Study design and population
The present longitudinal study used data from regular panel waves 1, 2, 4, 5, 6, and 7 of the Survey of Health, Ageing and Retirement in Europe (SHARE), a biannual survey recruiting individuals aged 50 or older from European countries and Israel [15, 16]. Wave 3 lacked data on HGS and was dismissed for the present study. Representativeness of SHARE waves stems from a multi-stage stratified sampling design in which included countries are divided into different strata according to their geographical area. The number of countries included in SHARE has been progressively increasing with each SHARE wave, thus there are countries with longer follow-up periods than others, and 50% of participants having 2 or more follow-ups. Municipalities or zip codes within these strata are considered the primary sampling units. Data used in SHARE were collected through home computer-assisted personal interviews from February 2004 to January 2019. Ex-ante harmonization was conducted to ease the comparability among countries and new respondents were added to compensate for the attrition bias due to losses from each wave. Participants aged 50 years or older and who were free from any prior heart attack or stroke diagnosis at study entry were considered in the current study (n = 122 676). Duplicated or overlapped observations as well as participants with missing values regarding time and death cause or unreliable values concerning covariates were excluded from the analyses (n = 1560). Missing values from included participants were estimated using multiple imputation (n = 30 691). Figure 1 shows more descriptive information of the study sample. The present study was reported according to Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) [17].
Handgrip strength (exposure)
HGS of both hands was measured twice using a handheld dynamometer (Smedley, S Dynamometer, TTM, Tokyo, 100 kg). According to the SHARE protocol [15], participants were instructed to maintain the elbow in a 90° angle flexion while either standing or sitting, with a neutral wrist position, and upper arm set vertical against the trunk. Interviewers verbally encouraged participants with standardized instructions to squeeze with maximum effort for a few seconds. HGS was defined as the maximum value reached in either hand. Because HGS in relation to BMI was identified as better predictor than absolute values of HGS alone for outcomes such as cancer [12], HGS was divided by BMI and thereafter standardized using sex-specific mean and standard deviation of the whole sample ([X − Mean] ÷ SD) [18]. For the purpose of this study, exposure of both absolute HGS, and relative HGS over all-cause and cardiovascular mortality were examined.
All-cause, cardiovascular, heart attack, and stroke mortality (outcomes)
Participants were followed throughout the study period to determine mortality status. When deceased, information concerning both date and cause of death was obtained from a proxy interview (i.e., a relative, a household member, a neighbour, or any other person close to the deceased participant); in such case, mortality was determined through the following question: “What was the main cause of respondent’s death?” The range of potential answers comprised cancer, heart attack, stroke, other cardiovascular disease related illnesses (heart failure and arrhythmia), respiratory, digestive, or severe infectious disease, and other causes. For all-cause mortality, participants were categorized into alive and deceased, whereas participants deceased due to heart attack, stroke, and other related-cardiovascular events were grouped as deceased due to cardiovascular mortality. Specific death due to either heart attack or stroke were also categorized as alive or deceased due to either of these causes.
Covariates
Based on a literature review, we explored potential causal and confounding pathways between HGS and all-cause and specific cardiovascular mortality using a directed acyclic graph (Supplementary material, Figures S1-S2). Self-reported age and sex, country of residence at the time of interview, education, body mass index, alcohol consumption, smoking habit, physical inactivity, fruits and vegetable consumption and high blood pressure were identified as critical potential confounders in the main model. Education was self-reported by participants and thereafter coded using the 1997 version of the International Standard Classification of Education [19]. BMI was calculated from self-reported height and weight and subsequently grouped into 4 categories according to standards proposed by the World Health Organization (WHO) [20]. Alcohol consumption was estimated through the following question: “How many days a week did you consume alcohol during the last six months?” and answers included the following possible options: “Almost every day”, Five or six days a week, “Three or four days a week”, “Once or twice a week”, “Once or twice a month”, “Less than once a month”, “Not at all in the last 6 months”, “Refusal to answer”, or “Don´t know”. Smoking habit was assessed through the following question: “Have you ever smoked cigarettes, cigars, cigarillos, or a pipe daily for a period of at least one year?”, with potential answers comprising “No”, “Yes”, “Refusal to answer”, or “Don´t know”. Physical inactivity was determined through two questions: “How often do you engage in vigorous physical activity such as sports, heavy housework, or a job that involves physical labour”, and “How often do you engage in activities that require a moderate level of energy such as gardening, cleaning the car, or doing a walk?”. Participants selecting the option of “Hardly ever, or never” in the two questions were considered to be physically inactive. Fruits and vegetables consumption was assessed with the following question: “In a regular week, how often do you consume a serving of fruits or vegetables?”, and potential answers comprised the following options: “Refusal”, “Don’t Know”, “Everyday”, “3–6 times a week”, “Twice a week”, “Once a week”, and “Less than once a week”, Hypertension and Diabetes diagnosed conditions were self-reported answering the question: “Has a doctor ever told you that you had any of the conditions (…)?”. If the participants responded affirmatively to the “high blood pressure or hypertension” or “diabetes or high blood sugar” listed options, they were considered to have either hypertension or diabetes respectively. Medication (i.e., medicines for treating chronic conditions) was assessed through the following question: “Do you currently use drugs at least once a week for problems mentioned on this card?” This variable was re-coded into the categories “None” for those who answered such option in the survey, and “Any” for those who took one or more of a lists of drugs.
Statistical analyses
We estimated the risk of the different types of mortality in relation to HGS. To address the time-varying confounding bias derived from the consecutive measurements of both exposure and covariates, we used an MSM [21]. This modelling approach was used because follow-up levels of time-varying covariates may simultaneously be confounders for later HGS and mediators for earlier HGS, and thus cannot be appropriately adjusted using standard methods. In the context of the current study, our model considered age at baseline, sex, and country as fixed (i.e., time-invariant) variables whereas the rest of covariates were assumed to possibly vary throughout the follow-up period. To account for time effects, natural cubic splines with knots placed at the 5th, 50th, and 95th percentiles of the time distribution and time-on-study in months variable were also included in the model. This model was fitted in a two-step process; first, we calculated each participant-specific inverse probability of treatment weights (IPTWs) based on the inverse of the predicted probability of a participant experiencing the exposure that they actually experienced. Secondly, the exposure–outcome association was estimated using a pooled logistic regression in which we modelled the probability that each individual was exposed in each wave using IPTWs stabilized weights. To account for informative censoring, we fitted logistic regression models to estimate inverse probability of censoring weights at each time interval. As with our IPTW, we derived the same models for the numerator and denominator of the stabilized inverse probability of censoring weights. The final stabilized weights were calculated by multiplying the exposure and censoring weights. Finally, we used the cluster option to derive robust standard errors allowing for clustering of effects within each participant. We conducted all statistical analyses in Stata version 16.1 (StataCorp, Texas, USA). The results were visualized as forest plots and estimations were provided as HRs and their 95% confidence intervals (CIs).
Sensitivity analyses
To further test the robustness of our estimates, we conducted three different sensitivity analyses. First, we adjusted the main model for disease-related confounders (i.e., medication and diabetes diagnosis) instead of lifestyle-related factors (i.e., physical inactivity and fruits and vegetables consumption) in the alternative model (Model 2) (Supplementary material, Figure S3). Second, as body mass index might be considered a potential mediator of the association between HGS and mortality, we carried out sensitivity analyses excluding it (Supplementary material, Figure S4). Finally, we repeated the main model with no imputation of missing values (i.e., with observed values only), (Supplementary material, Figure S5).