Skip to main content
Download PDF

Developing non-response weights to account for attrition-related bias in a longitudinal pregnancy cohort

BMC Medical Research Methodology volume 23, Article number: 295 (2023) Cite this article

Abstract

Background

Prospective cohorts may be vulnerable to bias due to attrition. Inverse probability weights have been proposed as a method to help mitigate this bias. The current study used the “All Our Families” longitudinal pregnancy cohort of 3351 maternal-infant pairs and aimed to develop inverse probability weights using logistic regression models to predict study continuation versus drop-out from baseline to the three-year data collection wave.

Methods

Two methods of variable selection took place. One method was a knowledge-based a priori variable selection approach, while the second used Least Absolute Shrinkage and Selection Operator (LASSO). The ability of each model to predict continuing participation through discrimination and calibration for both approaches were evaluated by examining area under the receiver operating curve (AUROC) and calibration plots, respectively. Stabilized inverse probability weights were generated using predicted probabilities. Weight performance was assessed using standardized differences of baseline characteristics for those who continue in study and those that do not, with and without weights (unadjusted estimates).

Results

The a priori and LASSO variable selection method prediction models had good and fair discrimination with AUROC of 0.69 (95% Confidence Interval [CI]: 0.67–0.71) and 0.73 (95% CI: 0.71–0.75), respectively. Calibration plots and non-significant Hosmer-Lemeshow Goodness of Fit Tests indicated that both the a priori (p = 0.329) and LASSO model (p = 0.242) were well-calibrated. Unweighted results indicated large (> 10%) standardized differences in 15 demographic variables (range: 11 − 29%), when comparing those who continued in the study with those that did not. Weights derived from the a priori and LASSO models reduced standardized differences relative to unadjusted estimates, with the largest differences of 13% and 5%, respectively. Additionally, when applying the same LASSO variable selection method to develop weights in future data collection waves, standardized differences remained below 10% for each demographic variable.

Conclusion

The LASSO variable selection approach produced robust weights that addressed non-response bias more than the knowledge-driven approach. These weights can be applied to analyses across multiple longitudinal waves of data collection to reduce bias.

Peer Review reports

Introduction

Longitudinal study designs allow researchers to establish temporality of exposure-outcome relationships by following samples of individuals over time with repeated measurements [1]. It is, however, common for participants in longitudinal cohorts to be lost to follow-up (i.e., attrition or censoring) [2]. While attrition over time is expected, it can contribute to biased exposure-outcome relationships depending on the nature of how and/or why individuals drop out of the study [3]. Attrition leaves researchers with challenges on how to address missing data, which will depend on why the data are missing, and has implications for analysis.

Several methods exist that aim to mitigate potential bias related to attrition. Complete case analysis and multiple imputation are used commonly, but both rely on assumptions related to how data are missing [4]. Attrition contributes to missing data that may be missing at random, missing not at random, or missing completely at random. Each changes the assumptions of how data are missing and the potential bias that may occur if one were to apply a complete case analysis [1, 4]. Missing at random values are conditional on observed data, missing not at random depends on unobserved data, and missing completely at random depends on neither [1, 4]. Another way to address attrition/censoring is to weight existing respondents using inverse probability of participation weights that are calculated based on the baseline information [5,6,7]. That is, the inverse of the probability of continuing in the study at subsequent waves of collection (i.e., those who have lower probability of continuing receive higher weights). This method accepts that individuals may drop out of longitudinal studies for various reasons and that these reasons can be modelled through weighting and using the existing data.

This study aimed to describe the process for developing and assessing the performance of weights in a pregnancy cohort that has spanned approximately 14 years to provide a statistical approach to account for attrition and the potential for selection bias. We describe two methods for developing a model to create the weights (one knowledge-based a priori model that is investigator derived and another data driven model using Least Absolute Shrinkage and Selection Operator [LASSO] regression), assessed the discrimination and calibration performance of each model, and then assessed the performance of the weights from each model. The best performing model was then applied to subsequent waves of data collection, and the performance of these weights was assessed to consider using these weights across all data collection waves in this cohort.

