加入新推出的
Discord 社区,展开实时讨论,获得同行支持,并直接与 Meridian 团队互动!
术语库
使用集合让一切井井有条
根据您的偏好保存内容并对其进行分类。
基准:在所有处理变量均设置为基准值的反事实情景下的预期结果。对于付费媒体和自然媒体,基准值为零。对于非媒体处理变量,基准值可设置为观测到的变量最小值(默认)、最大值或用户提供的浮点值。
贡献率:每个处理变量的增量结果占总结果的百分比。用于报告时,例如 ModelFit
的预期结果与实际结果对比图和 MediaSummary
贡献瀑布图,总结果是总预期结果。用于贡献率先验时,总结果是总观测结果。
控制变量:模型中不属于处理变量的变量。控制变量用于估计基准结果,控制变量的因果效应或贡献百分比无法估计。有关如何选择控制变量的实用建议,请参阅控制变量。此外,请参阅相关概念“中介变量”。
单位增量 KPI 的费用 (CPIK):总支出除以总增量 KPI。如果 KPI 不是收入,并且“每个 KPI 的收入”数据不可用,则 CPIK 等于投资回报率的倒数。
效果:增量结果除以媒体单位总数。
预期结果:所有处理变量都设置为实际历史值时的预期结果。这是基准结果加上所有处理变量的增量结果的总和。
排期模式:给定媒体变量的媒体单位在各个地理区域和时间段的相对分布。当渠道的总预算增加或减少时,此模式用于在不同地理区域和时间段之间分配媒体单位,这适用于预算优化和响应曲线。
增量结果:每个处理变量导致的预期结果变化。对于付费媒体和自然媒体,这是指一个变量设置为零时,预期结果的变化。对于非媒体处理变量,这是指一个变量设置为每个地理区域和时间段的基准值(观测到的变量最小值 [默认]、最大值或用户提供的浮点值)时,预期结果的变化。如需了解详情,请参阅增量结果。
KPI:模型的响应变量(目标变量、因变量)。它可以是收入、销量、转化次数,也可以是处理变量可能产生因果效应的任何其他指标。
滞后效应:之前时间段的处理变量对之后时间段的结果产生影响。Meridian 使用 Adstock 函数将滞后效应纳入模型中。
边际投资回报率 (mROI):响应曲线的导数,大致相当于在当前支出水平之上额外支出一个货币单位(例如美元)所带来的投资回报率。
媒体执行:泛指特定渠道在不同地理区域和时间段的媒体单位价值。
中介变量:受处理变量的因果影响,并对 KPI 有因果效应的变量。以控制变量形式纳入这些变量会导致处理变量对 KPI 的因果估计出现偏差,因此不应将其纳入模型中。
结果:Meridian 衡量处理变量因果效应的主要指标。这通常是收入,但如果 KPI 不是收入,并且“每个 KPI 的收入”数据不可用,则 Meridian 会将结果定义为 KPI 本身。它不一定是模型的响应变量(请参阅 KPI 定义)。
响应曲线:给定媒体变量的增量结果与支出水平的关系图。随着支出的变化,媒体单位将根据排期模式分布在不同的地理区域和时间段。
投资回报率 (ROI):Meridian 将 ROI 定义为增量结果除以支出。如果 KPI 是收入,或者“每个 KPI 的收入”数据可用,则增量结果就是增量收入。否则,增量结果就是增量 KPI。
收入:对于非收入 KPI,此值为每个 KPI 的收入乘以 KPI。对于收入 KPI,此值与 KPI 相同。如果 KPI 不是收入,并且“每个 KPI 的收入”数据不可用,则收入未定义。
每个 KPI 的收入:每个 KPI 单位产生的假定收入。这可能因时间、地理区域或两者共同影响而异。Meridian 会将增量 KPI 单位乘以每个 KPI 的收入,以估计处理变量的增量收入。
饱和度:Meridian 假定付费媒体和自然媒体的边际回报会递减,并且在给定时间段内媒体效应存在渐近极限。随着支出沿着响应曲线增加,mROI 会递减。随着支出增加、mROI 变低,渠道会被视为“已饱和”。“饱和度”是一个通用术语,没有定义特定的阈值。
处理变量:包含 MMM 会估计其因果效应的所有变量,即付费媒体、自然媒体和非媒体处理变量。“处理”一词源自因果推理领域,在其他地方通常与“干预”或“曝光”同义。
付费媒体变量:包含有支出数据的所有媒体渠道。这既包含在模型中纳入了单一变量(如支出、展示次数或点击次数)的渠道,也包含在模型中纳入了覆盖面和频次数据的渠道。
自然媒体变量:包含没有关联费用或费用未知的所有媒体渠道。这些渠道与付费媒体一样,在模型中将 Adstock 和回报递减效应纳入考量。主要区别在于,自然渠道的 ROI 和 mROI 无法进行衡量,因此也无法为自然渠道使用 ROI 和 mROI 先验。这既包含在模型中纳入了单一变量(如支出、展示次数或点击次数)的自然媒体渠道,也包含在模型中纳入了覆盖面和频次数据的渠道。
非媒体处理变量:包含任何非媒体策略(例如价格和促销优惠)。Meridian 会估计这些变量的因果效应,但会假定效应值是线性的,而不具有 Adstock 和回报递减效应。
如未另行说明,那么本页面中的内容已根据知识共享署名 4.0 许可获得了许可,并且代码示例已根据 Apache 2.0 许可获得了许可。有关详情,请参阅 Google 开发者网站政策。Java 是 Oracle 和/或其关联公司的注册商标。
最后更新时间 (UTC):2025-08-17。
[null,null,["最后更新时间 (UTC):2025-08-17。"],[[["\u003cp\u003eMeridian MMM (Marketing Mix Modeling) quantifies the impact of marketing activities, including paid media, organic media, and non-media tactics, on key performance indicators (KPIs).\u003c/p\u003e\n"],["\u003cp\u003eIt utilizes control variables to isolate the causal effects of marketing efforts and reduce bias in estimates.\u003c/p\u003e\n"],["\u003cp\u003eThe model incorporates concepts like adstock, diminishing returns, and saturation to reflect real-world marketing dynamics.\u003c/p\u003e\n"],["\u003cp\u003eKey metrics provided by the model include incremental outcome, contribution, ROI (Return on Investment), mROI (marginal ROI), and CPIK (Cost Per Incremental KPI), enabling data-driven marketing decisions.\u003c/p\u003e\n"],["\u003cp\u003eMeridian defines and calculates various marketing concepts, including baseline, expected outcome, response curves, and effectiveness, providing a comprehensive understanding of campaign performance.\u003c/p\u003e\n"]]],[],null,["# Glossary\n\n**Baseline:** The expected outcome under the counterfactual scenario where all\ntreatment variables are set to their baseline values. For paid and organic\nmedia, the baseline values are zero. For non-media treatment variables, the\nbaseline value can be set to the observed minimum value of the variable\n(default), the maximum, or a user-provided float.