Determinants of non-response in a longitudinal study of participants in the Women and Alcohol in Gothenburg project

ABSTRACT

Longitudinal assessment is useful for tracking patterns of alcohol use over time. Non-response is a common feature of longitudinal design and can bias estimates of alcohol use if there exist systematic differences between respondents and non-respondents. We investigated whether alcohol use, health status, and sociodemographic characteristics were determinants of non-response in a longitudinal cohort of women in the general population. We used data from a stratified, random sample of 479 women born in 1925, 1935, 1945, 1955, 1965, and 829 women born in 1970 and 1975, who were initially selected as participants in the Women and Alcohol in Gothenburg project. Results from multivariable logistic regression revealed that problematic alcohol use, depression, poor self-rated physical health, and basic education were associated with increased odds of non-response among women born in 1925, 1935, 1945, 1955, and 1965. Among women born between 1970 and 1975, older age and being unmarried increased the odds of non-response at follow-up. Surprisingly, problematic alcohol use and poor health were not associated with non-response in these younger birth cohorts. This study finding suggests that approaches to improve future survey response rates need to consider factors of greatest relevance to birth year and age.

Graphical abstract

KEYWORDS:

• Alcohol use
• health
• heavy episodic drinking
• high alcohol consumption
• non-response

Introduction

Alcohol use among women is a subject of public health concern. At equal levels of alcohol consumption, women tend to have higher blood concentration and are more likely to suffer intoxication and liver cirrhosis than men do (Bradley et al. 1998; Sugarman, Demartini, and Carey 2009). On the other hand, at lower levels of consumption, women are at greater risk of injury and all-cause mortality than men are (Bradley et al. 1998). Excessive alcohol use among women is a risk factor for breast cancer, stroke, and adverse pregnancy-related outcomes, such as infertility and spontaneous abortion (Bradley et al. 1998). Indeed, the increased health risk associated with women’s alcohol use calls for longitudinal assessment of alcohol use in order to better monitor patterns of alcohol consumption over time. This will help in early detection of high-risk women, and in developing appropriate and targeted preventive interventions (Greenfield and Kerr 2003).

Non-response is a major problem in longitudinal studies and can potentially affect estimates of alcohol use. Non-response results in biased estimates if those who remain in a study systematically differ from those who are lost to follow-up (McCoy et al. 2009). Non-response bias can lead to erroneous conclusions about the relationship between exposures and outcomes, limit extrapolation of findings, and compromise research resources. Addressing non-response will aid correct interpretation of study findings and provide useful information in making decisions regarding methods for dealing with non-response as well as possible measures for improving participation rates in future studies (Lamers et al. 2012; Nicholson, Deboeck, and Howard 2017).

Previous research has identified low education and being unmarried (Thygesen et al. 2008), female gender (Cunrandi et al. 2005), younger age (Lamers et al. 2012), poor self-rated health (Torvik, Rognmo, and Tambs 2012), and smoking (Clemens et al. 2007) as factors that may increase non-response in alcohol-related studies. Some studies also reported high alcohol consumption (Thygesen et al. 2008) and heavy drinking (Torvik, Rognmo, and Tambs 2012) as determinants of increased non-response. However, these studies combined data from male and female participants, ignoring the potential variations by gender in alcohol use (Wilsnack et al. 2009), which might modify the associations between alcohol use and non-response. Studies that conducted gender-stratified analysis did not find any associations between alcohol use and non-response among women (Caetano, Ramisetty-Mikler, and McGrath 2003; Goldberg et al. 2006). The sample in the study by Caetano, Ramisetty-Mikler, and McGrath (2003) constituted of couples and that in the study by Goldberg et al. (2006) was restricted to women aged 35–50, which raises the question of whether the findings can be extrapolated to other population groups. In the present study, we investigated non-response in the Swedish Women and Alcohol in Gothenburg project and assessed whether alcohol use, health status, and sociodemographic characteristics were determinants of non-response.

Materials and methods

The Women and Alcohol in Gothenburg (WAG) project is a longitudinal, population-based study that aims to increase knowledge about alcohol dependence and abuse (ADA) among women (Spak and Hällström 1996). Data were collected by using a 13-item alcohol-screening questionnaire (SWAG) followed by structured interviews.