Methods

Cohort description

This study used the data from All Our Families Cohort (formerly All Our Babies Cohort) [8]. This is a pregnancy cohort that recruited 3387 women at less than 25 weeks gestational age in Calgary, Canada. Initial recruitment took place from May 2008 and December 2010 [8]. Women completed one survey at < 25 weeks gestational age, one at 34–36 weeks gestational age, and one at four months postpartum [8]. Four more surveys were conducted when their child reached one year (2009–2012), three years (2012–2014), five years (2014–2016), and eight years of age (2017–2019). Finally, a survey was conducted during the COVID-19 pandemic between May 20 and July 15, 2020 [9, 10]. For the 8-year and COVID-19 surveys, data were collected and managed using Research Electronic Data Capture (REDCap) electronic data capture tools hosted at University of Calgary [11, 12], prior to REDCap, data were collected using physical surveys and TeleForm to scan and verify data. This study used the first survey at < 25 weeks gestational age as the baseline cohort and the 3-year follow-up to assess non-participation. Three-year follow-up is chosen as there was little loss to follow-up in this cohort during gestation and at birth; at the 1-year follow-up, there were administrative challenges that affected response rate, but for reasons not related to general attrition. STATA 16.0 statistical software (StataCorp, College Station, TX, USA) was used for all analyses while the ggplot2 package in R software [13] was used to generate figures.

Model development

We examined two models: a priori and LASSO variable selection method, described below. For both models, the first survey was used to identify variables for inclusion in prediction models that ultimately led to weight development. The first survey included 127 variables across multiple topics, including socio-demographics, prenatal physical and mental health, lifestyle, and pre-pregnancy and life events. For both models, we used the follow-up survey conducted at three years for the outcome point (i.e., women who did not attend the 3-year survey were considered lost to follow-up). We later applied these models to subsequent waves of data collection (5-year follow-up, 8-year follow-up, and once during the COVID-19 pandemic).

The first method for weight generation followed a knowledge-based variable selection approach. Investigators with subject matter expertise in pregnancy cohorts (KA, SM, SP, TMP) met several times and collectively identified possible variables for inclusion, including possible interaction terms that could be related to drop-out over time. Decisions on which variables to include were based on existing content expertise as well as the quality of variable data (i.e., high proportion of missing data).

The second method followed a LASSO variable selection method described by Schmidt et al. [14]. This method was used to develop weights in a child cohort and uses least absolute shrinkage and selection operator (LASSO) regression to select relevant variables [15]. Categorical variables were left in categories as they were initially coded with the addition of a category for missing in some cases. For those categorical variables with missing data, missingness was recoded so that ‘missing’ became a category of the variable itself. If a single level within a categorical variable had a large coefficient based on the initial LASSO regression, we retained the overall categorical variable as a candidate for the next step of variable selection. This meant that continuous variables were cut into relevant categories and another level of “missing” was created. Next, we split the variables of interest into seven relevant context themes: Sociodemographic Characteristics, Pregnancy History, Conception History, Prenatal Care, Lifestyle and Health Care Use During Pregnancy, Mental Health/Social Support, and Smoking/Drug/Alcohol Consumption (current and previous). We applied LASSO regression with 10-fold cross validation to each context theme such that the tuning parameter minimizes the out-of-sample prediction error [14, 16]. The three variables with the largest coefficients from the LASSO regression were fit in a multivariable logistic regression model and area under the receiver operating curve (ROC), sometimes referred to as C-statistic, was calculated from predicted probabilities. One at a time, the variable with the next largest coefficient was fit to the model and this process was repeated until the ROC was not significantly different (p > 0.05) from the previous ROC. This was completed for each context theme and all variables from each context them served as the candidate variables for the final model. Next, all of those top contributing variables, based on coefficient size, from each context theme were combined into a larger LASSO model. Only non-zero coefficients were selected for inclusion in the final logistic regression model.

