Interpreting analytic methods and reported findings

This lesson focuses on how to read the main analytical approaches used in Beard et al. 2017 and Fa-Binefa et al., and how those methods shape the conclusions each article can reasonably support.

1. Why the analytic method matters

Research findings are not just about the numbers reported at the end. The statistical method determines:

• what kind of relationship is being estimated,
• what the unit of analysis is,
• whether time is handled explicitly,
• whether the result is about association or risk over follow-up, and
• how cautiously we should interpret possible causation.

These two studies answer different kinds of questions. Beard et al. 2017 examines whether monthly population trends in smoking and alcohol-related behaviours move together over time in England. Fa-Binefa et al. examines whether age at smoking onset predicts later cardiovascular disease (CVD) events or mortality in a cohort followed over time.

2. Beard et al. 2017: ARIMAX time-series modelling

What question was the method used to answer?

In Beard et al. 2017, the authors used monthly survey data from April 2014 to June 2016 to assess whether changes in smoking-related prevalence were associated with changes in alcohol-related prevalence at the population level. The outcomes were monthly prevalence measures, such as smoking prevalence and high-risk drinking prevalence.

What is ARIMAX in this study?

The article states that all data were analysed using “Autoregressive Integrated Moving Average with Exogeneous Input (ARIMAX) modelling.” In practical terms, this is a time-series approach designed for repeated observations across time. It is useful when monthly values are not independent of one another and when the analyst wants to test whether one time-varying series is associated with another.

Key features of the method in this context are:

• Time-series structure: the data are monthly observations, not one record per individual outcome over follow-up.
• Autoregressive and moving-average components: these help account for the fact that a month’s value may be related to earlier months.
• Exogenous input: one series, such as smoking prevalence, is used to assess its association with another series, such as high-risk drinking prevalence.

What kind of evidence does this produce?

ARIMAX estimates whether fluctuations in one population trend are associated with fluctuations in another population trend after accounting for time-series dependence. So the result is about co-movement over time at the population level.

That means the method supports conclusions such as:

• when smoking prevalence decreases or increases across months, high-risk drinking prevalence tends to decrease or increase as well,
• the association is observed in aggregated monthly data, and
• some smoking-related and alcohol-related indicators may be linked at the population trend level.

According to Beard et al. 2017, the authors found a positive association between change in smoking prevalence and prevalence of high-risk drinking. They also reported no statistically significant associations for motivation to stop smoking versus motivation to reduce alcohol consumption, or for attempts to stop smoking versus attempts to reduce alcohol consumption.

How should the reported effect be interpreted?

Beard et al. 2017 gives an example showing that a 1 percentage point reduction in smoking prevalence corresponds to a very small absolute reduction in high-risk drinking prevalence. This helps readers distinguish between:

• relative movement in the series, and
• absolute percentage-point change.

So, even when the association is statistically significant, the practical size of the linked change may be modest.

What conclusions can and cannot be supported?

Supported conclusions:

• There was a statistically detectable population-level association between monthly smoking prevalence and monthly high-risk drinking prevalence in England during the study period.
• The same was not found for the motivation and attempt measures examined.
• The findings are consistent with the idea that smoking and high-risk drinking may behave as complementary behaviours at the population level.

Not fully supported:

• That individual smokers who cut down or quit also reduced drinking.
• That smoking changes caused alcohol changes.
• That a specific policy intervention definitely changed both behaviours, because this analysis was not an interrupted time-series evaluation of a defined intervention.

This caution is reinforced in Beard et al. 2017 itself, where the discussion notes that if the findings are confirmed and shown to reflect a causal association, policy models would need to take that into account. That wording shows the study is suggestive, not conclusive, about causality.

3. Fa-Binefa et al.: Cox regression with spline and bootstrap-based cut-point assessment

What question was the method used to answer?

Fa-Binefa et al. asked whether smoking onset age (SOA) was associated with later risk of cardiovascular disease events, cardiovascular mortality, and all-cause mortality. This is a prognosis or risk-over-time question, so the method needed to incorporate follow-up time and the timing of events.

What is Cox regression in this study?

The main modelling framework was Cox proportional hazards regression. This method estimates hazard ratios (HRs), which compare the rate of an event over follow-up associated with a predictor.

In this article, Cox regression allows the authors to estimate how SOA relates to later hazard of outcomes while adjusting for other variables. The table notes adjustment for factors such as age and, depending on the outcome, cholesterol, glucose, smoking status, pack-years, recruitment survey, and years since quitting smoking.

Why did they use splines first?

Fa-Binefa et al. did not assume automatically that SOA had a simple straight-line relationship with outcomes. Instead, they used a penalized smoothing spline (“pspline”) to assess both linear and non-linear components of the dose-response association.

