2.42)	2.57(0.56, 11.73)
50–54	3.17(2.40, 4.20)	2.36(1.97, 2.82)	2.79(2.09, 3.73)	1.93(1.60, 2.34)	2.33(1.29, 4.21)	3.27(1.66, 6.45)	4.22(2.08, 8.57)	2.09(0.69, 6.30)
55–59	3.57(2.86, 4.48)	2.55(2.19, 2.97)	3.53(2.76, 4.50)	1.97(1.67, 2.34)	2.28(1.41, 3.70)	3.47(2.10, 5.72)	2.39(1.22, 4.67)	4.76(1.41, 16.04)
60–64	4.66(3.83, 5.67)	3.61(3.14, 4.16)	3.81(3.07, 4.72)	1.97(1.67, 2.34)	2.40(1.62, 3.57)	1.62(1.10, 2.38)	2.62(1.40, 4.89)	2.95(1.01, 8.57)
65–69	3.42(2.83, 4.13)	3.46(2.98, 4.01)	3.34(2.73, 4.09)	2.57(2.19, 3.01)	2.36(1.54, 3.61)	2.91(1.82, 4.66)	3.93(2.26, 6.84)	5.84(1.36, 25.03)
70+	3.74(3.26, 4.29)	3.59(3.10, 4.15)	3.53(3.14, 3.98)	2.64(2.40, 2.91)	2.81(2.05, 3.81)	2.47(1.62, 3.76)	2.49(1.80, 3.44)	4.54(2.37, 8.71)

: Odds Ratio of Spouse Control, : Odds Ratio of Disease Control. Some subgroup estimates have wide confidence intervals due to small sample sizes, and these unstable estimates may reduce the reliability of subsequent regression modeling; therefore, the results should be interpreted cautiously

Selection of cross-walk model

Referring to cross-walk model, four forms of model were alternative to adjust the by . According to the selection criteria of the accuracy evaluation indexes, in urban area, and were selected for males and females, respectively. In rural area, was selected for males, and for females (Table 4). The regression lines of the four selected relational expressions were shown in the Fig. 2.

Table 4.

Accuracy evaluation indexes of cross-walk model

Region	Gender	Model	Evaluation indexes						95%CI of coefficients
			RMSE	MAE	ε	AIC	BIC	P-value
Urban	Male		0.70	0.52	0.81	23.00	23.20	0.37	(-0.67, 4.85)	(-0.40, 1.16)
			0.82	0.68	0.87	23.50	23.60	< 0.01	—	(0.79, 1.15)
			0.82	0.68	0.27	5.21	5.44	0.25	(-0.73, 1.63)	(-0.35, 1.57)
		*	0.24	0.20	0.26	3.94	4.09	< 0.01	—	(0.84, 1.12)
	Female		0.24	0.15	0.28	6.10	6.34	< 0.01	(-1.73, -0.25)	(1.38, 2.09)
			0.36	0.27	0.38	10.20	10.40	< 0.01	—	(1.14, 1.42)
		*	0.13	0.09	0.15	-3.54	-3.30	< 0.01	(-0.43, 0.07)	(1.17, 1.87)
			0.15	0.11	0.16	-3.31	-3.15	< 0.01	—	(1.14, 1.46)
Rural	Male		0.32	0.25	0.37	10.50	10.80	0.10	(1.21, 2.47)	(0.004, 0.36)
			0.82	0.78	0.88	23.60	23.80	< 0.01	—	(0.50, 0.85)
		*	0.13	0.10	0.15	-4.19	-3.95	0.03	(0.32, 0.80)	(0.07, 0.51)
			0.28	0.26	0.30	6.27	6.43	< 0.01	—	(0.56, 0.92)
	Female		0.57	0.45	0.66	19.60	19.90	0.25	(0.79, 2.91)	(-0.09, 0.45)
			0.97	0.78	1.04	26.30	26.40	< 0.01	—	(0.42, 0.82)
			0.23	0.18	0.27	5.12	5.36	0.23	(0.39, 1.01)	(-0.08, 0.40)
		*	0.48	0.32	0.51	14.90	15.00	< 0.01	—	(0.33, 0.87)

*: Models were selected for adjusting. RMSE: Root Mean Squared Error, MAE: Mean Absolute Error, ε: Residual, AIC Akaike’s Information Criteria, BIC Bayesian information criteria. The lower the above four indexes, the better the model fit. CI: Confidence Interval

Fig. 2.