Model assessment

We assessed the ability of the model to predict continuing participation through discrimination and calibration for both approaches. We assessed discrimination using area under the ROC. ROC plots are among the most common method of assessing discrimination and represents a curve of sensitivity over 1-specifictiy where sensitivity represents true positives (cases) while specificity represents true negatives (not cases) [17]. Values for ROC range from 0.5 (no better assessment than chance) to 1.0 (perfect discrimination). The following cut-offs are often suggested as guidelines to assess discrimination: ≤0.5 is no better than chance, > 0.5 and < 0.7 is poor, > 0.7 ≤ 0.8 is acceptable and > 0.8 is excellent [18]. Calibration of the model relates to the accuracy of predicted risk and has been defined as “for patients with a predicted risk of R%, on average R out of 100 should indeed suffer from the disease or event of interest” [19]. We assessed calibration through a combination of Hosmer-Lemeshow goodness-of-fit test, mean calibration, and calibration plots [20]. We then applied this model to the next wave of data collection (i.e., 5-year follow-up) to assess the temporal validity of the models and assess in the same way.

Weights assessment

Using the models derived from the a priori and LASSO variable selection method, we calculated predicted probabilities and stabilized inverse probability weights [21]. We applied stabilized weights as they typically result in less variance than non-stabilized weights [21, 22]. The means, standard deviations (SDs), and ranges of the weights were calculated and plotted. Weights were truncated at the 0.5th and 99.5th percentiles to avoid bias due to extreme weights [23]. Weight performance was measured by comparing baseline characteristics of those who continued in the study and those who did not, with and without the weights. It has been suggested that the standardized difference is the preferred measure for comparing weight balance between groups (continued in study vs. lost to follow-up) and that a difference between groups of less than 10% is negligible [24]. We use the “pbalchk” package in STATA to calculate the standardized difference. Standardized differences can be calculated for continuous and categorical variables and involve both means for the continued and lost-to-follow up groups and their variances; for more information on this calculation see Austin, 2009 [25].

Finally, the same model identified at the 3-year follow-up was then re-fit to develop weights for each of the subsequent waves (i.e., 5-year follow-up, 8-year follow-up, and the first survey during the COVID-19 pandemic). The performance of these weights was assessed as above.

Results

Based on the 3,351 singleton births in the All Our Families cohort, 1,990 (59.4%) continued participation at the three-year follow-up while 1,361 (40.6%) did not. Of note, the study population in follow-up waves differed slightly from baseline due to various reasons, such as child age eligibility for standardized developmental scales when data collection was initiated and associated funding and ethical constraints [9]. At the three-year follow-up 69% of participants from the two-year follow-up responded to the survey [9]. However, since some participants had dropped out at earlier waves (during pregnancy and at-birth waves), this represented 59% of the participants initially enrolled in the study. Ultimately, the a priori model contained 18 variables while the LASSO variable selection method model contained 22 (Table 1). The two models shared four variables (Education, Ethnicity, Physical Component Summary, and Previous History of Adverse Birth Outcomes).