\n\n**Contribution:** Each treatment variable's incremental outcome as a percent of\ntotal outcome. For reporting purposes, such as\n[`ModelFit`](/meridian/reference/api/meridian/analysis/visualizer/ModelFit)\nExpected versus Actual plot and\n[`MediaSummary`](/meridian/reference/api/meridian/analysis/visualizer/MediaSummary#plot_contribution_waterfall_chart)\ncontribution waterfall chart, total outcome is total *expected* outcome. For the\npurposes of\n[contribution priors](/meridian/docs/basics/roi-mroi-contribution-parameterizations#contribution),\ntotal outcome is total *observed* outcome.\n\n**Control variables:** Variables in the model that aren't treatment variables.\nControl variables are used to estimate baseline outcome, and it is not possible\nto estimate causal effects or contribution percentages for control variables.\nSee [Control Variables](/meridian/docs/advanced-modeling/control-variables) for\nimportant practical advice on selecting controls. Also, see the related concept\n*Mediator variables*.\n\n- **Confounding variables:** Variables that have a causal effect on both the\n treatment and the KPI. Including these as control variables debiases the\n causal estimates of the treatment on the KPI.\n\n- **Predictor variables:** Variables that have a causal effect on the KPI, but\n nothing else. Including these as control variables does nothing to debias the\n causal estimates of the treatment on the KPI. However, strong predictors can\n reduce the variance of causal estimates.\n\n**Cost Per Incremental KPI (CPIK):** Total spend divided by total incremental\nKPI. When the KPI is not revenue and revenue per KPI data is not available, then\nCPIK equals one over ROI.\n\n**Effectiveness:** Incremental outcome divided by total media units.\n\n**Expected outcome:** The expected outcome when all treatment variables are set\nto actual historical values. This is the sum total of baseline outcome plus the\nincremental outcome of all treatment variables.\n\n**Flighting pattern:** The relative distribution of media units across\ngeographic regions and time periods for a given media variable. This is used to\nallocate media units across geographic regions and time periods when the total\nbudget of a channel is scaled up or down, which applies to budget optimization\nand response curves.\n\n**Incremental outcome:** The change in expected outcome driven by each treatment\nvariable. For paid and organic media, this is the change in expected outcome\nwhen one variable is set to zero. For non-media treatment variables, this is the\nchange in expected outcome when a variable is set to its baseline value\n(observed minimum value of the variable (default), the maximum, or a\nuser-provided float) for every geographic region and time period. See\n[Incremental Outcome](/meridian/docs/basics/incremental-outcome-definition) for\ndetails.\n\n**KPI:** The response (target, dependent) variable of the model. It can be\nrevenue, sales units, conversions, or anything else that the treatment variables\nmay have a causal effect upon.\n\n**Lagged effect:** A causal effect of treatment variables from previous time\nperiods affecting the outcome in a later time period. Meridian models\nlagged effects using an adstock function.\n\n**Marginal ROI (mROI):** The derivative of the response curve and is\napproximately the ROI on the next monetary unit (such as dollar) spent beyond\ncurrent spend level.