This step matters because it asks:

• Is each additional year of delay in smoking onset linked to a similar change in risk across the whole age range?
• Or does the relationship bend, flatten, or change at certain ages?

If the non-linear component was not significant, the authors treated SOA as a continuous variable in the Cox model. If non-linearity was present, they looked for a cut-point where the slope changed.

What was the purpose of bootstrap-based cut-point assessment?

When the spline suggested a non-linear association, Fa-Binefa et al. used bootstrapping to identify the best cut-point where the linear dose-response relationship changed. They performed 1000 iterations and used the median observed cut-point as the best cut-point.

Conceptually, this means:

• the data suggested the effect of SOA might differ before and after a certain age,
• the bootstrap procedure was used to make cut-point selection more stable than choosing a threshold from a single sample result, and
• after selecting the cut-point, the analysis was stratified and separate linear associations were estimated within each age range.

What did this method find?

For cardiovascular mortality, the bootstrap analysis identified a cut-point at 12 years. The sample was then split into SOA ≤12 years and >12 years.

Fa-Binefa et al. reported that in the younger group, each year of delay in smoking onset was associated with a 42% decrease in CVD mortality risk (HR 0.58, 95% CI 0.37–0.90). In the older group, SOA was not associated with CVD risk (HR 1.00, 95% CI 0.93–1.08).

For all-cause mortality, the association was described as linear, and each year of delay in smoking onset was associated with lower mortality risk (HR 0.96, 95% CI 0.93–0.98).

How should the hazard ratios be interpreted?

In this framework:

• HR below 1 means later smoking onset is associated with lower hazard of the outcome.
• HR equal to 1 means no association is detected.
• Confidence intervals indicate the precision of the estimate and whether the result is compatible with no association.

So the study suggests that very early smoking onset is particularly important for cardiovascular mortality, because the association was strongest in the ≤12 years group.

What conclusions can and cannot be supported?

Supported conclusions:

• Smoking onset age was associated with later mortality outcomes in this cohort.
• The association with all-cause mortality appeared approximately linear.
• The association with cardiovascular mortality appeared non-linear, with stronger risk implications for onset at very young ages.
• Delaying smoking onset, especially beyond very early adolescence, was associated with lower later cardiovascular mortality risk.

Not fully supported:

• That the identified cut-point of 12 years is a universal biological threshold for all populations.
• That delaying onset alone, independent of all other life-course factors, definitively causes the reduction in later risk.
• That the findings apply beyond the study population without caution.

Even though Cox regression is stronger than a simple cross-sectional association for studying later outcomes, this is still observational research. Adjustment helps, but residual confounding and measurement limits remain possible.

4. Comparing what the two methods let the authors claim

Different unit of analysis

• Beard et al. 2017: monthly population prevalence values.
• Fa-Binefa et al.: individual participants followed for later events or death.

Different role of time

• Beard et al. 2017: time is the sequence of monthly observations; the method models trends and dependence across months.
• Fa-Binefa et al.: time is follow-up until an event or censoring; the method models hazard over time.

Different outcome language

• Beard et al. 2017: association between changes in prevalence series.
• Fa-Binefa et al.: hazard ratios for future disease or mortality outcomes.

Different substantive conclusions

• Beard et al. 2017 supports statements about population-level co-variation of behaviours across months.
• Fa-Binefa et al. supports statements about individual-level risk over follow-up associated with smoking onset age.

Different handling of non-linearity

• Beard et al. 2017 focuses on associations between time-series after accounting for temporal structure.
• Fa-Binefa et al. explicitly tests whether the exposure-outcome relationship is linear or changes shape, using splines and then bootstrap-based cut-point assessment.

5. A simple interpretation guide for students

When reading Beard et al. 2017, ask:

• Are the authors talking about individuals or monthly population averages?
• Does the method show behaviours moving together over time?
• Did they find association for all smoking and alcohol indicators, or only some?
• Do the findings imply causation, or only temporal association?

When reading Fa-Binefa et al., ask:

• Is the exposure being treated as continuous or does the relationship change across values?
• What does the spline say about linear versus non-linear association?
• What does the hazard ratio mean for each additional year of later smoking onset?
• Is the effect similar across the whole range, or concentrated among very early initiators?

6. Key takeaway

Beard et al. 2017 uses ARIMAX to test whether monthly population trends in smoking and alcohol-related behaviours are associated over time, supporting conclusions about linked population-level changes but not direct individual-level or causal claims. Fa-Binefa et al. uses Cox regression, with spline testing and bootstrap-based cut-point assessment, to study how smoking onset age relates to later hazards of CVD and mortality, supporting conclusions about differential long-term risk and possible non-linearity in that risk relationship.

In short: ARIMAX helps answer Do these population trends move together over time? Cox regression with splines and bootstrapped cut-points helps answer How does an exposure relate to future risk, and is that relationship linear or threshold-like?