Table 1 Variables included in participation models
Full size table

The a priori model had poor-acceptable discrimination ROC of 0.69 (95% CI: 0.67–0.71) while the LASSO variable selection method model had acceptable discrimination ROC of 0.73 (95% CI: 0.71–0.75). Hosmer-Lemeshow goodness-of-fit tests with 10 bins were non-significant for both the knowledge-based (p = 0.329) and the LASSO variable selection method approach (p = 0.242). A statistically non-significant goodness-of-fit test indicates no statistical difference in the observed cases from the predicted cases [29]. A non-significant goodness-of-fit result implies a well-calibrated model; however, a goodness-of-fit test alone may not be sufficient to assess calibration [19]. To this end, we considered the mean calibration where “the average predicted risk is compared to the overall event rate” [19]. In this case, the ‘event rate’ is considered as the proportion of individuals who continue in the study at the 3-year follow-up and is compared with the calculated proportion derived from the a priori and LASSO variable selection method models. Mean calibrations were 0.594 (95% CI: 0.58–0.60) and 0.594 (95% CI: 0.59–0.60) for the a priori and LASSO variable selection method models, respectively, compared with an observed proportion of continued participation of 0.594. Given that the mean calibrations in the two models were very similar to that of the observed proportion, both models appeared well-calibrated. In addition, we visually examined the calibration plots for each model (Figs. 1 and 2). The a priori and LASSO variable selection method models were re-fit on the next wave of data collection (5-year follow-up) and performed similarly to the previous wave with ROC of 0.69 (95% CI: 0.67–0.71) and 0.73 (0.72–0.75), non-significant goodness-of-fit tests of p = 0.567 and p = 0.307, and mean calibration of 0.596 (95% CI: 0.59–0.60) and 0.593 (95% CI: 0.59–0.60), respectively.

Fig. 1
figure 1

Results of Calibration Curve for LASSO variable selection method Model; AUC: Area Under the Receiver Operating Curve

Full size image
Fig. 2
figure 2

Results of Calibration Curve for a priori Model; AUC: Area Under the Receiver Operating Curve

Full size image

In calibration plots, an ideal plot (a diagonal line with slope of 1 and intercept of 0) is presented with a calibration curve derived from the model data and demonstrates how similar (or not) the estimated risk is to observed risk. The plot is assessed by examining the curve slope (target of 1.0) and by using a loess function to compare curve of predicted risk with the ideal plot [18]. Both Hosmer-Lemeshow tests and mean calibration suggested moderate calibration, as did the calibration plots; although, the LASSO variable selection method model seemed more well-calibrated at higher values than the a priori model.

The stabilized weights for the a priori model had a mean (SD) of 1.00 (0.58) and a range of 0.43–10.1. The LASSO variable selection method model had a mean (SD) of 1.00 (0.74) and a range of 0.42–23.1. After trimming, the LASSO variable selection method and a priori models had maximum weights of 4.8 and 4.9, respectively. This resulted in changes to 33 individual’s weights in both models. As well, mean (SD) for the a priori and LASSO variable selection method models were 0.99 (0.46) and 0.99 (0.51), respectively.

Weights performance

The absolute standardized differences were calculated across baseline demographic variables (chosen a priori) in the unweighted group were as large as 28.9% for home ownership and 27.5% for income (binary outcome split at $60,000) and a mean of standardized differences of 17.5%. In the a priori model, the largest absolute standardized difference was 13.1% (smoking history) with two variables having a standardized difference of 10% or greater and a mean of standardized differences of 4.6% (Fig. 3). In the LASSO variable selection method derived weights, the largest absolute standardized difference was just 5.4% (anxiety symptoms) with no variables greater than 10% and a mean of standardized differences of 2.5%. Comparisons of baseline characteristics are based on complete data at baseline; of the 15 variables measured, eight were missing data in ≤ 1% while the other seven (Income, Anxiety, Symptoms, Depression Symptoms, Maternal Age, New Canadian, and Household Size) ranged from 1 to 4.4%.

Fig. 3
figure 3

Comparing the unweighted absolute standardized differences with the stabilized truncated weights of a priori and LASSO variable selection method models

Full size image

Since the LASSO variable selection method weights appeared to perform better, weights were developed using this approach and applied to subsequent waves of data collection with performance evaluated in the same way (Fig. 4). Across each follow-up wave of data collection (3-year, 5-year, 8-year, and COVID-19 survey [approximately 12-years of follow-up]), absolute standardized differences remained below 10% for baseline demographic variables.