In 1986, the SWAG was mailed to all women (N = 3130) who were born in 1925, 1935, 1945, 1955, and 1965, (cohort I) and registered in Gothenburg (response rate, 78%). In 1995, the procedure was applied to women (N = 2910) born in 1970 and 1975 (cohort II). The response rate was 77%. Based on the scores on the SWAG, 479 and 829 women from cohort I and II, respectively, were selected by stratified, random sampling, and invited to a face-to-face interview (Hensing and Spak 2009). Women in cohort I had their baseline interview (T1) in 1989–1990 (response rate, 83%), with follow-up interviews in 1994–1998 (T2) and in 2000–2002 (T3). Written reminders were sent to non-responders and, if necessary, telephone calls were made. The response rates at T2 and T3 were 67% and 50%, respectively. The baseline interview (T1) for women in cohort II was conducted in 1995–1998 (response rate, 74%), with follow-up interviews in 2000–2002 (T2) and in 2013–2015 (T3). The response rates at T2 and T3 for this group were 58% and 39%, respectively. Health professionals and social workers with several years of practice conducted the interviews, either at the University of Gothenburg or at the respondent’s home. The interviews lasted between 1.5 and 3.0 hours. Respondents who could not complete the full interview were offered a shorter version of the interview focusing on alcohol and alcohol-related problems. These respondents were excluded from the current study, as they did not have information on all relevant study variables. The Ethics Committee of the Medical Faculty at the University of Gothenburg approved the study (Dnr: 320–85, 158–94, 591–99 and 955–11). Oral informed consent (in 1989–1990, 1994–1998 and 2000–2002) and written informed consent in 2013–2015 was obtained from the interviewees, and all were informed of available helplines and support services. The change to written consent was done as a response to the Ethical Review Act introduced in 2003 with a demand of documentary of consent. A written consent was not obligatory as documentation, but for this study, it was chosen as a relevant procedure. All data in WAG have been treated with full confidentiality and in line with ethical principles and laws since the start of the project.

Information on alcohol use, health status, and sociodemographic characteristics were obtained during all three measurement times. The frequency of any alcohol use in the past month (none, 1–3 times/month, ≥2 times a week, and daily or almost every day) and in the past year (none/less than once per month, 1–3 times/month, ≥2 times a week, and daily or almost every day) were assessed. Because of the low prevalence of those reporting ≥2 times a week, and daily or almost every day, both response categories were combined in the analysis to increase statistical power. Participants who reported drinking any alcohol were asked about the amount of alcoholic drinks consumed during a typical day in the past year (≤2 drinks, 3–4 drinks and >4 drinks). Information on the quantity and frequency of beverage-specific alcoholic drinks, e.g., wine, beer, or spirits, in the past month and in the past year were used to construct variables describing high alcohol consumption (HAC) and heavy episodic drinking (HED). HAC was defined as drinking at least 20 g of ethanol/day on average in the month before the interview. HED was defined as drinking ≥72 g/day of ethanol per occasion, at least once a month in the 12 months before the interview. Diagnosis of ADA (with and without) was made according to DSM-III-R and based on CIDI-SAM (Cottler, Robins, and Helzer 1989). Health-related variables included self-rated physical health (SRH) during the last year (good and poor), current tobacco use (yes and no), lifetime depression (with and without), and lifetime anxiety (with or without). The last two measures were assessed according to DSM-III-R in 1990 and 1995 and according to DSM-IV (American Psychiatric Association 1995) in 2000. Sociodemographic variables included age (grouped into three in cohort I and into two in cohort II), marital status (married/cohabiting and unmarried), education (primary, secondary and college/university), occupation (employed and unemployed), and personal income in Swedish kronor (<100,000, 100,000–200,000, and >200,000). The study outcome was non-response (respondents and non-respondents) at each follow-up interview, with non-respondents referring to women who, for example, were unreachable, did not reply to the invitation, or voluntarily declined participation.

Statistical analyses were conducted with SPSS Statistics version 25.0. All analyses were population weighted to correct for bias due to the sampling design. Chi-square test and multivariable logistic regression were performed. T1 characteristics were used to investigate non-response at T2, and the T2 characteristics to study non-response at T3 to minimize eventual bias due to time-varying characteristics. All sociodemographic and health-related variables that were significantly associated (p < .05) with non-response in the univariate analysis, or that have been previously reported as determinants of non-response in alcohol studies, i.e., age, education, occupation, and marital status (Lamers et al. 2012; Thygesen et al. 2008), were selected and simultaneously adjusted in the logistic regression model. The criteria for variable selection were used to ensure that only the most important determinants were studied since we had small sample size and needed to maximize statistical power. Because the WAG study focused on alcohol use among women, the independent role of each alcohol variable, irrespective of the p-value in the univariate analysis, was tested in a model containing the selected sociodemographic and health-related variables from the univariate analysis. Separate analysis was performed for women in cohort I and II since birth cohorts have been reported as having a strong influence on alcohol use (Kraus et al. 2015). Results of the regression analyses are presented as odds ratio (OR) with 95% confidence intervals (95% CI).