Fitted regression lines of cross-walk model among two kinds of odds ratios

Discussion

The primary contribution of this analysis is to demonstrate the quantitative relationship between different controls, which arise when they originate from non-comparable study bases. Referring to the idea of cross-walk models, a quantitative relationship between and was explored only as an illustrative exercise within this dataset. Four forms of cross-walk model among urban/rural, genders and 5-year age categories were established to explore the suitable association. Among the two kinds of OR, are commonly larger than in the urban areas, while the association in the rural areas was opposite. Finally, of males and of females in urban area, of males and of females in rural area were selected. However, this model choice reflects goodness-of-fit within the current dataset and does not imply that the resulting relationship is generalizable to other populations without independent validation. For the selected four models, the confidence interval for the coefficient in urban males includes 0 (equivalently, an odds ratio of 1 on the original scale), indicating that the difference between and is not statistically significant in this subgroup. This lack of statistical significance likely reflects limited precision due to reduced sample sizes after matching and stratification, rather than evidence of no underlying methodological difference. Notably, in some subgroups, such as Urban males, the estimated relationship between and was not statistically distinguishable from a 1:1 correspondence, suggesting that differences between control definitions may be minimal in certain contexts. As the primary objective of this study was methodological calibration rather than hypothesis testing within individual strata, these coefficients were interpreted in the context of the overall model structure and cross-stratum patterns, rather than in isolation. Accordingly, non-significant subgroup estimates were not over-interpreted, and conclusions were drawn based on the consistency and direction of effects across groups.

When alternative models showed small differences in model fitting indexes, model selection was guided primarily by epidemiological interpretability. The selected model captures a monotonic increase in with the increase of . Importantly, in some groups, the exclusion of the intercept term implies that as smoking exposure approaches zero, the estimated odds ratio approaches one. This reflects a biologically and epidemiologically meaningful boundary condition, consistent with the principle that no exposure corresponds to baseline risk, rather than an arbitrary modeling constraint. From both clinical and epidemiological perspectives, this specification ensures that model predictions align with established smoking-disease relationships while maintaining comparability across populations and study designs. Therefore, the selected model represents a balance between statistical adequacy and meaningful interpretability, but its role in the present study is limited to illustrating how differences between control types may be summarized rather than providing recommended adjustment tools for future studies.

In this study, the smoking prevalence of case was the highest in the three groups, then SC and DC in descending order in males, and DC and SC in descending order in females. However, were commonly larger than in the urban areas, and the quantitative relation in the rural areas were opposite. Several non-mutually exclusive hypotheses may account for these urban-rural differences. In urban areas, greater similarity in lifestyle and health-related behaviors between cases and their spouses may lead to correlated smoking exposures, potentially reducing exposure contrast between cases and spouse controls. In rural areas, different selection mechanisms may operate, whereby disease controls more closely reflect the underlying exposure distribution than spouse controls, particularly among females. Importantly, these explanations are speculative and cannot be empirically evaluated with the available data. The present analysis was not designed to disentangle these mechanisms, and the results should therefore be interpreted as exploratory and possible rather than causal. Beyond spousal co-exposure, additional factors may contribute to the observed patterns, including differences in smoking-related social norms, representativeness of control groups, gender-specific smoking behaviors, secondhand smoke exposure, and potential information or reporting biases. These factors likely operate in combination, and their relative contributions cannot be distinguished within the current study framework.