Fig. 4
figure 4

Comparing absolute standardized differences with the stabilized truncated weights derived from LASSO variable selection method model across data collection follow-up waves (3-, 5-, 8-, year follow-up and follow-up during COVID-19) in longitudinal cohort

Full size image

Discussion

This study aimed to develop non-response weights for a pregnancy cohort that has followed participants for more than 12 years. To accomplish this, we examined two approaches: one a priori and another LASSO variable selection method. The LASSO variable selection method approach produced robust weights that addressed non-response bias more than the a priori approach. The data driven approach, however still required content knowledge in how data were grouped, combined, or split. These weights can be applied to analyses across multiple waves of data collection to reduce bias. While the a priori model performed well, the weights themselves did not reduce differences in baseline characteristics to the same degree as the LASSO variable selection method model. While the models contained different specific variables, there was some overlap in that variables between the models captured similar concepts. For example, the a priori model used the combined variable of ‘history of drug/alcohol dependence’ while the LASSO variable selection method model included drug use per week and number of alcoholic drinks per day. While both the a priori and LASSO variable selection method models had access to the same calculated variables, their component parts, and interaction terms, the a priori model attempted to create simplicity and reduce the number of variables within an overarching theme, the nuance of more specific variables was ultimately found to be more informative. As well, the a priori variable selection ended after initial selection of variables. Typically, in developing prediction models, the investigators would examine performance and re-calibrate as necessary, but for the purposes of variable selection performance this was not done. Further, the weights derived from the LASSO variable selection method approach were robust across waves. That is, the balance achieved at the 3-year follow-up was generally maintained through the 5-year and 8-year follow-up as well as through the survey during the COVID-19 pandemic (12 years after baseline). This indicates that the same factors influence retention over time, and that one model can be used to develop weights, and then applied consistently over several waves of data collection.

Unweighted differences in baseline characteristics existed with respect to attrition status during follow-up, suggesting the potential for selection bias. However, while bias due to attrition is possible in cohort studies, and should be considered, bias is not guaranteed simply due to differing baseline characteristics of those who continue those who drop-out if those differences do not exist between groups as they relate to the exposure-outcome relationship of interest. Previous work by this group has used weighted and unweighted results showed a slight difference in magnitude of results but no difference in trends [10]. Further, recent work has demonstrated no difference in results comparing modelled results of complete case analysis and inverse probability weighting using missing at random, missing not at random, and missing completely at random data [30]. To better understand the extent of bias due to attrition, comparison of analyses with and without weights is suggested. The weights created for this sample balanced demographic characteristics of those who continued participation and those who did not and serve as another way to quantitatively examine the potential role of attrition in creating bias in our longitudinal study cohort.

There exist some methodologic challenges in creating effective weights while also ensuring no undue influence of extreme weights. There is no clearly defined point at which to truncate weights but it is important to consider both heterogeneity in order to achieve balance and the role of extreme weights. The use of a very small amount of truncation seemed to be effective for this particular sample. By truncating just at the 99.5th percentile, we see the range in the LASSO variable selection method weights drop from 23.1 to 4.8 which would indicate just a few “outliers” that could have spuriously influenced the weights.

Strengths of this study include examining two approaches to developing weights and the comparison of the two. As well, this study used a large sample of over 3,351 participants with 127 individual variables that were considered. The breadth of variables allowed us to consider a multitude of factors that could predict continuation in the cohort in later waves.

This study is not without limitations. The LASSO variable selection method approach used missing data as a level within categories. This allowed us to maintain a large sample size, but it also meant that variables that would normally be continuous were categorized to create this missing level. Categorization of continuous variables can result in loss of information given the collapsing of participant data into groups.