Results

Among the women in cohort I, 80 (17%) out of the 479 that were randomly selected for T1 interview did not participate. Of these, 21 declined further participation in the study, 34 declined participation at T1, seven died during time lag between screening and T1 interview, six emigrated, one was unable to speak Swedish, and 11 did not provide reasons for not participating. Figure 1 provides detailed information on the response status at each measurement time point for women in cohort I who participated in the full interview at T1. Fifty-four percent (n = 186) of the 345 participants at T1 completed all three full interviews. Of the women in cohort II, 214 (26%) out of the 829 randomly selected for T1 interview did not participate. Of these, 128 declined further participation, 25 emigrated while 61 could not be located. Figure 2 shows the response status at each measurement time point for women in cohort II who participated in the full interview at T1. Thirty-two percent (n = 175) of the 543 participants at T1 completed all three full interviews. In general, non-response in both cohorts was monotone, in that few women who left the study reentered at a later wave (data not shown).

Figure 1. Flow chart showing response status at each interview time for women in cohort I

Figure 2. Flow chart showing response status at each interview time for women in cohort II

Based on the result from the univariate analysis (supplementary Tables 1 and 2), age, education, marital status, occupation, self-rated health, and depression were selected and further evaluated in a multivariable logistic regression. Table 1 shows the result of the association between the variables and non-response for both cohort I and II. Among women in cohort I, those with a basic education had elevated risk of not responding at T2 compared with those with a college/university degree (OR, 3.01; 95%CI, 1.06–8.51). Those with depression had increased risk of not responding at T2 compared with those with no depression (OR 7.26, 95%CI 1.64–32.23). Women with poor SRH were more likely to be non-responders at T3 than those with good SRH (OR 3.38, 95%CI 1.47–7.95). Those belonging to age group 35/45 were less likely to be non-responders at T3 compared with those belonging to age group 25 (OR 0.21, 95%CI 0.06–0.69). Among women in cohort II, unmarried women had a higher likelihood of not responding at T2 than married or cohabiting women (OR 1.94, 95%CI 1.10–3.42). Compared with those belonging to age group 20, women belonging to age group 25 had higher likelihood of not responding at T3 (OR 1.79, 95%CI 1.02–3.16). Regarding the alcohol use variables, after controlling for selected sociodemographic and health variables, among women in cohort I, the individuals with HED had higher odds for not responding at T2 than those with no HED (OR 3.13, 95%CI 1.11–8.88). Women with HAC had higher odds of not responding at T3 than those with no HAC (OR 4.76, 95%CI 1.20–18.91). None of the alcohol use variables was associated with non-response among women in cohort II (Table 2).

Table 1. Associations between selected sociodemographic and health characteristics measured before the planned interview and non-response at follow-up interviews based on data from the Women and Alcohol in Gothenburg study: results from logistic regression

	Non-response at follow-up interview
	Cohort I		Cohort II
	T2	T3	T2	T3
	n= 51	n = 60	n = 137	n = 163
Sociodemographic and health variables ^a	Odds ratio (95%CI)	Odds ratio (95%CI)	Odds ratio (95%CI)	Odds ratio (95%CI)
Age in years (Cohort I)
35/45 vs. 25	0.99 (0.37-2.66)	0.21 (0.06-0.69)
55/65 vs. 25	0.74 (0.18-3.09)	0.44 (0.12-1.62)
Age in years (Cohort II)
25 vs. 20			0.89 (0.50-1.58)	1.79 (1.02-3.16)
Education
Basic vs. college/university	3.01 (1.06-8.51)	1.81 (0.58-5.68)	1.49 (0.60-3.69)	3.40 (0.87-13.34)
High school vs. college/university	0.91 (0.32-2.58)	1.07 (0.36-3.14)	1.08 (0.65-1.80)	1.28 (0.71-2.33)
Marital status
Unmarried vs. married/cohabiting	1.03 (0.41-2.58)	0.74 (0.27-2.00)	1.94 (1.10-3.42)	1.22 (0.65-2.28)
Employment status
Unemployed vs. employed	0.76 (0.23-2.45)	1.88 (0.60-5.92)	1.17 (0.69-2.00)	1.26 (0.64-2.48)
Lifetime depression
With vs. without	7.26 (1.64-32.23)	0.78 (0.14-4.33)	0.52 (0.21-1.25)	1.30 (0.55-3.11)
Self-rated physical health
Poor vs. good	1.87 (0.68-5.13)	3.38 (1.47-7.95)	0.90 (0.51-1.58)	0.70 (0.31-1.60)

^aAll variables were entered simultaneously into the model; T2, first follow-up interview; T3, second follow-up interview.