Across both control definitions, ORs among males were consistently higher than those among females within the same strata. This pattern likely reflects the substantially higher prevalence of active smoking among males in the study population. In contrast, lung cancer risk among females may be influenced by factors not directly captured by the binary smoking variable, such as secondhand smoke exposure or household environmental exposures (e.g., cooking-related fumes), as suggested in previous studies [15–18]. However, these factors were not measured in the present analysis and therefore cannot be formally assessed.

As for the choice and difference of DCs and SCs, the advantages, confounding control and bias of them are as follow. On one hand, the advantage of DCs is that selecting controls with UCOD irrelated to smoking can reduce the risk of underestimation, and SCs can naturally control for living environment and habit, thereby reducing confounding bias. On the other hand, DCs and SCs both exist selection bias, arising from differences in socioeconomic status, living environments, and habits with the same sex cases. Moreover, cases, DCs and SCs all have information bias, particularly related to recall by spousal recalling. SCs also introduce co-exposure bias, as spouses often have similar lifestyles; for instance, if one spouse smokes, the other may be exposed to secondhand smoke, or they may both smoke or not smoke together, which can diminish the smoking exposure difference between cases and controls, potentially underestimating the risk between smoking and lung cancer.

An important methodological issue that warrants explicit discussion is the potential violation of the study base assumption when comparing and . In case-control studies, valid estimation of exposure-disease associations relies on cases and controls being sampled from the same underlying source population. When this assumption is violated, observed differences in ORs may reflect selection mechanisms rather than true differences in etiologic effects. In the present study, SCs may not represent the same study base as DCs. Spouses tend to share socioeconomic characteristics, household environments, and health-related behaviors with cases, which may lead to correlated smoking exposure distributions and systematically different baseline risks. DCs, while often closer to the general patient population, may also deviate from the ideal study base depending on the disease spectrum and its association with smoking. Consequently, the observed differences between and should not be interpreted as purely correctable statistical discrepancies. Rather, they likely reflect a combination of study base violation and selection-related mechanisms inherent to different control definitions. Our analysis therefore aims to empirically characterize these differences and illustrate their potential magnitude, rather than to determine which control type yields a more valid causal estimate. This distinction is critical for interpreting the results and for understanding the limitations of applying any calibration across control types. These findings have practical implications for evidence synthesis: meta-analyses combining case-control studies should recognize that studies using spouse controls and disease controls are methodologically distinct. Sensitivity analyses or separate pooling strategies may therefore be necessary to avoid conflating estimates derived from non-comparable study bases.

In some studies [2, 14, 19], due to different definitions or consumption frequencies, methods are established to adjust them, and cross-walk model is one of the methods. Referring to the thinking of cross-walk, we illuminated to explore the association of OR between DC and SC by adopting this model. Via establishing regressions model, cross-walk used to adjust and by , which quantized the relationship of two controls. However, when differences arise from fundamentally different study bases, statistical cross-walk adjustments cannot resolve the underlying selection problem and can only describe empirical discrepancies observed in a specific context. It is worth mentioning that we should choose the cross-walk model according to different scenarios, which is not necessarily linear regression model. Besides, in the majority of systematic review studies, ORs might be pooled without considering their distinction or quantitative adjustment, which may cause some bias, and cross-walk model is a good attempt to quantify data relationship. The important condition of cross-walk model is that two or more definitions or consumption frequencies must be collected, which allows them to unitize and adjust. This study provides an exploratory illustration of how effect estimates derived from different control definitions may differ in case-control studies, highlighting that statistical adjustments cannot substitute for proper study base comparability in case-control design.