This study outlined two approaches to developing non-response weights to address bias that may be introduced due to attrition, with a LASSO variable selection method approach creating weights performing better that a priori approach in balancing baseline characteristics. The All Our Families cohort observes approximately 60% of participants returning to the study eight years after giving birth, in line with other major pregnancy cohorts [31,32,33]. The use of inverse probability weights considers the potential effect of non-response bias and the weights developed here can be applied to future studies using the AOF cohort data in secondary analyses and subsequent data collections; a further advantage of the use of these weights is that they can be easily applied to a variety of outcome models (i.e., linear regression, logistic regression, survival analysis). Importantly, the approach used in the present study in creating these weights could be applied in other cohorts, where the potential for selection bias exists due to attrition. Balancing the characteristics of participants at later cohort data collection waves to the sample recruited at baseline increases the confidence that temporal associations better reflect the experience of the target population.

Data Availability

The datasets analysed during the current study are not publicly available as they contain personal participant information but are available from the corresponding author, through the All Our Families Cohort Study, on reasonable request.

References

  1. Dohoo IR, Martin SW, Stryhn H. Methods in epidemiologic research. 2012.

  2. Kristman V, Manno M, Côté P. Loss to follow-up in cohort studies: how much is too much? Eur J Epidemiol. 2004;19:751–60.

    Article  PubMed  Google Scholar 

  3. Dettori JR. Loss to follow-up. Evidence-based spine-care Journal. 2011;2(01):7–10.

    Article  PubMed  PubMed Central  Google Scholar 

  4. Little RJ, Rubin DB. Statistical analysis with missing data. John Wiley & Sons; 2019. Apr 23.

  5. Seaman SR, White IR. Review of inverse probability weighting for dealing with missing data. Stat Methods Med Res. 2013;22(3):278–95.

    Article  PubMed  Google Scholar 

  6. Biele G, Gustavson K, Czajkowski NO, Nilsen RM, Reichborn-Kjennerud T, Magnus PM, Stoltenberg C, Aase H. Bias from self selection and loss to follow-up in prospective cohort studies. Eur J Epidemiol. 2019;34(10):927–38.

    Article  CAS  PubMed  Google Scholar 

  7. Howe CJ, Cole SR, Lau B, Napravnik S, Eron JJ Jr. Selection Bias due to loss to follow up in Cohort studies. Epidemiology. 2016;27(1):91–7.

    Article  PubMed  PubMed Central  Google Scholar 

  8. McDonald SW, Lyon AW, Benzies KM, McNeil DA, Lye SJ, Dolan SM, Pennell CE, Bocking AD, Tough SC. The all our babies pregnancy cohort: design, methods, and participant characteristics. BMC Pregnancy Childbirth. 2013;13(1):1–2.

    Google Scholar 

  9. Tough SC, McDonald SW, Collisson BA, Graham SA, Kehler H, Kingston D, Benzies K. Cohort profile: the all our babies pregnancy cohort (AOB). Int J Epidemiol. 2017;46(5):1389–90k.

    Article  PubMed  Google Scholar 

  10. Racine N, Hetherington E, McArthur BA, McDonald S, Edwards S, Tough S, Madigan S. Maternal depressive and anxiety symptoms before and during the COVID-19 pandemic in Canada: a longitudinal analysis. The Lancet Psychiatry. 2021;8(5):405–15.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Harris PA, Taylor R, Thielke R, Payne J, Gonzalez N, Conde JG. Research electronic data capture (REDCap)—a metadata-driven methodology and workflow process for providing translational research informatics support. J Biomed Inform. 2009;42(2):377–81.

    Article  PubMed  Google Scholar 

  12. Harris PA, Taylor R, Minor BL, Elliott V, Fernandez M, O’Neal L, McLeod L, Delacqua G, Delacqua F, Kirby J, Duda SN. The REDCap consortium: building an international community of software platform partners. J Biomed Inform. 2019;95:103208.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Wickham H. ggplot2: elegant graphics for data analysis. Volume ISBN 978–3–319–24277–4. Springer-Verlag New York; 2016. https://ggplot2.tidyverse.org.