Table 2. Associations between alcohol use variables measured before the planned interview and non-response at follow-up interviews based on data from the Women and Alcohol in Gothenburg study: results from logistic regression

	Non-response at follow-up interview
	Cohort I		Cohort II
	T2	T3	T2	T3
	n= 51	n = 60	n = 137	n = 163
Alcohol use variables ^a	Odds ratio(95%CI)	Odds ratio(95%CI)	Odds ratio(95%CI)	Odds ratio(95%CI)
Frequency of alcohol use last month ^b
1-3 times in a month	0.37 (0.11-1.18)	0.33 (0.09-1.34)	1.15 (0.51-2.61)	1.05 (0.44-2.48)
2 or more times in a week//daily	1.90 (0.65-5.59)	0.70 (0.16-3.13)	0.74 (0.32-1.74)	1.26 (0.54-2.94)
Frequency of alcohol use last year ^c
1-3 times in a month	1.02 (0.26-3.98)	0.81 (0.23-2.88)	1.15 (0.51-2.61)	1.75 (0.83-3.69)
2 or more times in a week/daily	3.82 (0.94-15.39)	1.76 (0.47-6.62)	0.74 (0.32-1.74)	1.70 (0.80-3.61)
Amount of drinks consumed a typical day last year ^d
3-4 drinks vs. 2 or less	1.36 (0.46-4.07)	1.03 (0.35-2.95)	0.87 (0.43-1.74)	1.08 (0.56-2.07)
>4 drinks vs. 2 or less	2.01 (0.60-6.71)	0.45 (0.09-2.12)	1.23 (0.61-2.49)	1.04 (0.48-2.26)
HAC last month ^e
Yes vs. no	1.58 (0.37-6.63)	4.76 (1.20-18.91)	0.67 (0.28-1.55)	0.80 (0.26-2.43)
HED last year ^f
Yes vs. no	3.13 (1.11-8.88)	0.61 (0.18-2.03)	0.60 (0.32-1.12)	1.29 (0.68-2.45)
ADA ^g
Yes vs. no	1.10 (0.28-4.37)	0.49 (0.10-2.42)	0.82 (0.38-1.79)	0.83 (0.22-3.18)

^aEach alcohol use variable was adjusted for age, education, marital status, employment status, depression and self-rated physical health.

^bReference category is none.

^cReference category is none/less than once a month.

^dReference category is ≤2 drinks.

^eHAC, high alcohol consumption.

^fHED, Heavy episodic drinking.

^gADA, alcohol dependence and abuse; T2, first follow-up interview; T3, second follow-up interview.

Discussion

We found that problematic alcohol use (HED and HAC), basic education, depression, and poor SRH significantly increased the odds for non-response in women born in 1925, 1935, 1945, 1955, and 1965, while being aged 35/45 years was associated with a decreased OR. Most of the characteristics, except for marital status and age, were not related to non-response among women born in 1970 and 1975.

That problematic alcohol use was associated with non-response only in the older birth cohorts was a surprise finding. One question is whether this finding might be driven by a possible increased risk of poor health and death among women in the oldest age group. We investigated this assumption by repeating the analysis after excluding women belonging to the age group 55/65 at baseline. Whereas the association between HED and non-response attenuated and became statistically non-significant (OR 2.25, 95%CI 0.81–6.26), the association between HAC and non-response at T3 remained unchanged (OR 4.27, 95%CI 1.21–15.09). Thus, suggesting that possible poor health or death among women in the oldest age group does not entirely account for the association between problematic alcohol use and non-response among women in the older birth cohort. Previous studies did not find any associations between high or frequent alcohol use and non-response among women (Caetano, Ramisetty-Mikler, and McGrath 2003; Goldberg et al. 2006). Our finding might be a generational effect, which may be explained by differences in norms between the birth cohorts. The oldest cohorts in our study were born and brought up during the rationing book period (1919–55) in Sweden, when alcohol use among women was perceived as an unacceptable behavior, and thus strictly regulated (Abrahamson and Heimdahl 2010). These women may have internalized their moral attitude toward alcohol use such that those who reported problematic alcohol use may perceive themselves as deviating from the norm and thus may present themselves with guilt and shame (Abrahamson 2012), which may have led to their refraining from participation in subsequent interviews. This may not be the case for the younger birth cohort who entered into adolescence and adulthood at a time of more permissible norms and diminished gender differences in alcohol use. In Sweden, women’s alcohol consumption began to increase in the early 1990s, particularly after Sweden joined the European Union in 1995 and adapted policies that encouraged increased availability of alcoholic drinks (Källmén et al. 2011; Ramstedt 2010).