Although the underlying data were derived from mortality surveillance conducted during 1986–1988, they remain the only national dataset containing information on both disease controls and spouse controls. Our findings suggest that and estimates in smoking-related lung cancer studies are not directly comparable, and any apparent quantitative relationship should be interpreted cautiously as dataset-specific and descriptive. Rather than proposing a validated adjustment framework, this work illustrates the extent to which effect estimates may vary when control groups originate from different source populations, underscoring a broader methodological concern relevant to meta-analyses and comparative epidemiologic studies. Therefore, the estimated relationships should be regarded as exploratory and context-specific, and should not be used for prediction, extrapolation, or routine adjustment without independent validation in larger and more representative datasets. Future research with improved data comparability may help clarify how differences in control selection influence estimated associations.

Limitation and strength

This study has several limitations. Firstly, the potential violation of the study base assumption, as SCs and DCs may not originate from the same underlying source population, fundamentally limiting the causal interpretation of differences between and . Secondly, the study relied on mortality data collected during 1986–1988, representing a historical smoking and disease context that differs substantially from contemporary populations. Changes over time in smoking prevalence, gender-specific smoking patterns, tobacco control policies, healthcare access, and disease classification mean that these data do not merely limit generalizability but constitute a fundamental constraint on the current relevance and applicability of the findings. Thirdly, due to data limitations and reduced sample sizes after stratification, no internal or external validation could be performed. Consequently, the derived relationships remain descriptive of this single dataset and cannot be recommended for predictive use or adjustment in other populations without independent validation. Fourthly, residual confounding from unmeasured factors such as occupation, alcohol use and smoking cessation may have affected the observed OR differences, despite matching and stratification on available covariates. Fifthly, matching and stratification by sex, region, and age group reduced sample sizes in some subgroups, resulting in wide confidence intervals and inclusion of these unstable estimates may have influenced regression coefficients. Therefore, all modeled relationships should be interpreted cautiously as exploratory and descriptive rather than as stable quantitative mappings.

Fortunately, the data perfectly included two types of control, and the OR was relatively stable in the past decades. Unlike previous applications of cross-walk models in GBD research, our study innovatively utilizes this approach to quantify the relationship of OR between different types of controls in case-control studies. This methodological exploration provides a new perspective for data harmonization in systematic reviews and meta-analyses. Moreover, future research should validate and update this model if more recent data is available.

Conclusion

were commonly larger than in urban areas, whereas the opposite quantitative association was observed in rural areas. Any modeling of the relationship between and is presented solely as a descriptive exercise for this historical dataset and should not be interpreted as a validated adjustment method or a solution to the methodological challenge identified. Therefore, this study primarily provides an empirical characterization of how effect estimates may vary across control definitions, highlighting a fundamental methodological caution for interpreting and synthesizing results from case-control studies.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

DC

Disease control

SC

Spouse control

GBD

Global burden of disease

OR

Odds ratio

RR

Relative risk

HR

Hazard ratio

UCOD

Underlying cause of death

ICD

International statistical classification of diseases and related health problems

CI

Confidence interval

RMSE

Root mean squared error

MAE

Mean absolute error

AIC

Akaike’s information criteria

BIC

Bayesian information criteria

Author contributions

ZMC organized and analyzed the data, interpreted the results and drafted the manuscript. XNZ and XW supported the data acquisition process. XW designed the study, guided its implementation. XZ and XW helped with the interpretation of the results and revised the manuscript. All authors read and approved the final manuscript.

Funding

(1) State Key Laboratory Special Fund (2060204); (2) Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences (2023-I2M-2-001); (3) Strengthen Capacity of Study and Application on the Burden of Disease in Health Care Systems in China: Establishment and Development of Chinese Burden of Disease Research and Dissemination Center (15–208) supported by the China Medical Board (CMB).

Data availability

The data is not publicly available, but is available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Ethical approval from the INSTITUTIONAL REVIEW BOARD of Chinese Academy of Medical Sciences was obtained [20]. This study was conducted in accordance with the Declaration of Helsinki and followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for case-control studies. Retrospective information collected was limited to smoking exposure only, and oral informed consent for this study was obtained.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.