  14. Schmidt SC, Woll A. Longitudinal drop-out and weighting against its bias. BMC Med Res Methodol. 2017;17(1):1–1.

    Article  Google Scholar 

  15. Tibshirani R. Regression shrinkage and selection via the lasso. J Roy Stat Soc: Ser B (Methodol). 1996;58(1):267–88.

    Google Scholar 

  16. Lee TF, Chao PJ, Ting HM, Chang L, Huang YJ, Wu JM, Wang HY, Horng MF, Chang CM, Lan JH, Huang YY. Using multivariate regression model with least absolute shrinkage and selection operator (LASSO) to predict the incidence of xerostomia after intensity-modulated radiotherapy for Head and Neck cancer. PLoS ONE. 2014;9(2):e89700.

    Article  PubMed  PubMed Central  Google Scholar 

  17. Cook NR. Use and misuse of the receiver operating characteristic curve in risk prediction. Circulation. 2007;115(7):928–35.

    Article  PubMed  Google Scholar 

  18. Hosmer DW Jr, Lemeshow S, Sturdivant RX. Applied logistic regression. John Wiley & Sons; 2013 Apr. p. 1.

  19. Van Calster B, Nieboer D, Vergouwe Y, De Cock B, Pencina MJ, Steyerberg EW. A calibration hierarchy for risk models was defined: from utopia to empirical data. J Clin Epidemiol. 2016;74:167–76.

    Article  PubMed  Google Scholar 

  20. Van Calster B, McLernon DJ, Van Smeden M, Wynants L, Steyerberg EW. Calibration: the Achilles heel of predictive analytics. BMC Med. 2019;17(1):1–7.

    Google Scholar 

  21. Robins JM, Hernan MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology. 2000;11:550–60.

    Article  CAS  PubMed  Google Scholar 

  22. Chesnaye NC, Stel VS, Tripepi G, Dekker FW, Fu EL, Zoccali C, Jager KJ. An introduction to inverse probability of treatment weighting in observational research. Clin Kidney J. 2022;15(1):14–20.

    Article  PubMed  Google Scholar 

  23. Cole SR, Hernan MA. Constructing inverse probability weights for marginal structural models. Am J Epidemiol. 2008;168:656–64.

    Article  PubMed  PubMed Central  Google Scholar 

  24. Austin PC, Stuart EA. Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies. Stat Med. 2015;34(28):3661–79.

    Article  PubMed  PubMed Central  Google Scholar 

  25. Austin PC. Using the standardized difference to compare the prevalence of a binary variable between two groups in observational research. Commun statistics-simulation Comput. 2009;38(6):1228–34.

    Article  Google Scholar 

  26. Ware JE Jr, Kosinski M, Keller SD. A 12-Item short-form Health Survey: construction of scales and preliminary tests of reliability and validity. Med Care. 1996 Mar;1:220–33.

  27. Sherbourne CD, Stewart AL. The MOS social support survey. Soc Sci Med. 1991;32(6):705–14.

    Article  CAS  PubMed  Google Scholar 

  28. Cohen S, Kamarck T, Mermelstein R. A global measure of perceived stress. J Health Social Behav 1983 Dec 1:385–96.

  29. Kramer AA, Zimmerman JE. Assessing the calibration of mortality benchmarks in critical care: the Hosmer-Lemeshow test revisited. Crit Care Med. 2007;35(9):2052–6.

    Article  PubMed  Google Scholar 

  30. Metten MA, Costet N, Multigner L, Viel JF, Chauvet G. Inverse probability weighting to handle attrition in cohort studies: some guidance and a call for caution. BMC Med Res Methodol. 2022;22(1):1–5.

    Article  Google Scholar 

  31. Piler P, Kandrnal V, Kukla L, Andrýsková L, Švancara J, Jarkovský J, Dušek L, Pikhart H, Bobák M, Klánová J. Cohort profile: the European longitudinal study of pregnancy and childhood (ELSPAC) in the Czech Republic. Int J Epidemiol. 2017;46(5):1379–f.