Our finding regarding basic education and increased non-response is consistent with previous research (De Graaf et al. 2013), likewise the findings regarding poor SRH (Torvik, Rognmo, and Tambs 2012). The OR for lifetime depression as predictive of non-response is remarkably high but the wide confidence interval makes it difficult for us to draw conclusions. Older women in cohort I had lower OR for non-response at T3 than those in younger age group. Caetano, Ramisetty-Mikler, and McGrath (2003) found that women aged 40–49 years had higher risk of becoming non-respondents than those aged 50 years or older. Goldberg et al. (2006) did not find any association between age and non-response. Since no clear picture emerged on the association between age and non-response among women, age remains an important factor to investigate. Being unmarried was associated with increased odds of non-response only among women born in 1970/75, which is consistent with a study by Young, Powers, and Bell (2006).

We lacked sufficient data on reasons for non-response at follow-up and were unable to investigate in detail whether the influence of the characteristics differed with respect to types of non-response. Due to lack of relevant characteristics, we were also unable to evaluate non-response among participants in the short interviews and those who did not participate in the baseline interview. Because of the small number of non-responders among participants in the full interviews, we had to limit our study to only those determinants that satisfied set criteria in order to maximize statistical power. The external validity of our study may be limited due to the restriction to only women that participated in the full interviews. We used data that were collected in 1990, 1995, 2000, and 2013. The main reason why the data collection took so long relates to approval of grants that in turn is connected to the possibilities for principal investigators to prepare grants. All parts of this study have been financed through external grants. All women were living in Gothenburg at the time of study recruitment. However, when reaching out to the women for baseline and follow-up interviews, some women had moved within or outside Sweden, thus making future contacts with the women difficult and time-consuming. During the time for this study, the number of cell phone users and unlisted telephone numbers had increased. This may have contributed to further difficulties reaching out to the women (Galea and Tracy 2007). These factors may explain why we recorded slightly higher non-response (>30%) at later waves compared to other survey-based studies (Caetano, Ramisetty-Mikler, and McGrath 2003; Clemens et al. 2007). It remains unclear what proportion of non-response or missing is deemed acceptable. Bennett (2001) stated that an analysis with more than 10% missing data is likely to introduce bias. Groenwold and Dekkers (2020) compared data with different proportions of missing (0%, 2%, 17% and 50%) and found a higher risk of bias in dataset with 2% missing compared to that with 50% missing, for example. Theysuggested that the proportion of missing data should not be used as a measure for assessing bias, as it does not provide sufficient information regarding the risk of bias associated with a study.

In this study, we evaluated non-response at two time points within a space of 10 years, which increased the chances of capturing information about people that may have dropped out at baseline but re-entered the study at later waves. Rather than using only the baseline measurements to assess non-response at the end of the follow-up, we used characteristics measured before the planned interview, thus reducing possible bias that may be associated with time-varying characteristics. Although 5 years have passed since the last data collection, we think that it is worthwhile to publish data from this study because the outcome (non-response) probably does not change in a five-year period.

In a study on alcohol use and its related consequences, the underrepresentation of women with problematic alcohol use, for example, could lead to inaccurate estimation of alcohol prevalence and incidence, figures that are important for planning of preventive efforts and for health and social care. Knowledge about the determinants of non-response can aid researchers’ decision-making in terms of methods for dealing with non-response in statistical analysis as well as inform on potential target groups to consider when planning programs to improve participation in future studies. Our study findings demonstrate that older women with problematic alcohol use, who have poor health, and basic education; and younger women who are unmarried may be potential target groups.

Acknowledgments

The authors thank Fredrik Spak for data collection.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author, [C.A.N], upon reasonable request.

Supplementary material

Supplemental data for this article can be accessed on the publisher’s website.

Additional information

Funding

The work was supported by the [Swedish Council for Social Research, Stockholm] under Grant [94-0130:1C] and [the Swedish Research Council for Health, Working Life and Welfare (Forte)] under Grant [2013-0632].