\n\n**Media execution:** A general term referring to the media unit values of a\ngiven channel across geographic regions and time periods.\n\n**Mediator variables:** Variables that are causally affected by the treatment\nand have a causal effect on the KPI. Including these as control variables causes\na bias in causal estimates of the treatment on the KPI. They should be excluded\nfrom the model.\n\n**Outcome:** The primary metric of interest that Meridian measures the\ncausal effect of treatment variables upon. This is typically revenue, but when\nthe KPI is not revenue and revenue per KPI data is not available, then\nMeridian defines the outcome to be the KPI itself. It is not necessarily\nthe response variable of the model (see the KPI definition).\n\n**Response curve:** A plot of incremental outcome versus spend level for a given\nmedia variable. As the spend varies, media units are allocated across geographic\nregions and time periods according to the flighting pattern.\n\n**Return on Investment (ROI):** Meridian defines ROI as incremental\noutcome divided by spend. When the KPI is revenue or revenue per KPI data is\navailable, the incremental outcome is incremental revenue. Otherwise, the\nincremental outcome is incremental KPI.\n\n**Revenue:** For non-revenue KPIs, this is the revenue per KPI multiplied by the\nKPI. For revenue KPIs, this is the same as the KPI. When the KPI is not revenue\nand revenue per KPI data is not available, revenue is undefined.\n\n**Revenue per KPI:** The assumed revenue generated per KPI unit. This can vary\nby time, geographic region, or both. Meridian multiplies incremental KPI\nunits by the revenue per KPI to estimate incremental revenue of the treatment\nvariables.\n\n**Saturation:** Meridian assumes that paid and organic media have\ndiminishing marginal returns, and that there is an asymptotic limit on the media\neffect over a given time period. As spending increases along the response curve,\nthe mROI diminishes. As the spending becomes large and mROI becomes small, a\nchannel is considered to have become *saturated* . *Saturation* is a general\nterm, and no specific threshold is defined.\n\n**Treatment variables:** Includes all variables for which the MMM estimates a\ncausal effect, namely paid media, organic media, and non-media treatments. The\nterm *treatment* comes from the field of causal inference and is elsewhere often\nused synonymously with *intervention* or *exposure*.\n\n- **Paid media variables:** Includes all media channels for which spend data is\n available. This includes both channels modeled with a single variable (for\n example, spend, impressions, or clicks) and channels modeled with reach\n and frequency data.\n\n- **Organic media variables:** Includes all media channels that don't have an\n associated cost, or where the cost is unknown. These channels are modeled with\n adstock and diminishing returns, just like paid media. The primary difference\n is that ROI and mROI cannot be measured for organic channels, and so ROI and\n mROI priors cannot be used for organic channels. This includes both organic\n media channels modeled with a single variable (for example, spend,\n impressions, or clicks) and channels modeled with reach and frequency\n data.\n\n- **Non-media treatment variables:** Includes any non-media tactics such as\n pricing and promotions. These are variables for which Meridian\n estimates a causal effect, but the effect size is assumed to be linear rather\n than having adstock and diminishing returns."]]