    PubMed  Google Scholar 

  32. Fraser A, Macdonald-Wallis C, Tilling K, Boyd A, Golding J, Davey Smith G, Henderson J, Macleod J, Molloy L, Ness A, Ring S. Cohort profile: the Avon Longitudinal Study of parents and children: ALSPAC mothers cohort. Int J Epidemiol. 2013;42(1):97–110.

    Article  PubMed  Google Scholar 

  33. Kooijman MN, Kruithof CJ, van Duijn CM, Duijts L, Franco OH, van IJzendoorn MH, de Jongste JC, Klaver CC, van der Lugt A, Mackenbach JP, Moll HA. The Generation R Study: design and cohort update 2017. Eur J Epidemiol. 2016;31:1243–64.

    Article  PubMed  Google Scholar 

Download references

Acknowledgements

The authors acknowledge the contribution and support of All Our Families participants and All Our Families team members.

Funding

All Our Families was funded through Alberta Innovates Interdisciplinary Team Grant #200700595 and the Alberta Children’s Hospital Foundation. TMP is supported by a Canadian Institutes of Health Research Doctoral Award (#187531). The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Author information

Authors and Affiliations

  1. Department of Paediatrics, University of Calgary, 28 Oki Drive NW, Calgary, T3B 6A8, Canada

    Tona M. Pitt, Suzanne C. Tough & Sheila McDonald

  2. Department of Epidemiology, Biostatistics and Occupational Health, McGill University, 2001 McGill College, Montreal, H3A 1G1, Canada

    Erin Hetherington

  3. Department of Community Health Sciences, University of Calgary, 3280 Hospital Drive NW, Calgary, T2N 4Z6, Canada

    Kamala Adhikari, Shainur Premji, Suzanne C. Tough & Sheila McDonald

  4. Provincial Population and Public Health, Alberta Health Services, 10301 Southport Rd SW, Calgary, T2W 1S7, Canada

    Tona M. Pitt, Kamala Adhikari & Sheila McDonald

  5. Centre for Health Economics, University of York, Heslington, YO10 5DD, York, UK

    Shainur Premji

  6. School of Psychology, Faculty of Social Sciences, University of Ottawa, 136 Jean- Jacques Lussier, Ottawa, K1N 6N5, Canada

    Nicole Racine

  7. Children’s Hospital of Eastern Ontario Research Institute, 401 Smyth Rd, Ottawa, K1H 5B2, Canada

    Nicole Racine

Authors
  1. Tona M. Pitt
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Erin Hetherington
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Kamala Adhikari
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Shainur Premji
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Nicole Racine
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Suzanne C. Tough
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Sheila McDonald
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

TMP drafted the main text of the manuscript and prepared all figures and tables. EH and NR developed the initial analysis plan and contributed to the background for this study. EH and TMP wrote the code for statistical analysis and TMP conducted the analysis. TMP, KA, SP, and SM contributed to the knowledge-based analysis and developed the final analysis plan. SCT and SM co-led the data procurement and ongoing data collection for the prospective cohort data used. All authors reviewed the manuscript. All authors reviewed the manuscript and contributed to the interpretation of findings.

Corresponding author

Correspondence to Sheila McDonald.

Ethics declarations

Ethics approval

Ethics approval for this study provided by the Conjoint Health Research Ethics Board at the University of Calgary (REB13-0868) and was completed in accordance within the guidelines and regulations of the Conjoint Health Research Ethics Board. This includes obtaining informed consent from all subjects included in the current study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests or conflicts of interest.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Pitt, T.M., Hetherington, E., Adhikari, K. et al. Developing non-response weights to account for attrition-related bias in a longitudinal pregnancy cohort. BMC Med Res Methodol 23, 295 (2023). https://doi.org/10.1186/s12874-023-02121-1

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12874-023-02121-1

Keywords

BMC Medical Research Methodology

ISSN: 1471-2288

Contact us